A SEMINAR REPORT ON INTELLIGENT WASTE CLASSIFICATION Submitted in partial fulfilment of the requirement for the award of the degree of BACHELOR OF TECHNOLOGY In COMPUTER SCIENCE AND ENGINEERING Submitted by P. KRISHNA PANCHAJANYAM – 20841A05C0 DEPARTMENT OF COMPUTER SCIENCE ENGINEERING AURORA’S TECHNOLOGICAL AND RESEARCH INSTITUTE Approved by AICTE and Affiliated to JNTUH Accredited by NAAC with ‘A’Grade Parvathapur, Uppal, Medipally (M), Medchal (D). Hyderabad – 500098 NOVEMBER, 2023 AURORA’S TECHNOLOGICAL AND RESEARCH INSTITUE Approved by AICTE and Affiliated to JNTUH Accredited by NAAC with ‘A’ Grade Parvathapur, Uppal, Medipally (M), Medchal (D).Hyderabad – 500098 DEPARTMENT OF COMPUTER SCIENCE ENGINEERING CERTIFICATE Certified that seminar work entitled “INTELLIGENT WASTE CLASSIFICATION” is a bonafide work carried out in the fourth year by PERAVALI KRISHNA PANCHAJANYAM (20841A05C0) in partial fulfilment for the award of degree of Bachelor of Technology in Computer Science Engineering from JNTU Hyderabad during the academic year 2023-2024. Ms. V. Shilpa Ms. A. Gowthami Ms. A. Durga Pavani Guide Coordinator HOD Associate Professor Assistant Professor Department of CSE Department of CSE Department of CSE II ACKNOWLEDGEMENT I would like to express my gratitude to Ms. V. Shilpa, Technical Seminar Guide, Associate Professor for her guidance, valuable suggestions and encouragement in completing the Technical Seminar work within the stipulated time. I would like to express my gratitude to Ms. A. Gowthami, Technical Seminar Coordinator, Assistant Professor for her guidance, valuable suggestions and encouragement in completing the Technical Seminar work within the stipulated time. I would like to express my immense gratitude to Ms. A. Durga Pavani, Head of the Department CSE, ATRI, Associate Professor for her guidance, valuable suggestions and encouragement in completing the Technical Seminar work within the stipulated time. I would like to express my sincere thanks to Dr. A. Mahesh Babu, Principal, Aurora’s Technological and Research Institute, Parvathapur, Hyderabad – 500098 for permitting me to do my Technical Seminar work. Finally, I would also like to thank the people who have directly or indirectly helped me, my parents and my friends for their cooperation in completing the Technical Seminar work. P. KRISHNA PANCHAJANYAM 20841A05C0 III ABSTRACT Globally waste is being increased a lot. Traditional waste classification methods rely heavily on manual classification, which is labour intensive, error-prone, and often inefficient. This report explores the transformative potential of intelligent waste classification system in addressing these issues. The accumulation of solid waste in the urban areas is becoming a great concern, and it would result in environmental pollution and may be hazardous to human health if it is not properly managed. It is important to have an advanced/intelligent waste management system to manage a variety of waste materials. One of the most important steps of waste management is the separation of the waste into the different components and this process is normally done manually by handpicking. To simplify the process, intelligent waste material classification system was introduced, which was developed with Robotic arm for picking the waste and the Masked Regional Convolutional neural networks(R-CNN) is a deep learning tool and serves as the extractor, used to classify the waste into different groups/types such as glass, metal, paper, and plastic etc. IV LIST OF CONTENT ACKNOWLEDGEMENT III ABSTRACT IV LIST OF CONTENT IV LIST OF FIGURES VII LIST OF TABLES IX 1. INTRODUCTION 1.1 Solid waste 1.1.1 Types of solid waste 1.2 Effects of solid waste 1 1 1 2 1.2.1 Effect of solid waste on environment 2 1.2.2 Effect of solid waste on living beings 3 1.3 Management of solid waste 1.3.1 Recycling of solid waste 2. WASTE CLASSIFICATION 2.1 Existing methods of waste classification 3 4 5 5 2.1.1 Hand Classification Method 5 2.1.2 Semi-hand classification Method 6 2.2 Industrial waste classification methods 6 2.2.1 Wasatch Integrated Waste Management District (WIWD) 7 2.2.2 TOMRA Sorting Solutions 7 3. EVOLUTION OF DEEP LEARNING 9 3.3 Convolutional Neural Network (CNN) 10 V 3.3.1 Convolutional layer 11 3.3.2 Pooling Layer 12 3.4 Fast Regional Convolutional Neural Networks (Fast R-CNN) 13 3.5 Regional Convolutional Neural networks 14 3.6 Masked Regional Convolutional Neural Networks 16 3.7 Residual Network-50 (ResNet-50) 17 3.7.1 ResNet-50 Architecture 18 4. INTELLIGENT WASTE CLASSIFIER 4.1 Components 4.1.1 Waste feeder 4.1.2 Trommel 20 21 21 22 4.1.3 Conveyor belt: 22 4.1.4 Camera 23 4.1.5 Robotic arm 23 4.1.6 Vacuum gripper 24 5. DEEP LEARNING WASTE CLASSIFICATION 27 5.1 Making Dataset 27 5.1.1 Synthetic complex dataset 28 5.2 Training of Masked R-CNN Model 30 5.2.1 Training Analysis 31 5.2.2 Training results of R-CNN 32 5.2.3 Obstacles Associated With Detecting And Classifying Solid Waste 34 6. WORKING OF INTELLIGENT WASTE CLASSIFIER 36 7. ADVANTAGES & LIMITATTIONS 37 7.1 Advantages 37 VI 7.2 Limitations 37 8. CONCLUSION 39 9. REFERENCES 40 VII LIST OF FIGURES Figure no. Description Page no. 1.1 Statistics of waste generation around world 1 1.2 Solid waste types 3 3.1 Convolutional neural networks architecture 10 3.2 Basic convolution operation 11 3.3 Max pooling operation 12 3.4 Architecture of Regional Convolutional Neural networks 15 3.5 Architecture of ResNet-50 18 4.1 Intelligent waste classifier robotic arm 20 4.2 Photograph of intelligent waste classifier in working state 21 4.3 Waste Feeder 21 4.4 Trommel in industry 22 4.5 A Conveyor belt and components 23 4.6 ZED AI based camera 23 4.7 ABB IRB360 delta robot (robotic arm) for picking waste 24 4.8 Vacuum gripper 24 4.9 Two types of air flow. (a) Vacuum suction (blue arrow). (b) 25 Compressed air (red arrow) used for disposal. VIII 5.1 The basic data set, including representations of each class of studied material: (a) aluminum (first row), paper and cardboard (second row), PET bottles (third row), and nylon (fourth row). (b) The augmented data set. 27 5.2 Four samples from the training data set presented with 29 annotated masks 5.3 Material classification using Mask R-CNN. 34 IX LIST OF TABLES Table no. Description Page no. 5.1 The training parameter configuration 30 5.2 The training and validation loss 31 5.3 The classification results of R-CNN 33 X 1. INTRODUCTION In our world waste is increasing everyday with time because of population which results increase in demand of daily use products due to this waste too is increasing. Let’s see in detailed manner about solid waste. 1.1 Solid waste Solid waste refers to any type of garbage, trash, refuse or discarded material. Figure 1.1: Statistics of waste generation around world It can be categorized according to where the waste is generated, for example as municipal solid waste, health care waste and e-waste. Over 2 billion tons of municipal solid waste are produced annually. The world generates 2.01 billion tonnes of municipal solid waste annually, with at least 33 percent of that—extremely conservatively—not managed in an environmentally safe manner. Worldwide, waste generated per person per day averages 0.74 kilogram but ranges widely, from 0.11 to 4.54 kilograms. Though they only account for 16 percent of the world’s population, high-income countries generate about 34 percent, or 683 million tonnes, of the world’s waste. The World Bank report showed that there are almost 4 billion tons of waste around the world every year and the urban alone contributes a lot to this number. In the next 25 years, waste accumulation in less developed countries will increase more drastically. 1.1.1 Types of solid waste The following are some common types of solid waste: 1) Municipal Solid Waste (MSW): This is the everyday waste generated by households 1 and businesses. It includes items like food scraps, packaging materials, clothing, appliances and other non-hazardous materials. 2) Construction and Demolition (C&D) Waste: This category encompasses waste produced during the construction, renovation, and demolition of buildings and infrastructure. It includes materials like concrete, wood, drywall, and metals. 3) Electronic Waste (E-Waste): E-waste consists of discarded electronic and electrical equipment, including old computers, mobile phones, televisions, and other electronic devices. 4) Recyclable Waste: Materials that can be collected, processed, and reused to make new products, such as paper, cardboard, glass, and certain plastics. 5) Green Waste: Organic waste from gardening and landscaping activities, such as grass clippings, tree branches, and leaves. 