Research on Robust Salient Object Extraction in Image a Dissertation Submitted to the Graduate School of Information, Production and Systems of Waseda University in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Object extraction is an essential branch of the computer vision, concerned with the analysis of images to extract useful information from the world. It covers wide range of image processing fields and plays an important role in computer vision. In this dissertation, contour detection is mainly discussed to realize object extraction, realized in three steps: edge detection, contour segment detection and object contour detection. In addition, an application of robust salient contour detection to 3D building reconstruction in aerial image is also discussed. Due to the complex scene of the natural image, it is necessary for us to develop robust contour detection methods against the noise and texture. And the computing complexity of the robust methods should also be considered. For the problem of salient edge detection, a multi-scale transition region based edge detector is proposed. According to existence of the " transition region " in natural image, a novel gradient computing method based on the transition region is designed. The central pixels of the transition region are prominent with low gradient magnitude while their neighborhood pixels are with high gradient magnitude. Through the multi-scale analysis by a series of the windows with variable sizes, we can detect the real edge pixels based on the relatively stable gradient distribution pattern of the transition region across the scales. Experimental results show that the proposed method performs better than Canny edge detector and an eigen-analysis based color edge detector with more connected edges. For the problem of salient contour segment detection, a scale invariance based line detector is proposed. A refined Nevatia-Babu line operator is firstly proposed based on the consecutive property of edge pixels on a line with simpler computing. Then a scale invariance based line detector is proposed based on the scale invariance property of the salient line segments on the object contour. According to the experimental results, the proposed method has almost no redundancy of the detected line segments compared with high redundancy level of the Progressive Probabilistic Hough Transform (PPHT) method-2-on the same experimental images. The proposed method also works more efficiently than PPHT methods, especially when the image includes many line segments. For the problem of salient object contour detection, a Constrained Delaunay Triangulation (CDT) graph based contour detection method is proposed. A structure called the Constrained Delaunay Triangulation (CDT) graph is introduced in order to enormously reduce the computing complexity. All the CDT edges are classified …