Boundary Detection Research Based On Stereo Vision

The automatic abstraction of image concept is a hot research topic in multimedia and computer vision. Within this field, how to conduct boundary detection is a decisive factor for concept abstraction.Boundary is different from what traditionally called edge. It defines the outline of a target in the image, indicates the target's difference from surrounding environment and demonstrates the change of pixel's ownership from one object to another. While edge just specifies the change of fundamental properties such as gray level, texture, brightness etc. It can be said that boundary is another form of edge, but from a higher level of concept.Traditional edge/boundary detection mostly employs single image. However, our algorithm starts with image pairs and lays emphasis on how to detect boundary from stereo image pairs. For this purpose, we take stereo image pairs with color difference as our research objects and do deep research on color correction, stereo matching and boundary detection. Finally a systematic scheme from image pairs to detected boundaries is presented.The main work and innovation point of this dissertation is as follows:(1) Image segmentation evaluation. Based on the difficulties of selecting a proper image segmentation algorithm in actual practice, we propose a subjective evaluation method. Firstly, participants are asked to score the segmentation results. Then the results are analyzed in a statistical way. In the end, algorithms' performance is compared from different aspects. In order to make our evaluation conclusion much more fair and convincing, we pay much attention to parameter selection. To be specific, we select 10 parameter combinations for each algorithm and the final evaluation procedure is conducted on the 10 representative parameter settings.(2) Color correction between stereo images. We propose a color correction algorithm based on image segmentation and keypoint matching to compensate the color discrepancy between stereo images. The main idea of our algorithm is to conduct color correction region by region. This avoids the contradiction that the overall compensation does not necessarily meet the need of sub-regions. Meanwhile, this algorithm does not need to employ standard color board or other complex equipments. That makes it easy to use in many applications.(3) Classifier design in pattern recognition. We propose a universal classifier design method based on piecewise linearization. Our method is not only an improvement of the minimax criterion, but also an approximation to Bayesian classifier. We apply this classifier design method to our boundary detection procedure and get a satisfying result.(4) Construction of image segmentation dataset. We construct a new dataset used for image segmentation and boundary detection evaluation. Different from existing datasets, this one is based on stereo image pairs. It contains both the ground-truth segmentation results and depth information. This makes it capable to be used as a new benchmark.(5) Boundary detection. We propose a boundary detection algorithm combing depth information and other low-level information. Different from traditional methods, we employ depth information from stereo images to fulfill the boundary detection task. This is a new idea for boundary detection.