6) Plastic Waste: plastic waste refers to the waste generated by the products which are made with plastic such as plastic cups, PET bottles, caps etc. 1.2 Effects of solid waste Waste is defined as unwanted and unusable materials and is regarded as a substance which is of no use. Waste that we see in our surroundings is also known as garbage. Garbage is mainly considered as a solid waste that includes wastes from our houses (domestic waste), wastes from schools, offices, etc. (municipal wastes) and wastes from industries and factories (industrial wastes). 1.2.1 Effect of solid waste on environment Improper disposal and mismanagement of solid waste have profound and detrimental effects on the environment. The pollution of land and soil, a consequence of landfills and illegal dumping, can contaminate the soil and hinder natural plant growth, diminishing biodiversity. Solid waste often leaches harmful chemicals into groundwater, resulting in water pollution. Moreover, when waste is not handled responsibly, it can be carried by surface runoff into rivers, lakes, and oceans, exacerbating water pollution. The open burning of waste releases toxic air pollutants and greenhouse gases, contributing to air pollution and climate change. Landfills, in particular, emit methane, a potent 2 greenhouse gas that adds to global warming. Wildlife and ecosystems suffer as animals ingest or become entangled in discarded waste, especially plastics, disrupting natural habitats and diminishing biodiversity. Furthermore, the presence of unsightly waste negatively impacts the aesthetics of an area and can expose humans and animals to health risks, such as respiratory issues and infections. Solid waste generation often relies on the extraction of natural resources, contributing to resource depletion. To mitigate these environmental impacts, it is crucial to promote responsible waste management practices, reduce waste generation, and enforce regulations to safeguard the environment and human health. 1.2.2 Effect of solid waste on living beings Figure 1.2: Solid waste types Without proper recycle of waste, especially plastic causes dangerous threat, not only to humans but also to other living beings. Mainly cows, buffaloes and monkeys are consuming plastic wrapped food and bags and are being died due to indigestion. Many marine creatures too are being suffered with solid waste, many whales and other aqua creatures are entangled with fishing nets and plastic ropes, which causes suffocation and dying. Hence it is necessary to recycle the waste to protect the environment. 1.3 Management of solid waste The main method of managing the waste is landfilling, which is inefficient and 3 expensive and polluting natural environment. For example, the landfill site can affect the health of the people who stay around the landfill site. Another common way of managing waste is burning waste and this method can cause air pollution and some hazardous materials from the waste spread into the air which can cause cancer. 1.3.1 Recycling of solid waste By considering 4R’s method i.e. reduce, reuse, recycle and recover, one can manage the waste in eco-friendly manner. One of the efficient way of managing solid waste is by recycling. So to protect environment and to reduce waste we need to recycle the waste. In recycling, one of the important stage is classification of waste according to their type. In very few countries various bins are used to dispose wastes according to their type. Mostly in many countries various wastes are disposed in single bin and buried under earth. Instead of burying under earth we can reuse the plastic by some methods to deform the existing plastic into required shape. 4 2. WASTE CLASSIFICATION Waste classification is a crucial aspect of modern waste management systems, aimed at categorizing and segregating different types of waste materials based on their characteristics, composition, and potential environmental impact. This process enables the proper handling, disposal, and, ideally, the recycling or recovery of waste materials, minimizing the negative effects on the environment. Waste is typically classified into various categories, including municipal solid waste (MSW), hazardous waste, industrial waste, electronic waste (e-waste), and more. Each category requires specific handling and treatment methods to ensure the safety of human health and the preservation of the environment. Waste classification is integral to sustainable waste management practices, as it allows for the identification and prioritization of materials that can be reduced, reused, recycled, or recovered. By differentiating between recyclable, hazardous, and non-hazardous waste, it becomes possible to divert materials away from landfills and incinerators, promoting resource conservation and reducing pollution. Proper waste classification also facilitates the implementation of recycling programs and the efficient classification of waste streams in recycling facilities. Through accurate and systematic waste classification, communities can work towards a more responsible and environmentally-friendly approach to waste management, which ultimately benefits both the planet and its inhabitants. 2.1 Existing methods of waste classification Mainly waste is classified to recycle the sorted waste or to reuse the sorted waste. Mainly they contain plastic bottles, aluminium cans, bottle caps, paper, plastic bags, etc. Existing methods to classify the waste is hand picking and semi hand picking, let’s them in more detailed manner in further sections. 2.1.1 Hand Classification Method Hand classification is a manual waste separation method where trained workers physically categorize waste into recyclables, compostable, and non-recyclables. Waste management relies heavily on hand classification. The advantages of hand classification include producing high-quality recyclables and being a low-cost, easily implemented 5 solution. However, it is labour-intensive, can expose workers to hazardous materials, and may not be cost-effective for large-scale operations. Efforts have been made in recent years to increase the efficiency of hand classification through the use of technology, such as handheld devices that can assist workers in identifying materials. Manual classification still plays a crucial role in the efficient operation of solid waste management systems because it has an influence on the efficiency and sustainability of- waste separation and recycling processes. Basically human power is enough for classification but, due to working in dump yards for more time, they are facing a lot of health problems like asthma, lung cancer, allergies etc. so, if we could automate the process of classification, it will be helpful to humans and may reduce solid waste and can recycle it and it contributes in saving environment. 2.1.2 Semi-hand classification Method It is a hybrid waste separation method that combines manual and mechanical classification techniques. It involves trained workers separating waste into different categories, but with the aid of technology such as conveyor belts, screens, and sensors. The main advantage of semi-hand classification is that it combines the benefits of both manual and mechanical classification methods. It can produce high-quality recyclables and reduce the health and safety risks associated with hand classification. However, there are several limitations to this method that should be considered, such as human error, labour-intensiveness, health and safety risks, limited classification accuracy, and limited scalability. Human error can lead to inconsistencies in the classification process due to human error, labour-intensiveness, health and safety risks, limited classification accuracy, and limited scalability. While semi-hand classification has been used effectively in waste management for many years, it has several limitations that need to be considered. These include human error, labour intensiveness, health and safety risks, limited classification accuracy, and limited scalability. As a result, waste management organizations should consider other more advanced and automated classification technologies to improve the efficiency and accuracy of their operations. 2.2 Industrial waste classification methods Sources of waste can be broadly classified into four types: Industrial, 6 Commercial, Domestic, and Agricultural. Recycling of waste product is very important as this process helps in processing waste or used products into useful or new products. Recycling helps in controlling air, water, and land pollution. It also uses less energy. There are a number of items that can be recycled like paper, plastic, glass, etc. Recycling helps in conserving natural resources and also helps in conserving energy. Recycling helps in protecting the environment as it helps in reducing air, water, and soil pollution. 