Research On Image Classification Of Location And Orientation Pooling

Now image classification is an important issue in Computer Vision and Pattern Recognition area. In image, classification and retrieval, the most popular and effective method is the model based on bag of words representation. Now a lot of models based on bag of words all improve the expression of coding to enhance image representation. But most of them ignore feature pooling after coding, so this paper presents:soft-assignment location and orientation pooling(SALOP) and high-order information of hard-assignment location and orientation pooling for image classification(H-LOP).This paper work includes:(1) Based on the local dictionary, we use location and orientation pooling between the local descriptors and the atoms of dictionary for image classification. Firstly, in order to improve the representation of images, this paper adds orientation information to the local descriptors and the atoms of dictionary, which is inspired by feature extraction method. Then each image contains richer information. Secondly, this paper combines a new pooling method with some popular pooling methods and image classification performance significantly improved. Finally, this paper use soft-assignment to replace hard-assignment to reduce the ambiguity and uncertainty of coding expression, which also improve the performance of image classification.(2) It turns out that using high-order statistical information local descriptors and atoms of dictionary can improve the performance of image classification, which can be verified from the instances of Fisher Vector(FV), High-order vector of locally aggregated descriptors (H-VLAD) and Second-order pooling(02P). So high-order information in image classification based on hard-assignment location and orientation pooling(LOP) are proposed, leveraging both location and orientation information between local descriptors and atoms of dictionary. Firstly, this paper compute first-order information between descriptors and atoms in each histogram bins. Then we concatenate first-order and hard-assignment location and orientation pooling information to represent images. Secondly, high-order information can improve the accuracy of image classification, because second-order can distinguish the manifold structure between descriptors and atoms.(3) Experiments and analysis are implemented on three standard database PASCAL VOC 2007, Caltech101, Scene 15 and show that the a higher results of location and orientation pooling. Firstly, experiments show that orientation information bring effective results. Then we compare the results of soft-assignment and hard-assignment on three databases with different sizes of dictionary and pooling methods. Finally, results on Scene 15 database show that high-order statistics is crucial to image classification. Firstly, verify the influence of first-order information on image classification. Then we compare with state-of-art methods with high-order information. It turns that high-order information has a good image classification consequence.