Global Academy of Technology Rajarajeshwari Nagar, Bengaluru-560098 Department of Electronics and Communication Engineering Project Phase I Presentation (21ECE76) On “Real-Time Sign Language Translation using CNN-LSTM Hybrid Architecture for Spatial-Temporal Gesture Synthesis and Grammar Refinement. ” PRESENTED BY: Ms C Saanvi Ms Varsha P V Mr Chidaksh Babu 1GA21EC028 1GA21EC165 1GA21EC032 Under the guidance of : Shubha GN 1 CONTENTS 1. Introduction 2. Literature survey 3. Problem definition 4. Objectives 5. Methodology (Option) 6. Expected Results 7. References 2 INTRODUCTION 1. This project aims to bridge the communication gap for individuals relying on sign language by translating gestures into text in real time using advanced deep learning techniques. 2. This project addresses critical challenges such as limited sign language proficiency among the general population, high costs, and limited availability of interpreters, as well as the inability of existing tools to recognize dynamic gestures. 3.LSTM (Long Short-Term Memory) is a type of recurrent neural network (RNN) designed to handle sequential data CNN (Convolutional Neural Network) is a deep learning model that specializes in extracting spatial features from images 4.By combining Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) for temporal sequence modeling, the system captures the complexity of hand gestures. 5.Attention mechanisms optimize the recognition of critical frames within gesture sequences, while the MediaPipe Holistic pipeline ensures accurate hand tracking by capturing 126 hand landmarks in three dimensions. 6.Users can train the system on custom datasets, allowing adaptability to various sign languages and dialects. Real-time processing enables seamless translation of gestures to text, with grammar correction performed using the language_tool_python library to ensure clarity and coherence. 7.The project’s applications span healthcare, education, customer service, workplace accessibility, and social integration, promoting inclusivity and accessibility for individuals with hearing and speech impairments. 3 LITERATURE SURVEY Sl Paper/Journal No.’; name Author Year of name publication Overview 1. Sign Language Production With Pan Xie, Taiying 2024 Latent Motion Transformer - Peng, Yao Du, Applications of Computer Vision Qipeng Zhang (WACV) This research introduces a novel method for Sign Language Production (SLP) that generates high-quality sign videos directly from sign gloss without relying on human poses. It uses a two-stage model: (i) an improved 3D VQ-GAN for learning latent representations and (ii) sequence-to-sequence attention for better feature ordering. To enhance semantic coherence, the model incorporates perceptual and reconstruction losses. Experiments on WLASL and NMFs-CSL datasets show superior performance over previous approaches. 2. Sign Language Translation with Sentence Embedding Supervision German Research Center for Artificial Intelligence 2024 Yasser Hamidullah Josef, van Genabith Cristina EspañaBonet This research introduces a novel gloss-free Sign Language Translation (SLT) method using sentence embeddings instead of gloss annotations, enabling multilingual learning without manual labeling. Evaluated on German (PHOENIX-2014T) and American (How2Sign) sign languages, it significantly outperforms other gloss-free approaches, narrowing the gap with gloss-based systems. 3. Towards Privacy-Aware Sign Phillip Rust, Bowen 2024 Shi, Skyler Language Translation at Scale Wang, Necati Cihan ACL-anthology This research proposes SSVP-SLT, a privacy-aware, two-stage SLT framework using self-supervised pretraining on anonymized videos and supervised fine-tuning. It achieves state-of-the-art performance on How2Sign, addressing both data scarcity and privacy concerns. 4. iSign: A Benchmark for Indian Abhinav,Sudeep Sign Language Processing Modi,Monali ACL-anthology Barbate This research presents iSign, a benchmark for Indian Sign Language (ISL) processing, featuring the largest ISL-English dataset. It introduces multiple NLP tasks like SignVideo2Text and Text2Pose, benchmarks them with baseline models, and provides linguistic insights, streamlining ISL research and enhancing accessibility for the NLP and sign language communities. 