Image Recognition Based On Graph Statistical Feature

With the development of advanced technology, images record more and more information of people’s life and their work. Therefore, image recognition draws more attention to researchers than ever before in pattern recognition. Image recognition methods based on the graph feature as a series of useful methods, has got significant progress due to researchers’ hard work. Focusing on the problem, that traditional methods could not describe the graphs under some non-rigid transformation adequately, three kinds of different methods are presented to recognize images more accurately after summarizing current researches and introducing related techniques.The main contributions and novelties in this thesis are follows:(1) Image recognition methods based on the graph feature are reviewed and then the current researches are introduced. These kinds of image recognition based on the graph feature can be divided into two key steps. One is the extracting and description of graph information. The other is the calculation of distance between two descriptors. From these two aspects, current methods are revealed in this thesis.(2)Edge Direction Histogram method, Earth Mover’s Distance, Shape Context and K-means algorithm are introduced in detail. These methods are theoretical foundation of the following chapters. Three new methods are presented based on the theoretical foundation.(3)The method based on the geometry statistical feature of graph is proposed for image recognition. In the method, graphs are built after calculating corner points. Then edge direction histogram and the edge distance histogram are counted to build the geometry statistical feature of graph. At last, fast and robust Earth Mover’s Distance is used to calculate similarity of graphs. The method describes graph with two different kinds of histograms so it can extend the content of graph.(4)The method based on structure context of graph is proposed for image recognition. Firstly, a sample point set is obtained by discrete sampling. Secondly, the graph structure context descriptor is presented based on the sample point set. At last, the improved Earth Mover’s Distance is used to measure the similarity between graph structure context descriptors. Different from traditional methods, the method uses2D-histograms to describe graph. Hence, it can perform better in experiments.(5) The method based on bag of words context of graph is proposed for image recognition. In this method, firstly graphs are built after corner points are calculated. Secondly, the graph entropy context of every sample point is used to form a codebook. Then the centers of K-means can be obtained to build the feature of graph. Thirdly, the graph bag of words context is calculated to describe the information of graph. Finally, Manhattan distance is used to calculate the similarity between descriptors. This method combines the bag of words and the graph entropy context, so it performs better than traditional methods.The results from retrieval and clustering experiments demonstrate that these three new methods describe graph more adequately, so they perform better than traditional methods. Images which have some structural features can be recognized correctly with these methods.