2.2.1 Wasatch Integrated Waste Management District (WIWD) The Wasatch is a government-run organization that provides waste management services for the Wasatch front in Utah. The district operates landfills, transfer stations, and recycling centres to handle the waste of more than 800,000 residents and businesses. WIWD employs cutting-edge technologies to reduce waste and advance ecological practices. Among these technological advancements is the implementation of energyfrom-waste systems, which transform waste materials into usable energy. The district also operates a composting facility that turns yard waste into a valuable soil amendment. The district's revenue is generated primarily through the sale of landfill space and services. WIWD also receives funding from grants and assessments of residential and commercial properties. The district uses its revenue to maintain its facilities, implement sustainability programs, and provide education and outreach to the community. The fiscal resources of the district come principally from service fees and the energy sales. The district imposes a routine charge on residential and commercial containers utilizing automated side-load mechanisms. For the intake of all other waste categories, a dedicated fee per unit of waste, known as a tipping fee, is implemented. In short, WIWD is a leader in waste management and sustainability, utilizing semi-automatic cutting-edge technology to reduce waste and generate revenue. Its commitment to sustainability and community education makes it a model for other waste management organizations. 2.2.2 TOMRA Sorting Solutions TOMRA is a leading provider of sensor-based sorting technology for the solid waste management industry. The company's advanced sorting methodology is designed to efficiently separate valuable resources from the waste stream and divert them from landfills to be recycled or reused. TOMRA's solid waste sorting methodology is based on 7 advanced sensor technology, including near-infrared (NIR) and X-ray fluorescence (XRF), which allows for highly accurate and efficient sorting of waste materials. The company's sorting systems can detect and separate a wide range of materials, including plastics, paper, metals, glass, and organic waste. The sorting process starts with the waste being fed into the machine, where it is separated into different streams based on material type and quality. The sensors then detect the properties of each material, such as its density, size, and colour, and a high-speed ejector system separates the material into different streams based on these properties. In terms of yearly income, TOMRA Sorting Solutions is a subsidiary of the Norwegian company, TOMRA Systems ASA, which is publicly traded on the Oslo Stock Exchange. According to its most recent financial statements, the company generated revenue of approximately 10.1 billion Norwegian Kroner (approximately $1.2 billion USD) in 2020, representing a year-over-year increase of 7%. This revenue is primarily generated through the sale of sensor-based sorting equipment. TOMRA Sorting Solutions has emerged as a prominent player in the domain of solid waste management, offering advanced sorting technology that enables its customers to increase recycling rates, reduce waste, and improve sustainability. 8 3. EVOLUTION OF DEEP LEARNING The evolution of deep learning can be traced back to the 1950s, when researchers first began to develop artificial neural networks. Neural networks are inspired by the structure and function of the human brain, and they are able to learn complex patterns from data. In the early days of deep learning, neural networks were very simple and could only learn simple patterns. However, over time, researchers developed more powerful neural network architectures and training algorithms. This led to a breakthrough in 2006, when Geoffrey Hinton and his team showed that deep neural networks could be used to achieve state-of-the-art results on a variety of image classification tasks. Since then, deep learning has revolutionized many fields, including computer vision, natural language processing, and machine translation. Deep learning models are now used in a wide range of applications, such as self-driving cars, facial recognition, and medical diagnosis. Here is a timeline of some of the key milestones in the evolution of deep learning: 1943: Warren McCulloch and Walter Pitts develop the McCulloch-Pitts neuron, a simple mathematical model of a neuron. 1950: Frank Rosenblatt develops the perceptron, a single-layer neural network. 1960: David Rumelhart, Geoffrey Hinton, and Ronald Williams develop the back propagation algorithm, which allows neural networks to learn from their mistakes. 1980: Yann LeCun develops the convolutional neural network (CNN), a type of neural network that is well-suited for image recognition. 1990: Jürgen Schmidhuber and Sepp Hochreiter develop the long short-term memory (LSTM) network, a type of recurrent neural network that is well-suited for sequential data such as text and audio. 2006: Geoffrey Hinton and his team show that deep neural networks can be used to achieve state-of-the-art results on a variety of image classification tasks. 2012: AlexNet, a deep CNN, wins the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), a prestigious competition for image classification. 2014: Google Translate begins using deep learning models to achieve state-of-the-art 9 results in machine translation. 2016: AlphaGo, a deep learning-powered Go program, defeats Lee Sedol, a professional Go player. 2017: Self-driving cars powered by deep learning begin to be tested on public roads. Today, deep learning is one of the most important and rapidly growing fields in artificial intelligence. Deep learning models are being used to solve a wide range of problems in many different industries. 3.1 Convolutional Neural Network (CNN) The convolutional neural network, often known as a CNN, is an example of a feed-forward artificial neural network. It belongs to the field of deep learning, which is a subset of machine learning, and can be seen as a rebranded version of Artificial Neural Figure 3.1: Convolutional neural networks architecture Networks (ANNs). CNNs have experienced significant advancement in the realm of image recognition. Consequently, they are particularly utilized for the purpose of analysing visual data and accomplishing tasks that go beyond simple classification. The key strength of CNN lies in its capability to derive meaningful features from an image and utilize these features as the foundation for its training process. Processing and feature extraction are performed differently compared to a traditional neural network using this system. For accurate identification of an image's contours, it is necessary to consider the local arrangement of pixels. CNN begins by detecting smaller local patterns on an image and then stringing those patterns together to form more complex shapes. Convolutional 10 Neural Networks consist of an input layer, an output layer, multiple hidden layers, and an extensive set of parameters that facilitate the learning of complex objects and patterns. CNN subsamples the input data using convolution and pooling processes before applying an activation function. The first layer of a standard CNN is the input layer, which is then followed by layers that use alternating combinations of pooling and convolutional operations. Ultimately, the fully connected layer at the end of the network is followed by the Soft Max activation function or output layer. 3.1.1 Convolutional Layer Convolution layer applies filters that conduct convolution operations to scan the input I according to its dimensions. It can be defined by adjusting the size of the filter F and the stride S. The output O of this operation is called feature map or activation map. The filter and stride layers make up the convolution layer, and it is the most important part of the CNN core. The convolution process is carried out by calculating the dot Figure 3.2: Basic convolution operation 11 product between the image pixels and the filter. Furthermore, the computed result gets added to the filter area. The term "stride" refers to the magnitude of the steps taken during changing. Following this, the filter proceeds to perform comparable operations across the complete image area by moving horizontally and vertically. 