Research On Image Recognition Based On Object Contour Feature

In recent decades, information interaction has been eventually transferred from conventional text to image expression. More and more image recognition techniques have been applied to daily life, military activities and industrial manufacturing. According to image features, there are texture based, color based and contour based image recognition methods. As cognitive psychology experts have claimed that the human vision is more sensitive to contour feature than color and texture features, the contour-based image recognition has attracted increasing attention. This dissertation, therefore, conducts the research about image recognition based on contour. Specifically, the de-noising, simplification, description and matching of image contour, are studied. The main contents are listed as follows.Firstly, smoothing image preprocessing algorithm and its application in edge preserving and denoising of image contour is studied. In this section, a novel image smoothing algorithm using the sparse feature of pixel intensity and gradient as the dual-constraints is proposed to achieve better edge-preserving effect. Firstly, a pixel intensity and gradient function based on L0 norm is defined as a constraint term of the smoothing model. Then, two auxiliary variables are introduced using half-quadratic splitting strategy to construct the final smoothing models. Finally, alternating minimization algorithm is taken to solve the model. Smoothing experiments on natural images show that our algorithm can better meet the requirements of the edge-preserving and denoising.Secondly, a contour simplified algorithm called new discrete curve evolution is studied. To solve the problem of contour noise and deformation, a new contour simplified method based on discrete curve evolution is proposed. Firstly, two threshold functions of evolution control are defined to improve discrete curve evolution. After N-DCE process, an important visual part composed of only hundreds of feature points is obtained, which will simplify the complexity of the contour description.Thirdly, a contour decription method based on common base triangle area is proposed. A common base triangle area descriptor of each contour point is defined based on the area functions of the triangles formed by the other contour points and its two neighbor points. Then the descriptor is local smoothed to keep more compact and robust. Then, a cost matrix is obtained by computing the common base triangle area descriptors of all the contour points on two contours. Finally, the distance between two contours is measured based on the cost matrix by DP algorithm. The experimental results indicate that this method is robust to the contour deformation, and the computational efficiency and the retrieval accuracy are all essentially improved.Fourthly, chord angle representation method for contour matching under occlusion is studied. In order to solve the problem of shape matching under partial occlusion, a novel method based on chord angle descriptor is proposed. Firstly, a chord angle descriptor is defined based on the angle between two chords for each contour point. Secondly, a match cost matrix is constructed by computing the L1 distance between descriptors of all the points on two open contours. Finally, the similarity between two contours is obtained by integral image algorithm and partial shape matching result is achieved. The experimental results on MPEG-7 and Kimia216 shape databases indicate that this method is robust to the partial occlusion, and the computational efficiency and the retrieval accuracy are both essentially improved.Finally, the proposed methods are used in the application of salient object recognition in natural image to verify their feasibility. This recognition processes system is as follow: Firstly, natural image is smoothed to remove background noise texture. Secondly the smoothed image is binarized and segmented. Thirdly a contour extraction is processed on the binary image and the coutour is then DCE simplified to obtain better contour. Finally, the contour feature is represented by CBTA descriptor and matching by DP in the templated image database. The experimental results indicate that the performance of the presented recognition system is efficient.