Effective and efficient techniques for contour-based corner detection and description for image matching

Finding the correspondences between similar visual contents in two images is a basic computer vision problem for many applications such as image registration, image retrieval, 3D reconstruction and object recognition. Even when two images do contain similar contents, this problem is still challenging as the contents in the two images might differ in factors like scale, illumination, rotation and viewpoint. Three main steps in the local feature-based approach to finding such correspondences are (i) detecting the locations of the features, (ii) selecting the local region (interest region) to build the descriptors, and (iii) building the descriptor to represent each feature. As the performance of finding the correspondences using the local feature-based approach depends on each of the aforementioned three steps, we have proposed new techniques, which are more effective and efficient, for the three steps. For the first step, we have developed two effective and efficient contour-based corner detectors to detect the feature locations. The repeatability of the proposed corner detectors is better than the existing contour-based detectors. In addition, the proposed detectors are the most efficient. For the second step, we have proposed a new method to estimate an appropriate interest region around a corner in order to describe that corner. Our method therefore addresses the limitation of existing contour-based detectors not having any inherent information to define such an interest region. We have shown that a descriptor built using the region estimated by our method performs better in finding the corresponding corners between two images. For the last step, we have proposed a more efficient technique for building a descriptor. Besides being more efficient to build, the descriptor can still achieve comparable image matching effectiveness compared to existing descriptors. We have also demonstrated how this description technique is also applied in our interest region estimation method to achieve more efficiency in the second step. According to our experimental study, the combination of our proposed techniques also achieves the best overall performance compared to various combinations of the existing state-of-the-art techniques.