3.1.2 Pooling Layer The pooling layer is a down sampling operation typically applied after a convolutional layer. In particular, the most popular types of pooling layer are max pooling and average pooling. It is also known as down sampling or subsampling operation, reduces the size of the output data from the convolution layer. This technique is employed to decrease the count of parameters that can be learned and the dimensions of the images, thus simplifying the learning process without compromising the information and features of the input. Additionally, it proves beneficial in preventing over fitting. When it comes to various pooling functions, two widely employed ones are max pooling and average pooling. These techniques effectively decrease the number of Figure 3.3: Max pooling operation 12 connections to the subsequent layer, which is typically a fully connected layer. The output of a convolution layer is a 4x4 matrix, and when a max-pooling operation is applied, the multiplication of those matrices begins at the top-left corner of the image. The max pooling operation returns the maximum value and then shifts to the right by an amount determined by the allocated stride value, is equal to 2. Since the channel is unaffected by the max-pooling operation, the output matrix will always be of the form 2x2. 3.2 Fast Regional Convolutional Neural Networks (Fast R-CNN) Fast R-CNN, a seminal deep learning model introduced by Ross Girshick in 2015, revolutionized object detection by significantly improving both speed and accuracy compared to its predecessor, R-CNN. This model addressed several inefficiencies in the original R-CNN architecture, making it more practical for real-world applications. The fundamental innovation of Fast R-CNN lies in its ability to share computation across multiple region proposals. Instead of processing each region independently, Fast R-CNN efficiently extracts features from the entire image using a single forward pass through the convolutional neural network (CNN). This shared computation dramatically reduces redundancy and accelerates the object detection process. The Fast R-CNN pipeline consists of several key components. The first step involves generating region proposals using an external algorithm, such as Selective Search. These proposals serve as potential bounding boxes for objects within the image. Unlike R-CNN, which processed each proposal separately, Fast R-CNN introduced the Region of Interest (RoI) pooling layer to extract features from all proposals simultaneously. The RoI pooling layer is a critical component that enables efficient feature extraction. It takes the convolutional feature map from the entire image and extracts features only from the regions specified by the proposals. This pooling operation ensures that the features extracted from different regions are aligned and can be fed into subsequent layers for classification and bounding box regression. Fast R-CNN incorporates a Region Proposal Network (RPN) into the model architecture. The RPN operates on the same convolutional feature map used for object detection, generating region proposals based on anchor boxes. This integration 13 streamlines the overall process, allowing end-to-end training of the model and eliminating the need for a separate proposal generation step. The model's loss function combines classification loss, which determines the object class, and bounding box regression loss, which refines the coordinates of the proposed bounding boxes. This multi-task loss is crucial for training the network to accurately classify and localize objects within the image. One notable advantage of Fast R-CNN is its improved speed. By sharing computation across the entire image, the model eliminates the need for redundant feature extraction for each region proposal, making it more computationally efficient during both training and inference. This efficiency made Fast R-CNN a practical choice for real-time object detection applications. Fast R-CNN has had a profound impact on the field of computer vision, serving as a foundation for subsequent models and advancements. Its success paved the way for Faster R-CNN, which integrated the region proposal generation directly into the model, further improving efficiency and end-to-end training. Despite its achievements, Fast R-CNN has limitations, such as the dependence on external proposal generation algorithms and the sequential nature of the training process. Researchers have continued to build upon the Fast R-CNN framework, addressing these challenges and pushing the boundaries of object detection performance. The legacy of Fast R-CNN persists in the ongoing quest for faster, more accurate, and more efficient deep learning models in the realm of computer vision. 3.3 Regional Convolutional Neural networks Region-based Convolutional Neural Networks (R-CNNs) represent a pivotal advancement in computer vision and object detection. Introduced by Ross Girshick et al. in 2014, R-CNNs address the limitations of traditional CNNs in handling object localization tasks. The fundamental challenge R-CNNs tackle is the accurate identification and localization of objects within an image. R-CNNs operate on the idea of dividing an image into distinct regions, which are then individually analyzed to determine the presence of objects. The process begins with the generation of region proposals, typically achieved through selective search, a method for identifying potential object 14 locations. Each proposed region is then passed through a CNN to extract relevant features. This feature extraction stage allows R-CNNs to understand the unique characteristics of different objects present in the proposed regions. The extracted features are then used to classify the objects within each region and refine the bounding box coordinates. One notable issue with traditional R-CNNs is their computational inefficiency, as they process each region independently. The subsequent development of Figure 3.4: Architecture of Regional Convolutional Neural networks (Sample figure) Fast R-CNNs addressed this concern by sharing convolutional features across multiple region proposals, significantly improving speed. Building upon Fast R-CNNs, Faster RCNNs introduced Region Proposal Networks (RPNs), allowing for an end-to-end trainable system. RPNs generate region proposals directly from the shared convolutional features, streamlining the overall process. The integration of RPNs contributes to the "faster" aspect of Faster R-CNNs, making them more efficient for real-time applications. R-CNN variants have been widely applied in numerous fields, including autonomous vehicles, surveillance, and medical image analysis. The introduction of R-CNNs has led to substantial improvements in object detection accuracy compared to previous methods. One significant advantage of R-CNNs is their ability to handle object detection in complex scenes with multiple objects and varying scales. Despite their success, R-CNNs have limitations, including high 15 computational demands during training and inference. Researchers have continued to explore improvements and optimizations, leading to the emergence of even more efficient models. The development of models like YOLO (You Only Look Once) and an image, improving both speed and accuracy. The interpretability of R-CNNs can be further enhanced by visualizing attention maps, providing insights into the regions of an image that contribute most to the model's predictions. The trade-off between model complexity and efficiency is a key consideration in the design of R-CNN architectures, influencing choices in network depth and parameter count. The development of R-CNNs has spurred research into domain adaptation, exploring methods to enhance the model's performance when applied to data from different distributions. The concept of self-supervised learning is gaining attention in the context of R-CNNs, allowing models to learn useful representations from unlabelled data, potentially reducing the need for extensive labelled datasets. The role of uncertainty estimation in R-CNNs is a topic of ongoing research, addressing the need for models to quantify their confidence in predictions, especially in safety-critical applications. 