Image Local Invariant Feature Extraction Algorithm

Recently, using computer to simulate the human visual perception has become possible. But there are still a number of difficulties, one of them is how to accurately recognize a target under different environments (such as viewpoint transformations, scale changes, illumination changes, occlusion and so on). Image features which describe the target are the basic, so extracting invariant features from the image has become a priority. Because of its good locality, stability, invariance, image local features gradually become the main direction of the image feature extraction techniques.In this paper, we mainly study how to extract stable region features from different images of the same scene. We started from analyzing the maximally stable extremal regions, and then researched how to further improve the invariance of the region features respectively from the scale-space theory, image color information and image edge information.Firstly, we took advantage of the scale information to improve the region features performance. We extended the MSER detector from single scale to multi-scale by detecting region features at every Gaussian scale levels. Moreover, we integrated automatic scale selection mechanism to deal with the existing of duplicate features. We defined a scale select function for every extremal region. By searching for the local maximal of this function, we selected the interesting scale level of the image local structure.Secondly, we introduced two methods of integrating image color information into the MSER. The first one was multi-color channel MSER. We discussed how to deal with the duplicate region features detected from the multi-color channel MSER, and then we proposed an new way to select the duplicate region features. The other method was MSCR which detecting stable regions by agglomerative clustering of image pixels based on proximity and similarity in color space. We showed how the color space and similarity measure affect the performance of MSCR algorithm. At last, we developed an edge-enhanced MSCR detector with the stable edge information to improve the invariance of image blur. Assigned edge probability intensity for every image pixel by mPb edge detector, and then extract the pixels which have greater edge probability intensity as the edge information. It had strong robustness to image blur. We utilized the edge information by strengthening the differences between edge pixel and region interior pixel. It would improve the region features’invariance to image blur.