Object Detection Based On Salient Contour Feature

With the development of computer science and technology,image processing technology and computer vision have been widely applied in daily life,industrial and military fields.As an important part of computer vision,object detection has always been a hotspot and difficult point in the research.Cognitive psychology experts found that human vision is more sensitive to contour compared with color,texture or other features.In addition,since contour feature is more stable and robust,the contour-based object detection has received wide attention.However,due to the complexity of the application scenario,the uncertainty of the object shape and the occlusion between the objects and the background,it is still difficult to detect object accurately.Specifically,with the financial support of the National Nature Science Foundation of China(No.51405320)and Suzhou Bureau of Science and technology(Natural Science Foundation of Suzhou,No.SYG201511),image smoothing,salient contour extraction and object detection are studied.The main contents are listed as follows.Firstly,one structure-preserving smoothing preprocessing method is proposed.The method consists of several steps.In first step,a scale-aware approach is used to generate a guidance image by introducing domain filter and bilateral filter into rolling guidance framework.Subsequently,constructing a sparse L0 gradient minimization model using input image and the guidance image as double data fidelity terms.Next,introducing two auxiliary variables into the model for solving easily.Finally,obtaining smoothed image using alternatively minimizing algorithm.The method effectively removes the details and noise on the premise of preserving the image structure as much as possible so creates good conditions for further contour extraction.Secondly,two salient contour extraction methods are proposed.The first method utilizes traditional algorithm,such as Canny detector,to extract contours from smoothed image obtained by the proposed structure-preserving algorithm.The second method uses g Pb algorithm as contour detector,then extracts contours by combing Otsu and dual threshold method.Compared with the first method,the second method takes more features in considering and has better performance in denosing,but has higher computation cost.Thirdly,since fan shape model is weak to represent the severe deformed part in object,a refined fan shape model is proposed.The model improves the robustness by detecting and removing unstable contour point.It can be verified that the model not only reduces the computation cost,but also improves the efficiency of subsequent contour matching.Finally,several contour matching methods are studied.Combined with the probability density function of multiple features,a matching function is constructed to locate the object.As the object turns or rotates,the angle of the contour points and their edge directing still maintain a certain constant relationship with the original values.Three similar contour matching functions,which share the same object shape model,are proposed to improve the universality of the algorithm.The three strategies locate object,mirror of object and rotated object,respectively.The validity of the strategies are proved by experimenting on ETHZ shape classes and INRIA horses datasets.In this study,the object detection in natural images based on contour features involves image smoothing preprocessing,contour extraction,model training and contour matching.A complete set of procedures from the image preprocessing to locate the object contour is presented.Moreover,the feasibility of the algorithms are verified by several experiments on the different image datasets.