Research On Local Invariant Features Based Class Specific Hyper Graphs Learning And Object Recognition

Image based object recognition is an active research domain in computer vision. Although some successful progress has been achieved over a number of decades, object recognition is still a challenging problem because of the complexity of object imaging. This thesis derives a canonical model of visual category or class which is accommodated to significant or complex variations in imaging conditions from the view of large variations in variant conditions of an object. Techniques of image representation, object modeling and training, and object recognition have been thoroughly studied. The main work of the dissertation includes the following aspects.(1) Attributed graph representation model and a similarity measure between two graphs are proposed. On the basis of selection of salient local invariant features, images are represented with attributed graphs by integrating the local information contained in those local features and their spatial information. Because the attributed graphs contain not only the local feature information, but also the spatial relationship among the features, the recognition performance of image models constructed using attributed graphs will be enhanced. After representation of images with attributed graphs, a feature matching algorithm that utilizing the global and local information comprehensively is proposed. Given the matching results, a similarity measure between graphs is constructed to be used in image modeling and object recognition.(2) A method for the construction of the class specific hyper graph (CSHG) and its efficient training method are proposed. The CSHG model, whose vertex are attributed graphs and hyper-edges represent the similarity information of attributed graphs, is constructed from a large corpus of multi-view images represented with attributed graphs. In the process of constructing a CSHG model, an initial CSHG model is firstly constructed, then a similarity propagation based graph clustering method is used to obtain class specific familiy tree of graphs (FTOG), all class specific FTOGs make up a CSHG model. A CSHG model is a comprehensive integration of the global and local information contained in those local features and can accommodate to significant or complex variations in imaging conditions. Training efficiency of the model is one of the most important problems. Borrowing the RSOM tree method, an efficient training algorithm of CSHG model is proposed, which makes it more feasible to object recognition in practice. In addition, the indexing method in RSOM tree is also used in later recognition algorithm of CSHG model.(3) Methods of optimization training, incremental training and weak supervised training of a CSHG model have been studied. The optimization training of a CSHG model includes two stages. In the first stage, an optimized CSHG model structure is obtained using the entropy function defined on the CSHG model. In the second stage, redundant graphs in FTOG clusters in the CSHG model are detected and omitted using an affinity propagation method. In the end, the optimized CSHG model is constructed. In incremental settings, new images will continue to be obtained after a CSHG model has been trained. It is necessary to incrementally train a CSHG model. An incremental training algorithm is proposed based on an RSOM tree. In the process of incremental training, the category attribution of newly-come images might be omitted or un-annotated. We also propose a weaklly supervised training algorithm, which can automatically detected new class specific FTOGs. In which case, the training system will actively ask for attribution annotation for such FTOGs. This process can be termed as weak supervision or half supervion. Incremental training and weak supervised training are promising properties of our CSHG model.(4) The techniques for object recognition and object annotation are researched. Based on the trained CSHG model, object recognition methods for images with simple or complex background clutters and challenging viewing conditions are proposed. An object region labeling algorithm is proposed. A design schema of a CSHG model based recognition system is proposed in Chapter VI.