Image Matching Based On Local Feature Descriptors

The goal of image matching algorithm is to find homogeneous regions between the images, and then establish their spatial correspondence accordingly. It is not only a hot topic in the field of computer vision research, but also the foundation of many vision ap-plications. However, the appearance of images may be affected due to the light changes, the perspective changes, target morphological changes, image noise, etc., which may bring many challenges to the image matching task. Therefore, how to effectively analyze and measure the homogeneity of the image content, and then acquire image matching algorithm with high accuracy and robustness is the essential issue of image matching tasks.Image matching based on feature descriptors is an important research direction in this field. Firstly, feature points in the image are detected by using various types of detection algorithms. Then, based on the low-level image feature extraction method and the image de-scriptor construction method, the area around each feature point will be described to acquire the local image descriptors. After that, by using point-set matching algorithm, correspon-dence between the feature points can be further established based on the similarity and the spatial relationship of the descriptors. Finally, matching results between images can be ob-tained.In this dissertation, critical technologies of the image matching algorithm have been studied, including low-level feature extraction, local descriptor construction and point-set matching. The main contents and innovations can be summarized as follows:1. The technical characteristics of image matching algorithm were analyzed. Firstly, traditional low-level features are calculated based on the comparison of discrete pixels with-in the interest region. However, these pixel-wise feature computing methods are sensitive to noise, while not suitable for the description of relatively larger patterns in the image. Sec-ondly, the construction strategies for local image descriptor were studied. In order to have an effective characterization of the spatial structure within the image, the conventional algo-rithms usually divided the interest region into different blocks based on the spatial template, but these methods do not take into account the distribution of image content, thus lack of robustness. While during the point-set matching, although some state-of-the-art algorithms can handle image distortion properly, but they failed to take into account the local point-set structures when performing iteration.2. A circular shaped low-level feature structure was proposed. Firstly, we designed a low-level feature structure based on local circular sampling method, which makes it possible to overcome some drawbacks of the classic low-level image feature extraction algorithms. Then, based on this structure we proposed the feature extraction algorithm by using differ-ential circular coding and discrete cosine transform coding, and corresponding analysis was performed. The local organization and continuously changes of the pixels can be well char-acterized by this low-level feature, thus relatively larger local patterns in the image can be obtained, which can provide a better basis for further image description method and image matching task.3. A new local image descriptors construction algorithm based on visual receptive field was proposed. Inspired by the structure and function of the visual receptive field in human nervous system, we first proposed a low-level feature organizational structure. This feature pattern takes advantage of the basic characteristics of the visual receptive filed, and can integrate into the low-level features extraction algorithm proposed earlier. In order to address the problems of the pooling strategy mentioned above, we further proposed a new membership pooling strategy based on fuzzy C-means. Therefore, by dividing the interest region into sub-blocks with common visual semantics, the spatial information of each image region can be encoded more effectively. Thus, we can simulate the spatial association property of the visual receptive fields in the same framework.4. A new point-set matching algorithm was proposed. Firstly, in order to acquire a more comprehensive characterization of the topology in the point-set, we proposed a shape extraction algorithm based on local structure, including the coding of basic patterns and the shape pooling based on relative coordinate system. Acquired local point-set shape code can effectively integrate the local spatial information, thus has a good ability for the description of the points-set’s topological structure. During the calculation, the point-set is normalized, and a corresponding look-up table is built according to pre-defined template, which can improve the computational efficiency. Then, we further integrate the local structure extraction algorithm into the point-set matching framework based on Gaussian mixture model, and constrain the matching process accordingly. What’s more, we further acquire the local shape and update the weight of the Gaussian component according to the posterior probability. Finally, the data points and the Gaussian components can be matched by maximum the posterior probability of the model. Therefore, the correspondence between two point-sets can be found under the proposed framework.