2D OBJECT RECOGNITION IN GRAY SCALE IMAGES

Abstract

This thesis contains a survey of techniques and problems related to pattern recognition in computer vision. It also includes some experimental results and describes an improvement to an existing recognition method. In addition, a new feature extraction technique is proposed. The survey comprised discussions on digital pattern recognition techniques, optical pattern recognition techniques, and the use of neural networks. The survey on digital techniques was biased towards Automatic Target Recognition (ATR). For the optical techniques, both feature extraction approach and correlation approach were discussed. The Synthetic Discriminant Function (SDF) filter was implemented as a result of the survey. A study was carried out to investigate the performance of SDF filters when a Difference of Gaussian (DOG) prefiltering was applied. The study showed that the DOG prefiltering improved the output correlation peak-to-sidelobe ratio when the background was of low-variation texture and when the noise was not more severe than some statistics for the images used, which were taken from model jeeps, trucks and tanks. In addition, a method to reduce the number of training images was also proposed and successfully implemented. It used a recurrent neural network to approximate the orientation of an input object before rotating and correlating with the SDF filter that was required to be synthesized only for a limited range of orientations. A survey of neural network techniques in computer vision was also done. The neurobiology of a human vision system was documented. In addition, the role of neural networks in pattern recognition applications was also studied with selected examples. In this project, neural networks were used in the experiments carried out. The survey done in the initial phase of the project led to the conclusion that no method was best for all problems. Therefore, in the assessment of the efficiency of algorithms, criteria had to he set. Simplicity and potential for parallel implementation were two of them. It was with this philosophy that a position, scale and rotation invariant feature extraction technique was proposed. The feature vector is extracted by observing the correlation values obtained as one part of the object is rotated with respect to another. The method is inherently position and rotation invariant. Although it is not strictly scale invariant, it can he shown that in real life situations, the scale invariant property holds. The test images used was taken from model jeeps, trucks, tanks and aircrafts under strong and non-uniform lighting that resulted in inconsistent distribution and saturation of intensity values. A multilayer neural network was used as a classifier to resolve the problems caused by the lighting. The accuracy of the system was above 90% for the images used.