Multi-scale Attributed Graph Kernel For Few-shot Image Categorization

With the rapid development of science and technology,and the progress of Internet technology,computer software and hardware technology have been developed rapidly.In this context,artificial intelligence and machine learning algorithms have been greatly developed.Image categorization has always been a basic problem in the field of artificial intelligence.It is also the basis of other advanced visual tasks such as face recognition,license plate recognition and network image management.Therefore,in view of the importance of image categorization in the field of computer vision,it is of great theoretical and practical significance to study image categorization algorithms with robustness and accuracy.There are still many problems in image classification,such as irregular change of non-rigid objects,change of image perspective,scale change,change of illumination and occlusion.All of these problems will bring challenges to image categorization.In order to classify images better,after decades of development,researchers have developed a large number of image categorization algorithms,which can be roughly divided into two categories:image categorization algorithms based on traditional machine learning,such as image categorization methods based on Bag of Word model,image categorization methods based on Spatial Pyramid and image categorization methods based on graph representation.These methods have the explicit theoretical exposition and proof,the complexity of the algorithm are relatively low and the accuracy need to be improved as well.The other is image categorization algorithms based on deep learning,such as Convolution Neural Net-work and Residual Neural Network.The advantage of these methods are that they can extract image features through the deep network,which are competent for large-scale image categorization tasks.The disadvantage of these methods is that the model needs a lot of data training,and a large number of parameters should be optimized by network.This paper studies the few-shot image categorization,and the main contributions are as follows:(1)Spatial Pyramid Matching has been widely used in scene recognition and image retrieval.It divides the image into a series of sub-regions and counts the local features in each sub-region.However,the Spatial Pyramid Matching method does not describe the spatial relationship of the image local features.In order to represent the spatial correlation of the image local features at different scales,this paper constructs the multi-scale attributed graph from image which contains Bag of Word labels,calculate the distance between the corresponding attributed graphs of any two image regions,and find the optimal matching.Then,the distance of attributed graphs at different scales is pooled to construct the kernel matrix for image categorization.Compares five different image categorization algorithms on four common data sets,Caltech 101,Caltech 256,Scene Categories and Six Actions.The experimental results show that the method is effective in dealing with image categorization problems.(2)Fast and refined multi-scale attributed graph kernel for image categorization is proposed.PCA-guided K-means method is firstly used to cluster the image features and construct the vocabulary tree as the label of multi-scale attributed graph.Then,the attributed graph of layer 0 is refined to obtain the refined attributed graph.Then,for the scale of the first level and above,the distance of corresponding attributed graphs of any two image regions is calculated,then the optimal matching is found.The distance of attributed graphs at different scales is pooled to construct the kernel matrix.This method optimizes the computational complexity of multi-scale attributed graph kernel for image categorization.Compares with six different image categorization algorithms on four common data sets,the experimental results show that the proposed method achieves better accuracy and reduces time complexity in image categorization.