Research On Image Matching Based Sets Of Local Features And Its Application

As the basis of the area of Computer Vision (CV), image matching technology gives the approximation of correspondences between two images, it has been used in quite varied applications including:three-dimensional reconstruction, object categorization, pattern recognition, robot localization, and image retrieval. Compared with Image matching technology based on global features, Image matching technology based on sets of local features has better performance and therefore is more widely used. Based on the research on theories of local feature extracting algorithms, this paper proposes dimension partition pyramid matching algorithm (also called dimension partition pyramid matching kernel algorithm), it has state-of-the-art matching performance and very fast matching speed, and therefore can be used in above varied applications. This research mainly consists of the following parts:(1). In the view of recent related algorithms of local features, including invariant regions detectors and descriptors for those regions, gives a snapshot of the state of the art in detectors and descriptors, and compares their performance on a set of test images under varying scene type, imaging conditions; also establishes a benchmark against which future detectors can be assessed.(2). For the drawbacks of recent image matching technology based on sets of local features, proposes a new image matching algorithm called dimension partition pyramid matching kernel algorithm, its main idea is consistently dividing the feature space into two subspaces while generating several levels, In each subspace of the level, it maps local feature sets to multi-resolution histograms and computes a weighted histogram intersection in order to find implicit correspondences based on the finest resolution histogram cell where a matched pair first appears. Then a weighted sum of every subspace at each level is made. Compared with other related algorithms based on sets of local features, which need square or even cubic computational time, DP-PMK needs only linear computational time in the number of features, has no constraints on number or dimension of feature sets, and meanwhile overcomes the high-dimension performance decline problem caused by original PMK algorithm.(3). By model optimization, this paper applies proposed DP-PMK to area of image classification, gives an empirical study using multi-class SVM method. The entries of kernel matrix K consists of matching values between sets of features, the diagonal entries of K mean the self-correspondences. By introducing a base kernel which has kernel matrix diagonal significantly larger than the off-diagonal entries, K becomes a positive semi-define kernel, and therefore can be used as kernel function of SVM to solve classification problem. Experimental results on datasets Caltech-101 and ETH-80 show that:compared with other stat-of-the-art algorithms which need tens of times of original computational time, it takes only about 4-6 times of original computational time to obtain the same classification accuracy by using the method of DP-PMK.