Feature Extraction Based On Biological Visual Cognition And Its Applications On Computer Vision

Feature extraction is an indispensable and important part of the computer vision systems, which plays an important role in a variety of visual applications. The merits of feature extraction will directly determine the system’s performance. Although many feature extraction algorithms and recognition models have been proposed, there are still a considerable performance gap between the current computer vision systems and human vision system. On the other hand, studies on biological visual systems have made great progress in recent years. The researchers have made a preliminary understanding of the processing paths of visual information, functional features of visual cortical areas and some important perception mechanisms by exploring the primate’s visual cortex area of the brain in-depth. These results provide a new way for constructing computer vision algorithms which simulate the human visual perception.Based on the biological vision theory, this dissertation focused on developing a good way for processing visual information, that is, the way for feature extraction. Through the study of information processing mechanisms and important functions in visual cortex, we simulated the different vision processes in visual cortex and then proposed feature extraction model based on the characteristics of complex cells, feature pooling mechanism by simulating the convergence of cell’s response, bio-inspired local feature descriptors and the model of color feature extraction. These models have been used on many practical tasks. The main work of this dissertation can be summarized as follows:Firstly, we proposed a feature extraction model which combining the topological encoding of image to analog the properties of complex cells. The model improved the filter groups learned by independent subspace analysis from nature images, and extracted filter banks which have obvious orientation and topology, and can simulate the properties of complex cells in a simple and efficient way. We further pool the filter’s response by the cooperating cortical pooling operation, and extracted features with certain invariance. The features were used for multi-class recognition and showed an obvious advantage.Secondly, we made in-depth study on the convergence way of cell response in visual cortex, and analysed the merits and defects of some current feature pooling strategies proposed in previous studies, then proposed a new feature pooling mechanism which can give a more reasonable explanation of biological experimental phenomena. The pooling operation cooperated with other method, and achieved invariance to the local transformations of input stimulus. We used the invariant features extracted by the proposed pooling operation for object recognition, and the experimental results have confirmed that it is more robust to the local transformations of an object.Then, different from the conventional descriptors which are extracted from some statistics or mathematical analysis in image content, the dissertation explored the way that the cerebral cortex characterizes the local area, and proposed a biologically inspired local descriptor (BILD), which mainly simulated the visual information processing mechanism in ventral stream of primate cerebral cortex. After combining the properties such as inhibition, enhancement, and response normalization between cells, our BILD extracted robust features for small local changes. By further integrating the orientation and spatial structure, the BILD achieved strong distinction. Compared to widely accepted SIFT and SURF descriptor, our descriptor has achieved better performance on image matching and object recognition, which shows the advantage of biologically inspired visual algorithms. The local descriptor can give a more detailed description of local area, and has a wide range of applications and research value.Finally, to solve the lackness of color information in most feature extraction methods, we proposed a color feature extraction model based on human color perception mechanism. Based on the research of Trichromatic theory and Opponent Color theory, we constructed the feature exteaction structure which contains the color information and the shape information simultaneous by simulating the color information processing path from pyramidal cells in retina to the opponent cells in visual cortex. We extended this feature to some existing visual algorithm, and validated its performance on several datasets. The results show that the proposed color feature can achieve better integration of color information, and effectively enhance the performance.