OBJECT RECOGNITION USING NEURAL NETWORKS

Abstract

The main aim of this project is to investigate the problem of invariant 2D object recognition for input gray-level images which may or may not have textural content. There are two steps : (a) segmenting an object from the background by intensity thresholding or by textural boundary detection. (b) recognition of an object subjected to translation, scale change and rotation. Over the past few years, there has been a resurgence of interest in the field of neural networks. A neural network's capability to learn and to adapt, together with its inherent parallelism and robustness, have made it a natural choice for machine vision application. We have thus decided to pursue the neural network approach with the hope of achieving full neural network implementation. The entire system consists of two modules, i.e., the texture segmentation module (TSM) and the object recognition module (ORM). For the TSM, the outputs of Gabor filters are used as features to extract the object from the background. The family of Gabor elementary functions has been shown to be localized maximally both in spatial and spatial frequency domains. The competitive and co-operative mechanism of the Boundary Contour System (BCS) is utilized to smooth the feature maps produced. A binary image with segmented regions is produced and fed to the object recognition module. In the ORM, edge strengths of an input image from the previous module are extracted using oriented receptive fields. Using a selective attention mechanism the location of an input object is detected. This object is fed into a pyramid structure which normalizes its size. Rotated versions of this normalized object are generated and used in the matching process. The matching is done using a template which is a two-layer hybrid neural network. The performance of this system has been evaluated by software simulation. The input images include geometric shapes with textural content and also mechanical tools of various kinds. The results show that the TSM is capable of producing satisfactory segmentation which is then used by the ORM for the purpose of recognition. Besides the transformations mentioned above, the ORM has also demonstrated its tolerance to deformations and occlusions.