The Research Of Image Representation Based On Visual Perception And Its Application

The complex external environment information can be expressed by visual system with a small amount of neurons, the sparse response mechanism of which has a major impact on visual computing, image sparse representation, signal compression perception and other areas. Among them, the image parse representation simulated visual system can express the essential characteristic and internal links of images by the most concise way, and improve the processing efficiency and real-time performance of image relevant problem, which has attracted much public attention and became research hotpot.Based on the hypothesis of visual system is the result of long-term evolution in external environment, using natural images statistical properties as the breakthrough point, in accordance with visual neural physiological mechanism, the systematically research will be carried out about three core problems research of image sparse representation and its application:the sparse representation model, the extraction and screening of basis function set, its application in image processing. The main results are as follows:(1) The major visual neural mechanisms associated with sparse representation will be analyzed, for instance, receptive field properties, sparseness, correlation, over-complete and synchronous oscillation, etc. We research and dissect the inner relationship between visual neural mechanisms and natural image property, which offers the basis and foundation for solving the problem of image sparse representation based on visual perception.(2) Based on visual neural mechanisms and related theories of visual computing, simulating basis function extraction mechanism of visual complex neurons, we construct the sparse representation model of independent subspace analysis and topographic independent component analysis, and present related training algorithm of natural image basis function, and build the basis function complete set; then we construct the PCNN model simulated synchronous oscillation mechanism of visual cortex, and present basis function selection algorithm based on neuron response coefficient.(3) On the basis of aforementioned complete set model, according to the properties that visual system can increase the robustness of system through proper redundancy for reducing the noise and error sensitivity, we further make the feature dimensions extracted from natural image are greater than the dimensions of the image data, to form the over-complete set which has different orientation, phase and frequency. Aiming at the three restricted conditions of sparseness, correlation and quasi orthogonality bring from over-complete set, we present sparse representation model of OCTICA and the related training algorithm of basis function.(4) Based on the computing models of image sparse representation mentioned above, the key algorithm of covert-target recognition, moving-target tracking and image denoising will be presented, and the related experiments are designed for verifying the algorithm. The experiment results show that the above-mentioned algorithm can solve the corresponding problems, the accuracy rate, real-time in the process of recognition and the denoising effect have more preponderant than traditional methods.