Cycle Structure Feature In Image And Its Applications

Feature extraction which plays a key role in computer vision attracts more and more attention by researchers in recent years. Local features usually convert the complicated matching problem to the measure of feature vectors, thus can greatly reduce the complex-ity of the algorithm. This paper proposes a cycle structure feature which is invariant to translation, scaling and rotation and it can be used for image recognition and registration. The main work of this paper includes:1. Extraction and description of cycle structure feature. Combined with the concept of cycle in graph, the algorithm to extract and describe cycle structure feature in images is proposed in this paper. To overcome the dependence on the segmentation for registration or recognition, a multiwavelet kernels and multiscale hierarchical decomposition method is adopted to generate multiscale binary maps at varying image resolutions in different levels of details. The multiscale segmentation results are further processed for better feature extraction. Cycle detection algorithm based on Breadth First Search strategy is then proposed to detect the cycle structure feature composed of bifurcation points, intersection points and the connected lines. Finally, the feature is described as feature vector using normalized angles and lengths of the branches.2. Retinal image registration based on cycle structure feature. Matching two cycle structure features with the same number of feature points can be implemented by minimizing the similarity measure between feature vectors. It tries to find the most accurate matching cycle structure features which can be used for the parameter esti-mation of the following similarity transformation. VARIA database is used to eval-uate and compare the performance of our method for retinal image registration. In order to evaluate the results quantitatively, the Skeleton Alignment Error Measure (SAEM) is defined and the final registration result is 0.938 pixel.3. Scallop image identification based on cycle structure feature. The scallop image identification database is constructed by reference image database and identifying image database. The similarity measure is applied to match the feature vectors between the images from these two databases. Then the reference image, which has the minimum value of similarity measure with an identifying image, is treated as the detected scallop image. For the constructed scallop image identification database, the identification success rate reaches to 83.3%, which means this method is feasible and effective for identification purpose.The cycle structure feature consisting of bifurcation points, intersection points and the connected line is proposed, which is applied in the retinal image registration and scal-lop image identification in this paper. The proposed feature and method have vital signif-icance and can also be used for other applications such as identification and verification.