3.4 Masked Regional Convolutional Neural Networks Mask R-CNN, a ground breaking deep learning model, extends the capabilities of traditional object detection by incorporating instance segmentation. Introduced by Kaiming He et al. in 2017, Mask R-CNN builds upon the success of Faster R-CNN, adding a dedicated branch for predicting segmentation masks alongside bounding boxes and class labels. The primary objective of Mask R-CNN is to precisely delineate the contours of individual objects within an image. This is crucial for applications where understanding the exact boundaries of objects is essential, such as in medical imaging or robotics. The model achieves this by introducing a parallel mask prediction branch that operates in tandem with the existing object detection pipeline. The architecture of Mask R-CNN comprises three main components: the backbone network, the Region Proposal Network (RPN), and the Mask Head. The backbone network is responsible for feature extraction, capturing hierarchical representations of the input image. The RPN generates region proposals, and the Mask Head refines these proposals by predicting both the class labels and the segmentation masks. Training Mask R-CNN involves a multi-task loss 16 function that combines object classification loss, bounding box regression loss, and mask segmentation loss. The model is trained end-to-end, leveraging annotated datasets with pixel-level segmentation masks. One notable advantage of Mask R-CNN is its ability to handle overlapping and occluded objects, providing a more detailed understanding of complex scenes. This makes it well-suited for applications like autonomous vehicles, where accurate instance segmentation is critical for safe navigation. The success of Mask RCNN has led to its widespread adoption in various domains, including image and video analysis, where a granular understanding of object boundaries is essential. Additionally, the model's versatility extends to interactive image segmentation tasks, enabling users to interactively refine or modify segmentation masks. Despite its effectiveness, Mask RCNN has computational demands due to its intricate architecture. Researchers continue to explore ways to enhance the model's efficiency and speed without compromising its segmentation accuracy. As deep learning in computer vision progresses, Mask R-CNN remains a pivotal model, inspiring further innovations in instance segmentation and object understanding. 3.5 Residual Network-50 (ResNet-50) ResNet-50 is a type of convolutional neural network (CNN) that has revolutionized the way we approach deep learning. It was first introduced in 2015 by Kaiming He et al. at Microsoft Research Asia. ResNet stands for residual network, which refers to the residual blocks that make up the architecture of the network. ResNet50 is based on a deep residual learning framework that allows for the training of very deep networks with hundreds of layers. The ResNet architecture was developed in response to a surprising observation in deep learning research: adding more layers to a neural network was not always improving the results. This was unexpected because adding a layer to a network should allow it to learn at least what the previous network learned, plus additional information. To address this issue, the ResNet team, led by Kaiming He, developed a novel architecture that incorporated skip connections. These connections allowed the preservation of 17 information from earlier layers, which helped the network learn better representations of the input data. With the ResNet architecture, they were able to train networks with as many as 152 layers. The results of ResNet were ground breaking, achieving a 3.57% error rate on the ImageNet dataset and taking first place in several other competitions, including the ILSVRC and COCO object detection challenges. This demonstrated the power and potential of the ResNet architecture in deep learning research and applications. Let’s look at the basic architecture of the ResNet-50 network. 3.5.1 ResNet-50 Architecture ResNet-50 consists of 50 layers that are divided into 5 blocks, each containing a Figure 3.5: Architecture of ResNet-50 set of residual blocks. The residual blocks allow for the preservation of information from earlier layers, which helps the network to learn better representations of the input data. 1) Convolutional Layers The first layer of the network is a convolutional layer that performs convolution on the input image. This is followed by a max-pooling layer that downsamples the output of the convolutional layer. The output of the max-pooling layer is then passed through a series of residual blocks. 2) Residual Blocks Each residual block consists of two convolutional layers, each followed by a batch normalization layer and a rectified linear unit (ReLU) activation function. The output of the second convolutional layer is then added to the input of the residual block, which is then passed through another ReLU activation function. The output of the residual block is 18 then passed on to the next block. 3) Fully Connected Layer The final layer of the network is a fully connected layer that takes the output of the last residual block and maps it to the output classes. The number of neurons in the fully connected layer is equal to the number of output classes. 4) Concept of Skip Connection Skip connections, also known as identity connections, are a key feature of ResNet-50. They allow for the preservation of information from earlier layers, which helps the network to learn better representations of the input data. Skip connections are implemented by adding the output of an earlier layer to the output layer. 19 4. INTELLIGENT WASTE CLASSIFIER Intelligent waste classifier is an autonomous system that uses computer vision and deep learning to identify and categorize recyclable materials. The system consists of two main parts: fig 4(a) robotic manipulator and a vision-based material detection and categorization module. The robotic manipulator is responsible for physically separating the waste into different bins according to the material type. The vision-based material detection and categorization module uses cameras to capture images of the waste and then uses deep learning to identify the material type of each item. This information is then sent to the robotic manipulator, which picks up the item and places it in the appropriate bin. The robotic sorting machine offers several advantages over traditional manual sorting methods. First, it is more accurate and consistent. Second, it can work 24/7 without Figure 4.1: Intelligent waste classifier robotic arm breaks. Third, it is less prone to injury. Fourth, it can handle a wider variety of materials. It has the potential to revolutionize the waste sorting industry. It could help to increase the efficiency and effectiveness of waste sorting, which could lead to a decrease in the amount of waste that is sent to landfills. It could also help to create new jobs in the robotics industry. 20 Figure 4.2: Photograph of intelligent waste classifier in working state 4.1 Components 4.1.1 Waste feeder Waste feeders are most suitable for handling light, fine, abrasive and free-flowing materials. Especially these waste feeders are used to dump the solid waste into sorting machine. They are available as horizontal and inclined Belt Feeders for extracting bulk materials from under a dump hopper or regulating feed to a crusher or screen. Belt Feeders help to provide volumetric feed and even metering to prevent flooding. The belt that carries the material may be a fabric reinforced belt or a steel cable reinforced belt. The belt is propelled by a Drive Pulley and Shaft. The drive shaft is rotated by a Drive Figure 4.3: Waste Feeder 21 Mechanism. A supplementary equipment may include a feed chute, skirt board, a discharge chute and a High-Pressure Lubrication System. Belt Feeders are a trusted and proven economical means of transportation for many industries. A dual feeder shed waste at a controlled rate onto the belt. The material came from a waste processing installation that transferred uncontrolled urban waste or a newly developed stream located above the conveyor to feed the system a controlled waste mix for experimentation purposes 4.1.2 Trommel First waste is carried through garbage feeder and feeded into trommel then trommel drum rotates with waste, which at bottom tiny holes are present from that tiny particles like dust, sand, mud is separated. It is a pre-processing method used to remove Figure 4.4: Trommel in industry tiny particles from waste. Then the solid waste is flowed on to conveyor belt. 4.1.3 Conveyor belt: A conveyor belt is a looped belt that is driven by and wrapped around one or more pulleys. It is powered by an electric motor and supported by a metal plate bed or rollers upon which the conveyor belt rests. The pulley that powers a conveyor belt is referred to as the drive pulley and has an unpowered idler pulley. Pulley drives at the discharge end of a conveyor belt are referred to as head drives, while ones located at the in feed end are known as tail drives. The preferred type of pulley drive is a head drive located at the discharge end and uses pull force to move a conveyor belt. Typically, industrial waste processing units transfer recyclables on conveyor belts. It is a replica of such a setup through the installation of a 22.5-m long, 1-m wide (usable width: 0.8 m) belt with a 22 Figure 4.5: A Conveyor belt and components speed of up to 0.15 m/s and an optical encoder system so that the speed could be monitored. 