Contour Detection Via Multi-scale Spatial Inhibition And Contextual Modulation Based On Biological Visual Perception

Contour detection from complex scenes is one of the most important and challenging tasks in computer vision. Although a variety of contour detection algorithms have been developed in recent years, most of them cannot detect contour completely and accurately as human visual system which is good at accomplishing such task. Hubel and Wiesel have found that the majority of neurons in primary visual cortex are extremely sensitive to lines or edges with a certain orientation at a given position in the classical receptive field(CRF). However, the response of a neuron is not only related to its CRF, but also modulated by the neurons located in the non-classical receptive field(NCRF) that surrounds it. The existing studies show that the modulation of NCRF includes both inhibition and facilitation, and their interaction is crucial for eliminating background noises.In this paper, based on the perception mechanism of primary visual cortex(V1) and inhibition between neurons with the same and different frequency sensitivity, a contour detection model which contains multi-scale spatial inhibition and contextual modulation is proposed. At first, responses of V1 cells to contours and orientations are simulated by multi-scale Gabor filters, and multi-scale elementary contours are obtained by filter vectors. Then, a three-dimensional(3D) DOG filter extended from its 2D counterpart is used to convolute with the multi-scale elementary contours to accomplish inter-scale and intra-scale inhibition. After that, the inhibition results and orientations of multi-scale are combined respectively to obtain integrated contours and integrated orientations. Finally, contextual modulation based on both inhibition and facilitation is performed on the nonlinear mapping of integrated contours, and collinear facilitation and texture inhibition are realized by constructing distance and orientation weighting function, and then the integrated contours are further improved. Experiments on natural images demonstrate that the proposed model can improve the overall performance effectively in terms of both noise inhibition and accurate contour detection compared with other methods.Compared with previous studies, the contributions in this thesis are as follows:(1) The sum of filter vectors is used to obtain more detailed orientation information, which is beneficial to computing orientation difference in contextual modulation.(2) A three-dimensional(3D) DOG filter is constructed to simulate both inter-scale and intra-scale spatial inhibitions simultaneously, which can suppress background noise more effectively than only processing in single scale.(3) Combined with nonlinear mapping of Weber-Fechner Law, the contextual modulation, which can realize collinear facilitation and texture inhibition, takes both distance and orientation into consideration is introduced to improve the contour detection performance.