Structure Extracting and Matching Based on Similarity-Pictorial Structure Model for Microscopic Images

An improved pictorial structure method (PS) is presented in this paper for structure detecting and matching in microscopic images of the micro structure fabricated using micro-fabrication technology. The obtained structure information of the micro objects is important for the following 3D visual calculation process. Although PS method can extract and match features successfully by evaluating the appearance and shape under slightly varying illumination, while the microscopic images are taken under considerable varying illumination, appearance description of traditional PS model performs poor for our application. In this paper, self-similarity descriptors were used for features extracting and appearance describing for images of micro objects under large changed illumination. The special local features on the micro objects ensure the feasibility of self-similarity method. In the proposed method, the self-similarity descriptors of patches which reflect the information of local object structure are extracted. At the same time, the geometric relationship among patches is modeled by PS. By using our similarity-pictorial structure method (combination of local self-similarity descriptors and pictorial structure model), the structure of object can be extracted and matched exactly. Finally, contrast experiments between PS method and our method for structure extracting and matching on object structure in microscopic images of MEMS components under large changed illumination are presented. Experiment results show the effectiveness of our proposed method.