4.1.4 Camera A stereo, full-high-definition ZED camera enabled the automated categorization of recyclable materials. The camera was placed 148 cm in front of the robot (in the direction in which the recyclables were moving), at a height of 75 cm above the conveyor belt, looking downward. To ensure constant light conditions during data acquisition and Figure 4.6: ZED AI based camera system operation, the camera was placed inside a box with LED equipment. The camera’s field of view covered the entire width of the conveyor belt, providing information about the shape and colour of the transported waste. 4.1.5 Robotic arm An ABB IRB360 delta robot was installed above the conveyor belt to enable the picking and transfer of waste to bins. The robot consisted of three high-torque servomotors mounted on a rigid frame, each linked to a different arm. The three arms 23 Figure 4.7: ABB IRB360 delta robot (robotic arm) for picking waste were connected to a central platform, driving it to move very fast and accurately in 3D space. The robot had a payload of 6 kg, which was appropriate for repetitive and rapidly completed applications. 4.1.6 Vacuum gripper Figure 4.8: Vacuum gripper To enable automated picking and placing, we attached a vacuum gripping module 24 to the end effector of the robot; it consisted of a vacuum blower that provided highvolume suction to pick up and hold selected materials. The use of vacuum technology provides a robust and low-cost solution for material transfer. The majority of robotic systems for recyclable sorting rely on a pick-and-place (PnP) process to select objects and physically transfer them to the appropriate material bin. Capitalizing on delta robots, they developed a composite system that performs fast Figure 4.9: Two types of air flow. (a) Vacuum suction (blue arrow). (b) Compressed air (red arrow) used for disposal. recyclable sorting by adopting the PnP approach. Here venture generators use compressed air flowing through a conical orifice to develop a pressure difference. Their main advantages include a short response time that enables fast picking and a powerful gripping force once the suction cup has sufficiently sealed on an object. The use of blowers provides an alternative. Blowers spin synchronized rotors in a chamber to generate a vacuum that is relatively weak but that has a high volume (typically more than 500 m /h). 3 The problem with using blowers for recyclable picking lies in the equipment’s slow response time, which may significantly decelerate waste processing. Industrial conditions, it was occasionally impossible for the suction cup to form a seal with objects, 25 resulting in failures to grab items. To compensate for this, the present work incorporates a blower. The high volume of pumped air increases the system’s ability to pick up significantly deformed objects when there is no perfect suction cup seal. This comes with the cost of slowing the processing speed because the blower valve used to activate/ deactivate the vacuum at the end effector has a sluggish response, on the order of 500 ms/cycle. Therefore, we examined methods to gain speed. When picking, the delay was reduced by sending the vacuum activation command at a properly timed moment while the robot was moving toward the goal. Moreover, to decrease the material disposal delay, we developed a solution that forced compressed air toward an item to detach it from the suction cup. Specifically, a custom T-shaped tube enabled fast switching between the blower vacuum (negative pressure) and the compressed air (positive pressure), right on the suction cap (see Figure 4.1.6). The joint activation/deactivation of binary valves controlling the two air streams (vacuum versus compressed air) significantly reduced the time needed for item disposal. When the robot’s end effector reached the bin, the two valves turned off the vacuum and activated the compressed air, thus making items detach from the suction cup in much less time. 26 5. Deep learning waste identification For identifying the waste we are using deep learning algorithm. Accordingly, any autonomous recovery system needs to accurately categorize recyclables based on their material type. One way to achieve this is to use hyper spectral cameras, although this significantly increases the cost of composite systems. Alternatively, one may exploit recent developments in deep learning to come up with an efficient and much cheaper solution. CNNs trained with deep learning algorithms have been widely applied in demanding computer vision applications. Interestingly, CNNs have also been used to perform waste classification, although they were applied to a relatively small data set with single-item images and without any type of occlusion. In contrast, the present work targets “instance segmentation” to accomplish multiple object identification and labeling. To this end, the well-known Mask Regional CNN (R-CNN) open source network was employed since it has been particularly successful at similar tasks. Mask R-CNN provides a scalable means for categorizing recyclables into a high number of classes. 5.1 Making Dataset Figure 5.1: The basic data set, including representations of each class of studied material: (a) aluminum (first row), paper and cardboard (second row), PET bottles (third row), and nylon (fourth row). (b) The augmented data set. Making a dataset which is more precise is a crucial step for accurate results. So already there were two open source best datasets were available from Trash net and taco 27 but those are not much suitable for industrial use because the garbage which they sort was filled with dust and many objects are piled one up on another so these datasets are not suitable. So they made their own dataset, to make it first they took photographs of single object by allowing to run on conveyor belt with black background as shown in figure 5.1 (a).Then these photos were made as a basic dataset and again made another data set this time they captured the objects which were used in first dataset with the combination of other objects and with different background in that some are piled upon one another as shown in figure 5.1 (b). Typical image data collection methods assume that photos of objects of interest will be taken against a variety of backgrounds and manually labeled using annotation tools. This is a complicated and time-consuming approach that can hardly support the collection of very large data sets. The steps we took to develop a rich waste data set are outlined in the following. The resulting mask was assigned as an image annotation to clearly indicate the region of interest; it was also used for Mask R-CNN training. In the second step, we randomly applied three geometric transformations (translation, rotation, and scaling) to the image and the mask. This resulted in a large data set of single-object images against a black (i.e., a conveyor belt) background. Finally, pairs of objects from the second step were randomly selected (with their masks) and arbitrarily placed over new images with colourful backgrounds to develop complex problem instances. In the case of recyclable object data collection, this step provided a means to examine cases with multiple, possibly overlapping, objects across a range of diverse backgrounds. 5.1.1 Synthetic complex dataset They followed the semi-automated procedure summarized earlier to develop the data set for training Mask R-CNN. They started by collecting 400 images for each of the material types considered in the study, namely, aluminum, paper and cardboard, polyethylene terephthalate (PET) bottles, and nylon. Then, we followed a manual procedure to collect red–green–blue images. All images had the same size, i.e., 800 × 800 pixels, and depicted a single recyclable against a black background. It is noted that, to have a representative data set of recyclables images, we also considered deformed and 28 dirty materials. Figure 5.2: Four samples from the training data set presented with annotated masks These images were processed by following step one (mentioned previously) to identify masks describing regions of the identified objects. We called this the Basic data set. Then, we applied step two (from the preceding) to gather 2,000 images of each material type. We called this the Synthetic Single data set. Finally, following the third processing step, they developed a data set that artificially described complex cases of multiple and overlapping objects. This was the Synthetic Complex data set. In particular, they developed 16,000 images for the training set and 5,600 for the validation set. Figure 5.1.1 includes four random images from the data set that have been successfully created and annotated following the data augmentation process. The first column contains original images. The 29 second and the third columns show each material mask and its class identification as a subtitle. 