Research On Image Irregular Region Matching Algorithm

Region matching is a key and classic issue in the field of computer vision, and it is popularly used in applications of 3D reconstruction, object recognition, image registration, and so on. Generally, typical region matching consists of three steps: Firstly, extract the feature region from two or more images respectively; then describe the feature region; finally match images by the corresponding relationship between feature regions.In recent years, the study of regional problems have obtained some breakthrough, but due to extract the feature region exist many factors, such as all kinds of irregular shape change and region exist repetitive texture, which will lead to irregular region hard to divide and descript. The extraction and matching of image region is still a hot research topic in computer vision, this paper makes a deep research for this problem, the main innovation of this paper gives the following content:(1) Since MSER regions have irregular shapes different with each other, it is hard to match these regions directly using popular descriptors. Usually, an extra fitting step is performed before descriptor construction of the irregular region, which introduces fitting errors and thus results in poor performance of the descriptor, especially when great viewpoint change happens. For the above problem, we present a novel matching algorithm based on intensity order. Firstly, we extract the irregular region by using maximally stable extremal regions(MSER). Then adopt a method which can avoid the ellipse fitting error to determine the support region, and we use gray level histogram statistical method excludes repetitive texture from the support region to express distinctive information. In addition, the support region is divided into sub-regions based on intensity order. Finally construct irregular region descriptor(IOMSRD) by using the mean-standard deviation algorithm.(2) For the problem of matching irregular region like MSER, a single support region exist some similar feature will increases image mismatch, this paper proposes the intensity order multi-support region descriptor, which pools many support regions construct descriptor based on the intensity order signal support region descriptor. Firstly, the method determines the multiple single support region of object region, and then divides each single support region into several sub-regions based on intensity order of the pixels, and constructs signal-support region descriptor, finally many single-support region descriptors compose the multi-support region descriptor. The experiment shows that the proposed method has better matching performance than the signal-support region descriptor. The experimental results show that the proposed algorithm has better matching performance than single support region descriptor in rotation, illumination change and viewpoint change image.(3) In this paper, the novel matching method called the distance transform irregular region descriptor(DTIRD) is proposed, which is based on distance transform for irregular region matching. The proposed algorithm Firstly give a division method based on distance transform, for the problem that each sub-region has different pixel points, we use mean standard deviation(MS) algorithm to construct descriptor. The experiment shows that the proposed method has better robust performance under the conditions of image rotation, illumination change, blur change and JPEG compression change.Experiments show that the proposed three kinds of irregular region descriptor(IOMSRD MRMSD DTIRD) have high matching accuracy and excellent matching performance, which can effectively solve the problem of irregular region match.