Biologically Inspired Computer Vision/ Applications of Computational Models of Primate Visual Systems in Computer Vision and Image Processing

Biologically Inspired Computer VisionApplications of Computational Models of Primate Visual Systems in Computer Vision and Image Processing Reza Hojjaty Saeedy Abstract Biological vision systems are remarkable at extracting and analyzing the information that is essential for vital functional needs. They perform all these tasks with both high sensitivity and strong reliability. They can efficiently and quickly solve most of the difficult computa- tional problems that are still challenging for artificial systems, such as scene segmentation, 3D/depth perception, motion recognition, etc. So it is no surprise that biological vision systems have been a source of inspiration for computer vision problems. In this research, we aim to provide a computer vision task centric framework out of models primarily originating in biological vision studies. We try to address two specific tasks here: saliency detection and object classification. In both of these tasks we use features extracted from computational models of biological vision systems as a starting point for further processing. Saliency maps are 2D topographic maps that catch the most conspicuous regions of a scene, i.e. the pixels in an image that stand out against their neighboring pixels. So these maps can be thought of as representations of the human attention process and thus have a lot of applications in computer vision. We propose a cascade that combines two well- known computational models for perception of color and orientation in order to simulate the responses of the primary areas of the primate visual cortex. We use these responses as inputs to a spiking neural network(SNN) and finally the output of this SNN will serve as the input to our post-processing algorithm for saliency detection. Object classification/detection is the most studied task in computer vision and machine learning and it is interesting that while it looks trivial for humans it is a difficult problem for artificial systems. For this part of the thesis we also design a pipeline including feature extraction using biologically inspired systems, manifold learning for dimensionality reduction and self-organizing(vector quantization) neural network as a supervised method for prototype learning