5.2 Training of Masked R-CNN Model The development of a vision-based module capable of categorizing recyclables into different material types was accomplished using Mask R-CNN, which has been successfully employed to tackle a wide range of object identification and categorization tasks. Mask R-CNN is a deep CNN that simultaneously predicts recyclable object bounding boxes, masks, and material types. We used a public Mask R-CNN implementation that was trained on the data sets in the “Vision-Based Material Categorization” section, using 30% of the images for model learning validation and the remaining 70% for model training. Table 5.1: The training parameters configuration To increase the image processing frame rate, we incorporated ResNet-50 for feature extraction across entire photos. Moreover, a region proposal network (RPN) was employed to scan images in a sliding window manner to identify areas that contained objects. The RPN employed a set of boxes, called anchors, with predefined image locations and scales (there were five scales and three aspect ratios in our implementation) to figure out the size and location of an object on the feature map. Specifying the correct 30 anchor span is critical for obtaining successful classification results. To estimate recyclable objects’ bounding boxes, masks, and material types, dedicated subnet works known as heads work on identified regions of interest to shape the final output. After extensive experimentation with the validation set, we identified the Mask R-CNN training parameters that worked effectively. Before using Mask R-CNN for recyclable identification and classification, a learning process was applied to improve the applicability of the deep neural network to the problem. To bootstrap the learning process, a network pertained on the Common Objects in Context data set was used. This approach, introduced in, speeds up learning and ensures the minimum quality of the results in a reasonable training time. Following this, only the Mask R-CNN heads are typically trained for a problem, keeping the general object feature extraction mechanism unchanged. To adapt Mask R-CNN to the classification task, we developed and tested four customized implementations: Net 1, Net 2, Net 3, and Net 4. They were trained on a high-performance computing infrastructure with Tensor Flow GPU support. 5.2.1 Training Analysis Table 5.2: The training and validation loss Mask R-CNN training is based on a complex loss function that is calculated as the 31 weighted sum of different Partial losses at every training state. The partial losses considered in the current implementation are described in the following. 1) mrcnn_bbox_loss: corresponds to the success at localizing the bounding boxes of objects that belong to a given class. This loss metric is increased when the object categorization is correct but the bounding box localization is not precise. 2) mrcnn_class_loss: describes the loss due to the improper categorization of an object identified by the RPN. This provides an indication of how well Mask R-CNN recognizes each class of an identified object. This metric is increased when an object is detected in an image but mis-classified. 3) mrcnn_mask_loss: It is the metric that summarizes the success at implementing masks across a processed image that correspond to the pixel area covered by the identified objects. The training data set includes masks for all objects of interest in input images, thus providing the ground truth for the evaluation of masks predicted by Mask RCNN. 4) rpn_bbox_loss: corresponds to the localization accuracy of the RPN, in other words, how well the RPN identifies individual objects. High values indicate miscalculations when an object is detected but the bounding box needs to be corrected. 5) rpn_class_loss: It is an RPN performance metric that is increased when an object is undetected at the final output and decreased when the RPN successfully identifies an item. Besides tuning the training procedure parameters, there are conditions that may affect the quality of the obtained results. These include the data set size, the number of the learning steps per epoch, and the use of additional data augmentation strategies to improve generalization. Accordingly, several Mask R-CNN training variations were examined for their training and validation results. 5.2.2 Training results of R-CNN 1) Net 1: We began with the relatively small Basic data set, which consisted of singleobject, black-background images (see the “Vision-Based Material Categorization” section). The training process ran with 50 training steps per epoch, for a total of 200 epochs (this setup was also followed in the next two cases). Although the network was successfully trained (Table 5.2.1), an assessment against real industrial images revealed 32 many limitations that rendered this solution useless in practice. This is clearly indicated in Table 5.2.2, which highlights the recall, precision, and F1 performance of the network for each material and in total. The limited success of this version provided the driving force for developing artificial data augmentation algorithms. 2) Net 2: The next attempt considered the much larger Synthetic Single data set for training Mask R-CNN. The solution yielded slightly better results for identifying individual items, as demonstrated in Table 5.2.2. However, the average F1 score remained low, proving poor overall performance. This was mainly because the training data set presented recyclables against a black (conveyor belt) background. Thus, when this implicit hypothesis was violated (i.e., due to object occlusion), the network performance dropped dramatically. 3) Net 3: A major improvement to the data set concerned the synthetic placement of objects against colorful backgrounds that could also contain other objects. The Synthetic Complex data set was used to train a new Mask R-CNN (the loss metrics are summarized Table 5.3: The classification results of R-CNN in the third row of Table 5.2.2), which provided the first vision-based recyclable identification and categorization solution that could be potentially applied in the real world. As shown in Table 5.2.2, the network’s success rate in demanding industrial conditions significantly increased over that of Net 1 and Net 2. 4) Net 4: To further enhance the Mask R-CNN performance, we made changes in two directions that improved generalization and robustness. First, random affine 33 transformations. 1. R-CNN model had shown average precision of 89.1% 2. It had shown average recall of 91.8% 3. It had shown average F1 score of 89.4% 4. Approximately masked R-CNN was trained with an accuracy of 90% (Which you can see in figure 5.2.3) 5.2.3 Obstacles Associated With Detecting And Classifying Solid Waste 1) Variability in waste types: Waste materials are diverse and can vary greatly in their physical and chemical properties. This variability makes it challenging to develop accurate waste detection and classification models that can identify different types of waste materials. 2) Complexity of waste objects: Waste objects can be of different shapes and sizes, and they can be irregularly shaped, making it difficult to accurately detect and classify them. Environmental factors: Environmental factors such as lighting conditions, dust, and noise can impact the accuracy of waste detection and classification systems. For example, low Figure 5.3: Material classification using Mask R-CNN. 34 lighting conditions can make it difficult to identify waste objects accurately. 3) Cost of equipment: Developing and deploying waste detection and classification systems can be costly. It requires specialized equipment such as sensors, cameras, and other technologies, which can be expensive. 4) Training data availability: To develop accurate waste detection and classification models, large datasets of labelled waste materials are needed. However, these datasets may not be readily available, making it challenging to develop robust and accurate models. 