Research On Automatic Curve Matching Technique Based On Texture Feature

Image feature matching plays an important role in many tasks of computer vision, and it has an extensive application in object detection and recognition, image registration, video understanding, 3D reconstruction and so on. Three basic steps of image feature matching are introduced: Firstly, extract image features from two or more images respectively. Then, describe theses extracted feature parameters. Lastly, match images by establishing the corresponding relationship between features.In the past years, feature matching technique based on descriptor has made great success. The basic idea of this method is to construct the description vector based on texture information around features: firstly, determine the curve support region(CSR) and partition CSR into several sub-regions; then, construct the descriptor by statistics the invariant features in each sub-region; finally, compute the similarity between the descriptors to complete the match. Although the point descriptors based on texture features have many and make a series of achievements, the curve descriptors based on texture features only make little progress. Since curves have different lengths, incorrect position of endpoints and contain lots of relative texture around neighbor, the research of curve matching based on texture feature is still a challenging topic. Therefore, this thesis gives the following content.(1) This paper objectively analyzes and summaries the existed curve matching algorithms. Emphasis on analysis of the most relevant algorithm—MSCD(mean standard deviation curve descriptor), and aim at the exits problem of MSCD descriptor, this paper proposes three curve descriptors based on texture features.(2) The intensity order mean standard deviation descriptor(IOMSD) is proposed. For MSCD like methods, the sub-regions partition based on fixed shape causes a boundary error, This paper presented the IOMSD descriptor, which is based on the intensity order partition and the mean standard deviation information. Under different viewpoints, image deformation distorts the region shape. However, image deformation hardly changes the permutation of intensities in a region. Therefore, IOMSD not only can overcome the boundary errors of image deformation, but also it is invariant to monotonic intensity changes.(3) Different from the traditional gradient or intensity feature, this paper introduces intensity order to represent the texture feature of the curve’s neighborhood, and develops a novel texture-based curve matching method called IOCD. IOCD descriptor is inherently rotation invariant, which is not only invariant to monotonic intensity changes, but also can handle more complex illumination changes.(4) Aim at the lost of relative texture around the curves’ neighbor, this paper proposed ternary contextualized histogram pattern(TCHP). Local ternary homogeneity patterns are firstly adopted to represent the texture information of curve neighbor by the method of histogram contextualization, which not only encodes spatial information into histogram, but also solves the dimension problem caused by high order histogram.Experiments show the proposed three descriptors(IOMSD, IOCD, and TCHP) perform robust and has high correct ratio, which can be competent for curve automatic matching.