Biologically Inspired Edge Detection And Its Applications On Automatic Target Recognition Systems

Edge detection is the most fundamental issue in the field of image processing, and also one of the classical technical problems. It has great impact on many high level applications, such as feature extraction and description, object recognition and image understanding, etc. However, due to the blurring and deformation of images caused by the projection, mixture, distortion and noise in imaging process, though a great number of edge detection algorithms have been proposed, they could not achieve satisfactory performance compared to human visual system either on the detection of edges itself or on other advanced visual tasks such as image matching and object recognition based on edge detection. On the other hand, much progress concerning researches of biological visual has been made recently. By studying neural physiology and psychology, researchers preliminarily acquired processing pathways of visual information, functional characteristics of different areas in visual cortex, and some cognition rules of human vision. All of these study results provide new guidance for the researches of edge detection.Based on theories of biological vision, in this thesis, we first simulates the perception characteristics and recognition methods of human vision to carry out researches on several basic issues such as edge detection, contour detection and contour grouping, and then study the applications of edge information in some relevant issues in automatic target recognition. The major contribution of this thesis is concluded as following:First, inspired by different receptive field properties and visual information flow paths of neurons, an edge detection algorithm based on receptive field properties is proposed. This algorithm simulates the response characteristics of LGN, simple cells with non-classical receptive field properties and complex cells to detect edges. By studying the formation mechanism of simple cell's receptive fields, a CORF model combined with non-classical receptive field properties is presented to simulate the responses of simple cell's receptive fields. Compared to the classical model, the proposed one is able to better imitate simple cell's physiologic structure with consideration of facilitation and suppression of non-classical receptive fields. Experimental results validate the robustness of the proposed algorithm to noise and background interference.Then, a contour detection algorithm based on multi-level visual clues is proposed. Different from traditional methods implemented on pixel levels, this method first produces contour candidates on the basis of superpixel segmentation. Guided by relevant process and theories of biological vision, multi-level visual clues are extracted for the evaluation of the above mentioned contour candidates where the final contours are selected. Experimental results show that this method is effective for salient contour detection of objects in images, and outstanding in balancing the performance and computational timings.Third, guided by Gestalt cognitive rules, a further study of multiple object contour grouping is carried out on the basis of contour detection according to images containing multiple objects. Taking results of contour detection as grouping elements, a multiple object contour grouping method based on Gestalt grouping constraint and spectral clustering is presented. The affinity matrix constructed by grouping constraint is designed to automatically estimate the number of objects, and the grouping of different object contour is achieved by spectral clustering. From the results of experiments, this method is proved to be effective on object number estimation and contour grouping of different objects.Finally, aiming at the objective area selection problem in ATR applications, a selection method of candidate objective area based on edge orientation distribution is proposed. Inspired by pooling function of complex cells for invariant feature extraction, this method constructs a feature to represent edge orientation distributions. According to the defects of traditional saliency methods applied in remote sensing images, a saliency map of edge orientation distribution is proposed to select candidate areas. Tests on multiple remote sensing images validate the effectiveness of the proposed method.