5. Unsupervised Sign Language Zhengs Xhui, 2024 Translation and Generation Xing Wang Camgöz 2024 USLNet is an unsupervised sign language translation and generation model that learns from non-parallel text and video data. It employs single-modality reconstruction and cross-modality back-translation, tackling text-video discrepancies with a sliding window method. Experiments on BBC-Oxford and Open-Domain ASL datasets show its 4 effectiveness, rivaling supervised models. LITERATURE SURVEY Sl No. Paper/Journal name Author Year of name publication 6. End-to-End Sign Language Video Yu Duan Generation with Latent Diffusion Changsheng Xu -CVVSR Huaiwen Zhang 7. LLMs are Good Translators-CVVSR Sign Overview · 2024 · This research pioneers end-to-end text-to-sign video translation, capturing full-body and facial expressions for enhanced expressiveness. It improves diffusion models with multimodal conditions and a motion-aware framework for better text-visual alignment. Experiments on multiple datasets demonstrate superior semantic accuracy, naturalness, and expressiveness compared to state-of-the-art methods. Language Jia Gong, Lin Geng 2024 Foo, Yixuan He SignLLM leverages large language models (LLMs) for Sign Language Translation (SLT) by transforming sign videos into a language-like representation. It introduces vector-quantized visual sign tokens and an alignment module to enhance semantic compatibility. Experiments show state-of-the-art gloss-free performance on SLT benchmarks, improving translation readability and accuracy. 8. YouTube-ASL: A Large-Scale, OpenDomain American Sign LanguageEnglish Parallel Corpus -Arrve David Uthus, Tanzer, Georg 2024 YouTube-ASL is a massive ASL dataset with ~1000 hours of videos and >2500 unique signers, significantly larger than previous datasets. It enables state-of-the-art fine-tuned ASL-to-English translation and achieves the first nontrivial zero-shot results. This resource advances machine learning research for ASL by addressing data scarcity and supporting diverse signers. 9. An Isolated American Sign Language Thad Starner, Sean 2024 Dataset Collected via Smartphones Forbes, Matthew -Oxgh So, David Martin PopSign is a smartphone-based game that helps hearing parents of deaf infants learn sign language by integrating sign recognition into gameplay. The PopSign ASL v1.0 dataset, collected from 47 Deaf adult signers using Pixel 4A cameras, features 250 ASL signs with over 175,000 verified examples. A baseline LSTM model achieves 84.2% accuracy, demonstrating its potential for educational use. 10. ASL Citizen: A Community-Sourced Aakasha Desai, 2024 Dataset for Advancing Isolated Sign Lauren Berger Language Recognition -Oxgh ASL Citizen is the first crowdsourced dataset for Isolated Sign Language Recognition (ISLR), featuring 83,399 videos of 2,731 ASL signs from 52 signers in diverse settings. Designed for ASL dictionary retrieval, it enables users to match their demonstrated signs via webcam. Supervised models trained on this dataset achieve 63% accuracy and 91% recall-at-10. Garrett Manfred 5 LITERATURE SURVEY Sl No. Paper/Journal name Author Year of name publication Overview 11. Leveraging Large Language Models Ryan Wong for Gloss-Free Sign Language Richard Brodi Translation-Argv , 2024 Sign2GPT is a framework for gloss-free sign language translation that integrates pretrained vision and language models using lightweight adapters, addressing data scarcity and computational constraints. It introduces a novel pretraining strategy using pseudo-glosses to enhance sign representation learning. Tested on RWTH-PHOENIXWeather 2014T and CSL-Daily, it achieves state-of-the-art performance. 12. An Explicit Multi-modal Fusion Cong Hu, Biao Fu, 2024 Method for Sign Language Translation Pei Yu -Argv Sign Language Production (SLP) faces significant multimodal semantic gaps due to limited cross-modal alignment labeling, unlike sign language translation. Generating complex visual semantics from text remains challenging. Similar issues are addressed in cross-modal retrieval, vision-language understanding, and multimodal sentiment analysis by aligning and comparing cross-modal semantics in rich multimodal data. 13. Unsupervised Sign Language Bian Wu, Chenyong 2024 Translation and Generation-WordXl Guan USLNet is the first unsupervised model for sign language translation and generation, learning from single-modality data without parallel datasets. It combines singlemodality reconstruction and cross-modality back-translation, addressing text-video alignment issues using a sliding window method. Tested on BOBSL and OpenASL, it demonstrates competitive performance against supervised models. 14. Uncertainty-aware sign language video Ronglai retrieval with probability distribution Fangyun We modeling-WordXl UPRet introduces an uncertainty-aware probability distribution framework for sign language video retrieval, addressing the complexity and uncertainty in sign language. By enabling flexible mappings between video and text, it overcomes limitations of finegrained annotation scarcity. UPRet achieves state-of-the-art performance on How2Sign (59.1%), PHOENIX-2014T (72.0%), and CSL-Daily (78.4%). 