Natural Scene Recognition Based On Graph Edit Distance

The image recognition is an important branch in the field of pattern recognition.In view of its widely use in medical, aerospace, military and other fields, and theimportant role, image recognition has always been the focus of researchers. Due tothe diversity and ambiguity of classification, natural scene image recognition is oneof the difficulties of the field of image recognition, has not been able to form a clear,unified classification and identification system standard. Feature extraction anddescription of the natural scene images plays a vital role in the recognition process.Owing to the diversity of the natural features of image content and the sensitivity ofthe illumination, image size and other factors, traditional single feature imagedescription method is severely restricted.On the basis of summarizing the previous studies, this paper proposes a naturalscene recognition method based on graph edit distance. This method uses thetechnology of multi-feature fusion, and considers the color feature, the shape featureand the texture feature of the natural scene image comprehensively. By these, themethod effectively avoided the impact of the external condition such as illumination,the image size, in the process of recognition. In the process of feature extraction, thedominant color descriptor and the edge orientation histogram descriptor in theMPEG-7has been used.Moreover, a labeled graph based representation method of natural scene imageis proposed. In this method, the natural scene image is divided into several areas first.And then, select some meaningful region for recognition, and extract its features byusing the extraction method mentioned before. In the meantime, abstract each regionto vertex and its features extracted before is taken as the label of responding vertex.Finally, natural scene images were abstracted to be labeled graphs. After these, theproblem of measuring the similarity between two scene images is transformed intolabeled graph matching problem.On the basis of previous work, this paper uses the diffusion kernel function toconvert the similarity matrix into graph kernel. Finally, take the kernel graph into theSVM for image recognition.The research of the paper mainly includes the following three aspects:(1)Research the description method of natural scene recognition and combine with the graph based representation method of image. Summarize the advantagesand disadvantages of traditional representation method of natural scene image, andpropose a suitable comprehensive representation method for natural scenerecognition.(2)Define a suitable edit cost function, and transform the problem of measuringthe similarity between two scene images into a labeled graph matching problem.(3)Research the application of the kernel theory in the graph area. Learn themethods of graph kernel construction, and select one of them which is suitable tonatural scene recognition based on graph edit distance. Improve it.Finally, experiments based on two graph datasets are carried out respectively.One of them consists of two categories of natural scene, country and city. The othercontains eight categories: highway, tall building, open country, street and mountain,inside city, forest and coast. Experiment results show that natural scene recognitionmethod based on graph edit distance proposed in this paper has a strong versatility.Meanwhile, the recognition method is robust to both the illumination and the size ofthe image, and obtains reasonable recognition results.