GROUP MEMBERS:
MUBASHIR SHAH
MAMOONA USMANI
November 2015
Conference: IEEE Asia-Pacific World
Congress on Computer Science and
Engineering 2015
At: Plantation Island, Fiji

Introduction

Conceptual Framework of Entire System

Feature Extraction

The ANN Classifier

Location of Extracted Image

Conclusion & Results

Demo

Object detection and recognition using vision sensors and
its implementation on SCORBOT ER-4U (robotic arm
platform) [9] which will be refurbished and utilized to sort
electronic components such as resistors and capacitors for
laboratory technicians robotic arm

Object detection (IP) Algorithms require huge processing
time for successful implementation

The technique includes feature extraction algorithm followed
by a classifier to recognize the objects under test

Once the object is finalized the result would be the type of
the object along with its coordinates
Work done by previous researchers include:
The work presented by T.P. Cabre et el, used of basic algorithms like
image enhancement, noise reduction and a visual loop algorithm (based
on trial and error approach) prior to applying image processing techniques

IP algorithms developed to reduce response time and increase in the
efficiency for the object recognition tasks.

Employing a parallel programming approach called object surface
reconstruction method, ten times faster.

Communication of vision system (webcam) via USB and through MATLAB,
the system is enabled to perceive environment via IP algorithms
 Usage of multi-stereo vision technique for the detection of 3D Object.
Eliminating the background/ least req information, using opening and
closing morphological techniques.
 Use of Viola-Jones IP method, one of the Robust Object detection
algorithm basically developed for face detection, using cascaded
architecture of the strong classifiers arranged in the order of complexity.
Along with the classification part, the concept of Feature Extraction (FE)
is also studied.
 Feature extraction via Contour matching , one of the best methods to
detect objects. The trained shape is matched according to a
probabilistically motivated distance measure which enhances the shape
comparisons within the framework. Also noise reduction and other
image optimization via segmentation and other IP techniques were
used.
 Trainarp algorithm (derived from ANN). It also presents the method to
migrate from the statistical approach to Artificial Neural Networks
(ANN). The author has stated the efficiency as 95% and response time
of 94ms.
Object detected, cropped and resized
according to classifier specification
Training and validation of classifier result in data labels
Required object along with its coordinates are computed
Inverse kinematics results in pick & place
Eliminating Hue & Saturation info while preserving intensity info
Intends to minimize intra-class variance of black&white pixels.
Pre-req for blob analysis.
Use of rectangular SE to enlarge detected edges
Pre-req for blob analysis, more clear & visible edges outlines
Rectangular bounding boxes around detected objects; cropped
& resized to 20x20 pixels & conversion to gray scale for classification
Binary Image
Dilation
Edge Detection
BLOB analysis




Acquiring 312 images with different objects
(bounding boxes around each)
Dividing this datasets to 70% training &
30% test sets
Labeling of training data
Use of gradient descend back-propagation
algorithm
◦ Decide ANN layers(3 layers; An input(400),a hidden(var)
& output layer(as per no. of outputs)
◦ To train classifier, initialize weights and perform feedforward & update weights by back-propagation (20000
iterations)

2 Models of classifier
1. Converts cropped image of bounding box to gray-sale and resize
as per classifier requirement 20 x 20 pixels
2. Cropped image is placed at center of white back ground; only the
dimension that exceeds 20 x 20 is resized ie in 35 x 18 only 35
will be resized

Cross-validation of Classifier- 5 fold method used on
both models
400 Input Layers
25 Hidden Layers
2 Output Layers


Blob analysis resulted in centroid points (xc,yc)-Known
Conversion factor to obtain cm
values from pixels
Xc,yc


Centroid values treated as
end effectors posn
Inverse Kinematics will be
used for pick and place
operation
Orignal Image
Grey Scale Image
Edge Detection
Dilation
BLOB Analysis
Binary Image
Image Filling
 Feature Extraction Accuracy The feature extraction algorithm was
tested using a separate set of test data and its accuracy was 83.6443%.
 Classifier Accuracy .
 Final Results is the table of 32 scenes which altogether has a total
of 448 objects inclusive of both capacitor and non-capacitor
images.
RESULTS FROM THE PAPER
 The accuracy yielded by feature extraction is
83.6443% and the classifier accuracy is
99.33% (upon Cross-validation)
 The best model classifier has an input layer
consists of 400 neurons, 25 hidden neurons
and 2 output neurons