5) Human errors: Human errors can also impact the accuracy of waste detection and classification systems. For example, errors in labelling or sorting waste materials can lead to incorrect identification and classification of waste. 6) Continuous improvement: As new waste materials are introduced and as waste materials change over time, waste detection and classification models need to be continuously improved to ensure their accuracy and effectiveness. This requires ongoing research and development, which can be time-consuming and costly. Therefore, waste detection and classification pose several challenges that require innovative solutions to ensure accurate and effective waste management. 35 6. WORKING OF INTELLIGENT WASTE CLASSIFIER Step 1: First solid waste was picked up with cranes and dumped into garbage feeder. Step 2: then waste in the feeder is lifted up and dropped into trommel. Step 3: then rotating trommel filters tiny particles and drop the solid waste on to conveyor belt. Step 4: the moving conveyor belt move with constant velocity with solid waste towards robotic arm. Step 5: robotic arm sorts the object which was internally connected to system which identifies waste with the trained Mask R-CNN deep learning algorithm. It monitors in real time with ZED quad camera. Step 6: The object is picked with robotic arm which uses vacuum grip to hold the object firmly. Step 7: then finally the robotic arm drops the picked waste into right bin by turning off vacuum. 36 7. ADVANTAGES & LIMITATTIONS 7.1 Advantages 1) Robots can sort waste more accurately than humans, which can lead to a reduction in contamination and an increase in the quality of recycled materials. 2) Robots can sort waste much faster than humans, which can lead to significant time savings and increased productivity. 3) Robots can automate the waste sorting process, which can lead to significant labor cost savings. 4) Robots can work in dangerous or difficult conditions, such as in hazardous waste facilities, which can help to protect workers from injury. 5) Reduced environmental impact: Intelligent waste sorting technology can help to reduce the environmental impact of waste management by improving the efficiency of the sorting process and reducing the amount of waste that is sent to landfills. 6) Increased recycling rates: Intelligent waste sorting technology can help to increase recycling rates by making it easier and more efficient to sort recyclable materials. 7) Improved public health: Intelligent waste sorting technology can help to improve public health by reducing the exposure of workers and the public to harmful waste materials. 8) Reduced greenhouse gas emissions: Intelligent waste sorting technology can help to reduce greenhouse gas emissions by reducing the amount of waste that is sent to landfills, which produce methane, a potent greenhouse gas. 9) Increased energy production: Intelligent waste sorting technology can help to increase energy production by converting waste materials into biofuels or other forms of renewable energy. 7.2 Limitations 1) It cannot sort with 100% accuracy still research need to be done. 2) Climatic conditions may effect the sorting process. Especially in rainy season the waste will be so, this may reduce the efficiency of machine. 3) It cannot open bottle caps to remove left over water from bottles. 37 4) Robotic waste sorting systems are typically designed to sort a specific type of waste, such as plastic, metal, or paper. This means that they may not be able to sort all types of waste, which can require additional sorting steps. 5) The production and operation of robotic waste sorting systems can have an environmental impact. For example, the robots may require energy to operate, and they may produce waste materials during their manufacturing and disposal. 6) The adoption of robotic waste sorting technology may lead to job losses in the waste management industry. It is important to consider the social impact of this technology when making decisions about its adoption. 38 8. CONCLUSION Current solid waste management practices are no longer sustainable. With the exploitation of state-of-the art computer vision and robotics technology, the goal of finding innovative ways to attain sustainable waste management becomes more achievable and realistic. This work presented an intelligent autonomous system to support the recovery of recyclable materials. The system was gradually integrated into the operation of an MRF to support the treatment of waste processed on a daily basis. In comparison to previous works, this article accomplished the following: 1) It introduced a blower-based vacuum system to improve the ability of robotic systems to manipulate recyclables. 2) It provided a new data set of recyclable images and a group of processing tools to facilitate deep learning research into recyclable categorization. 3) It developed a composite autonomous system that successfully identified, localized, and categorized recyclables in a demanding industrial environment. The technology proposed here could significantly impact future waste treatment plants, which are envisioned to be highly automated facilities where no humans will be in direct contact with waste and where almost all recyclables will be recovered. 39 9. REFERENCES [1]. Ellis, C. (2018, September 23). World Bank: Global waste generation could increase 70% by 2050. Waste Dive. Retrieved October 5, 2022, from: https://openknowledge.worldbank.org/entities/publication/d3f9d45e-115f-559b-b14f 2a. [2] Kaza, Silpa; Yao, Lisa C.; Bhada-Tata, Perinaz; Van Woerden, Frank. 2018. What a Waste from: https://www.wasatchintegrated.org/material-recovery-and-transfer-facility/. [3] J. Bobulski, J. Piątkowski, PET waste classification method and Plastic Waste DataBase WaDaBa, Advances in Intelligent Systems and Computing, vol. 681, Springer Verlag, 2018, pp. 57-64. [4] Mitra, A. (2020) Detection of waste materials using deep learning and image processing. [5] Shahab, Sana, et al. “Deep Learning Applications in Solid Waste Management: A Deep Literature Review.” International Journal of Advanced Computer Science and Applications, vol. 13, no. 3, The Science and Information Organization, 2022, From: https://doi.org/10.14569/ijacsa.2022.0130347. [6] Bhandari, Sishir. “Automatic Waste Sorting in Industrial Environments via Machine Learning Approaches - Trepo.” Automatic Waste Sorting in Industrial Environments via Machine Learning Approaches - Trepo, 1 Jan. 2020, from: https://trepo.tuni.fi/handle/10024/123574. [7] Wilts, Henning, et al. “Artificial Intelligence in the Sorting of Municipal Waste as an Enabler of the Circular Economy.” MDPI, 29 Mar. 2021, from: www.mdpi.com/20799276/10/4/28. [8] Sultana, Rumana & Adams, Robert & Yan, Yanjun & Yanik, Paul & Tanaka, Martin. (2020). Trash and Recycled Material Identification using Convolutional Neural Networks 40 (CNN). 1-8, from: www.10.1109/SoutheastCon44009.2020.9249739. [9] Pieber, Simone & Meierhofer, M & Ragossnig, Arne & Brooks, L & Pomberger, Roland & Curtis, Alexander. (2010). Advanced Waste-Splitting by Sensor Based Sorting on the Example of the MT-Plant Oberlaa. [10] R. A. Aral, Ş. R. Keskin, M. Kaya and M. Hacıömeroğlu, "Classification of TrashNet Dataset Based on Deep Learning Models," 2018 IEEE International Conference on Big Data (Big Data), 2018, pp. 2058-2062, doi: 10.1109/BigData.2018.8622212. [11] Kubic, W. L. J., Moore, C. M., Semelsberger, T. A., & Sutton, A. D. (2021, September 23). Recycled paper as a source of renewable jet fuel in the United States. Frontiers. Retrieved April 9, 2023, from https://www.frontiersin.org/articles/10.3389/fenrg.2021.728682/full. [12] Peter Chazhoor, Anthony Ashwin, et al. “Deep Transfer Learning Benchmark for Plastic Waste Classification.” Deep Transfer Learning Benchmark for Plastic Waste Classification, 28 Jan. 2022, from: https://www.intellrobot.com/article/view/4550. [13] Buzby, P. by J. (2022, January 24). Food waste and its links to greenhouse gases and climate change.USDA. Retrieved April 9, 2023, from: https://www.usda.gov/media/blog/2022/01/24/food-waste-and-its-links-greenhousegasesandclimatechange#:~:text=EPA%20estimated%20that%20each%20year,42%20coal %2Dfi red%20power%20plants. [14] Maria Koskinopoulou, Fredy Raptopoulos, George Papadopoulos, Nikitas Mavrakis: Robotic Waste Sorting Technology: Toward a Vision-Based Categorization System for the Industrial Robotic Separation of Recyclable Waste, Published in: IEEE Robotics & Automation Magazine, Issue: 2, DOI: 10.1109/MRA.2021.3066040. 41