Research On Shape Recognition Based On Shape Contour

With the popular usage of personal electronic devices and the improvement of device performance, users need to comprehend and handle a large number of images. Due to the continuous improvement of computer technology, people want to understand and handle the images or data with computers. In order to manage and make out these data effectively, the low-level characteristics of images (color, texture, shape, etc.) become the most important features when solving this problem. As a significant perceptible feature, shape becomes one of the key targets for researchers. Along with the increase of practical application (such as character recognition, trademark logo retrieval, activity recognition, human pose estimate) which associated with shape matching and recognition, more efficient shape matching and recognition algorithms are needed to handle these problems in researches. Thus, the study on shape matching and recognition is crucial among research area both in theory and practice.Since the contour based algorithms shows a good performance and has tremendous potential, it has been one of the most big-hit research topics in shape matching and recognition. However, there are still many challenges. This thesis concentrates upon the theory of shape matching. From the point of computer vision, the research gives an in-depth study on shape matching. To aim at the problem of the existing algorithms, this thesis completes the following works:(1) As for the contour based shape matching, it was almost using one to one correspondence schemes. One to one scheme shows a good performance in most algorithms, but still has some problems. With out considering the position attribution of the point on the contour, it is easy to mismatch in one-to-one based matching algorithms. In order to reduce the possibility of mismatch, the relationship between pair of points was incorporated into shape matching schemes. This thesis proposed a pairwise correspondence scheme to measure the similarity between shapes and used kernel function to denote the affinity of pairwise correspondence. For the purpose of improving recognition ratio, a new method was proposed to measure the distance between given shape and model according to the recognition characteristics. The shape to class distance effectively improved the recognition accuracy.(2) Polar coordinate system is a widely used two-dimensional coordinate system. Inspired by the polar model shape matrix and shape context, they were combined in this thesis to be a new descriptor called global shape context. From the point of human visual perception, angle gradient was used to simple shape and set to reduce the effects of discrete pixelization of the contour. Since the minor variations along the edge were ignored, the global shape context descriptor extracted from the simplified shape was more robust against noise.In order to verify the performance of the proposed scheme by applying them to shape matching and retrieval, the experiments are conducted on a common used standard test set of MPEG-7CE-Shape-1, which contains various shapes. The researcher measures the retrieval rate by the Bulls-eye test. Experimental results demonstrate that the proposed algorithms were robust to scaling and rotation. In the process of shape retrieval, the retrieval rate of pairwise correspondence shape matching algorithm is93.1%, beyond the shape context and inner-distance shape context. The retrieval rate of global shape context is85.1%, beyond the SC and proportion to the IDSC. In terms of running time, GSC’s extracting time is below the SC’s and IDSC’s.