15. Signer Diversity-driven Data Honcho Fu, Guan 2024 Augmentation for Signer-Independent Che Sign Language-Agvrv Zuo, 2024 The Signer Diversity-driven Data Augmentation (SDDA) method enhances signerindependent SLT by enriching signer diversity without using costly signer identity labels. Combining adversarial training and diffusion-based augmentation, SDDA generates diverse signer examples. Consistency and discrimination losses further improve signer-independent feature learning, achieving state-of-the-art SLT 6 performance in signer-independent settings. PROBLEM STATEMENT The aim of this project is to develop a system using machine learning techniques like a hybrid CNN-LSTM model which uses hand tracking to enable real-time sign language recognition and translation into text, improving communication for individuals with hearing impairments and enhancing social inclusion. Additionally, it incorporates grammar correction to ensure the accuracy of the translated sentences. This Project combines improves upon the literature survey as given below: CNN-LSTM model with attention mechanisms providing: 1. Better Temporal and Spatial Feature Extraction: CNNs handle spatial features, LSTMs manage temporal sequences, and attention. It selectively prioritizes critical frames 2. Enhanced Long-Range Dependency Handling: Attention focuses on key gestures without the need for exhaustive sequence processing like TCNs 3. Real-Time Efficiency: The combination reduces computational complexity, making it more suitable for real-time translation tasks compared to the heavier TCN approach. In summary, by integrating LSTMs and attention, this project effectively addresses the vanishing gradient issue while maintaining computational efficiency and accuracy, outperforming TCN-based models in both performance and speed. 4. Enhanced Frame Prioritization: The attention mechanism improves recognition of critical frames, something bidirectional LSTMs lack by default. 7 8 Advantages & Disadvantages Advantages Disadvantages 1. Customizable Data Collection: Users can create personalized sign language datasets through a data collection process, enhancing the model's adaptability to various sign languages. 2. Real-Time Predictions: The system offers real-time hand gesture recognition using hand landmarks, facilitating immediate translation of sign language into text. 3. Accurate Hand Tracking: The use of MediaPipe's Holistic pipeline ensures precise hand tracking, which is crucial for effective sign language recognition. 4. Grammar Correction: Integration with the language tool python library allows for grammar correction of predicted sentences, improving the readability and accuracy of translations 1. Data Collection Requirement: Users need to collect and label their own datasets, which can be timeconsuming and may require substantial effort to achieve a comprehensive dataset. 2. Potential Performance Variability: The accuracy of the model is highly dependent on the quality and diversity of the training data, insufficient data can lead to inefficient performance. 3. Technical Proficiency Needed: Setting up and training the model may require a certain level of technical expertise in machine learning and programming, which could be a barrier for some users. 9 OBJECTIVES A. Spatial-Temporal Gesture Analysis: Building a deep learning framework using CNN for capturing spatial features of hand shapes and LSTM for tracking temporal sequences of gestures to accurately map sign language movements. B. Gesture Recognition: Developing a system that performs accurately even with variations like different lighting conditions, hand positions, or gesture speeds by enhancing the adaptability of the neural network architecture. C. Grammar-Aware Text Generation: Integrating Natural Language Processing (NLP) tools to ensure the translated text maintains proper sentence structure and grammar, making the output easily understandable. D. Real-Time Hand Tracking: Using MediaPipe Holistic for precise hand landmark detection and motion tracking to ensure quick and accurate gesture input processing. E. Data Enhancement Techniques: Applying techniques such as flipping, scaling, and frame shuffling to generate diverse training data, improving the model's ability to recognize different types of gestures. F. 10 Methodology The block diagram below depicts the methodology for sign language translation: Fig 1:Block diagram depicting the methodology 11 1. Create a Dataset • This step involves collecting and annotating gesture sequences to build a labeled dataset. • Tools like MediaPipe Holistic are used to extract hand landmarks and store the data in a structured format. • This data serves as the foundation for training and testing the model. 2. Spatial Feature Extraction Using CNN • The CNN layers process image data to extract spatial features such as hand shapes, finger orientations, and movement patterns. • These features are critical for identifying the static aspects of gestures. 3. Temporal Dependencies Are Captured Using LSTM • LSTM layers are used to analyze the temporal relationships between frames in a gesture sequence. • This step captures how gestures evolve over time, allowing for interpretation of dynamic gestures. 