Image Matching Algorithm Based On The Local Invariant Features

Local invariant features are distinctive, robust to various image transformations, such as geometry transformations and illumination changes, and they have low redundancy and do not require the segmentation before the image matching. Because of these characteristics, local invariant features have proven to be very successful in applications such as image matching, object recognition, image classification and image retrieval, and so on. Using local features in image matching can convert multifarious image matching problems to the feature vector measurement, so as to improve the efficiency and robustness of algorithms. These approaches usually first detect the feature point sets, then compute the descriptors for these features making use of the point sets and their local neighborhood and at last measure the similarity of feature vectors to complete the image matching problem. However, the test images may exist geometry transformations and illumination changes, or have a blur background and partial occlusion,therefore,It is still a long and arduous task to lucubrate the local invariant features technology.In the problem of image matching based on local features, local feature detection and description are very important. The performance of feature detectors and descriptors directly determine the efficiency and precision of the result. This thesis concludes and studies previous achievements in local features, elaborates the general process of image matching methods based on the local invariant technology. Some advanced research and analysis about methods used to extract and describe the feature also are analyzed here. Additionally, two novel local descriptors are presented and be made use of image matching problem. The main contribution of this paper are:Firstly, the thesis elaborates the general process of image matching methods based on the local invariant technology, describes and analyzes the principle and performance of some classic detectors and descriptors, then evaluates their performance to provide a theoretical basis to select the appropriate detector and descriptor for the matching algorithm.Secondly, due to the limitation of image match introduced by one kind of feature, a feature descriptor using gray and edge information is proposed. The descriptor combines geometrical features with gray information. It first calculates the contrast context histogram of each feature point, and edge information is introduced to enhance the edge matching. The experimental results indicate that the results of the proposed algorithm are better than those of other methods that only use the gray information or edge information, and reduces the rate of mismatch in a certain extent, at the same time satisfies the invariance of rotation and illumination and are not sensitive to gray change and noise with a good match performance.Thirdly, because some feature extraction methods based on the LBP highly depend on the location of the pixels, an image matching algorithm based on the Local Sequence Patterns is proposed in this section. Feature points are detected by the Sift detector in the beginning, then a method combining the techniques of extended LBP and histogram is used to describe the feature points, generating the feature vectors. Finally these feature vectors are matched by using the Euclidean distance. This algorithm is effective with the invariant attributes of brightness, rotation, compression and acquires good performance on viewing transformation in some degree. In addition, it is robust to noise to a certain degree. The experimental results on the standard image databases demonstrate the effectiveness of our approach with a good practical value.