Development of a neural network based joint transform correlator for multi class distortion invarient pattern recognition

Abstract

This thesis is concerned with the application of neural network function to the classassociative target detection technique to get distortion invariant detection for the objects of different classes by using optical conelator. Here neural network function generates a synthesized reference image with the help of weighted sum of the training set containing the original image and the possible distorted versions of the original image and that is controlled by fringe-adjusted joint transform correlator (FJTC) function at the output layer. A class of objects may be defined as a group of objects with similarity or dissimilarity among them, As the target members of the class are known, the neural network based reference images can be generated in off line by computer and fed dircctly to thc dctection systems in real time. Classassociative target detection employs a multi-target detection algorithm (MTDA) facilitating multiple reference images for which more than one dissimilar objects or images of different class can be detected. In MTDA several combinations of input spectra are produced using each reference and input scene separately. For distortion invariant class-associative target recognition, all the conventional references are replaced by neural network along with fringe-adjusted JTC generated composite image, Thus employillg the generated composite image as the reference, the class associative detection scheme provides distortion invariant multi-class target identification. A simplified version of one step single layered neural network is proposed, which is found successful in the implementation. The conditions for generating neural network based composite image and employing into class associative fringe-adjusted filter have also been suggested. The simulation shows satisfactory performance or the proposed scheme in getting distortion invariant detection of the members of the classes mainly with scale variation, in-plane rotation, missing data and 3D distortion giving clear correlation peaks for all the distorted and undistorted images while no or negligible correlation peaks for the non-targets both in noisy and noise free conditions. This scheme works successfully for both binary and gray level images.