4. Capturing Hand Gesture Sequences • This step involves recording a series of gesture frames over time to form complete sequences. • These sequences are essential for understanding the context and meaning of the gestures. 5. Feature Extraction • The extracted features (both spatial from CNN and temporal from LSTM) are combined and refined. • This integrated feature set is used to classify gestures accurately. 6. Sentence Generation • Recognized gestures are mapped to corresponding words, which are then structured into sentences. • This involves grouping predictions into meaningful outputs. 7. Grammar Correction • The generated sentences are passed through a grammar correction module (e.g., using language_tool_python). • This step ensures the final output is grammatically correct, enhancing readability and usability. 12 APPLICATIONS 1. Helps individuals who use sign language communicate with people who do not understand it by translating sign language into text or speech in real time 2. Can be used in classrooms or workshops to aid in communication between teachers and hearing-impaired students. 3. Students learning sign language can use the system to validate their signs and improve accuracy. 4. Enhances communication between healthcare providers and patients with hearing impairments. 5. Collects data to analyze different sign languages and their variations, contributing to linguistic and AI research. 6. When combined with speech-to-text or text-to-speech tools, this system can facilitate communication across different spoken languages, breaking barriers for sign language users. 13 EXPECTED RESULTS Gesture Recognition:The system will effectively classify various hand gestures with high precision, supporting dynamic and complex sign language recognition. Feature Extraction:The CNN will capture hand shapes and positions, while LSTM models will learn gesture sequences over time, enhanced by attention mechanisms focusing on key frames. Real-Time Processing:Integration with MediaPipe Holistic allows real-time hand tracking and low-latency processing for smooth user interaction. Text Translation: Recognized gestures will be translated into grammatically correct text using automated refinement tools. 14 REFERENCES 1 Kumar, P., Verma, A., & Raj, S. (2023). Gesture Recognition in Sign Language Using CNN and Bi-directional LSTM. IEEE Transactions on Pattern Analysis, 45(5), 1568-1580. 2 Yang, T., Zhang, H., & Li, Q. (2023). Gesture Translation in Sign Language with Deep Reinforcement Learning. AI & Robotics, 29(4), 220-234. 3 Zhang, W., Liu, B., & Wang, T. (2023). Hand Gesture Recognition for Sign Language Translation Using CNN and Attention Mechanism. Journal of Vision and Graphics, 27(2), 60-74. 4 Zhou, Y., & Chen, H. (2023). Efficient Hand Gesture Recognition for Sign Language Translation Using Transformer Networks. Computer Vision and Image Understanding, 214, 1-12. 5 Li, M., & Zhao, X. (2023). Sign Language Recognition Using 3D Convolutional Neural Networks. IEEE Transactions on Image Processing, 32(5), 1356-1367. 6 Wang, J., & Hu, L. (2023). Cross-Domain Sign Language Translation Using Generative Adversarial Networks. IEEE Transactions on Cybernetics, 53(7), 4873-4884. 7 Zhang, S., & Liu, Z. (2023). A Comprehensive Survey on Sign Language Recognition and Translation: Methods, Applications, and Future Directions. IEEE Access, 11, 24560-24577. 8 Singh, A., & Gupta, N. (2022). Deep Learning-Based Gesture Recognition for Sign Language Translation. Journal of Machine Learning Research, 23(4), 123-135. 9 Kumar, R., & Sharma, P. (2022). Multi-Modal Sign Language Translation Using Deep Learning. IEEE Access, 10, 12578-12589. 10 Kumar, A., & Sharma, R. (2022). Vision-based and Wearable-based Sign Language Recognition: A Comparative Study. Pattern Recognition, 132, 107830. 11 Soni, R., Jain, P., & Gupta, S. (2022). Real-Time Sign Language Translation with Machine Learning. Journal of Real-Time Systems, 58(3), 98-112. 12 Li, F., & Liu, P. (2021). Hybrid CNN-LSTM Models for Gesture Recognition in Sign Language. International Journal of Machine Learning, 43(3), 71-82. 13 Lee, S., Choi, J., & Kim, D. (2021). End-to-End Sign Language Recognition Using Deep Learning. IEEE Transactions on Neural Networks, 32(10), 2053-2067. 14 Zhang, H., Li, Y., & Wei, Z. (2021). Sign Language Recognition with Attention-Based CNN-LSTM Model. Pattern Recognition Letters, 145, 43-52. 15 Liu, J., & Wang, X. (2021). Sign Language Recognition Using Temporal Convolutional Networks. IEEE Transactions on Signal Processing, 69, 15 2642-2653. THANK YOU 16
0
You can add this document to your study collection(s)
Sign in Available only to authorized usersYou can add this document to your saved list
Sign in Available only to authorized users(For complaints, use another form )