A Research Of Image Classification Method Based On Spatial Pyramid Model

Image classification, as a key technique of image retrieval, image filtering and imagerecognition, has becoming one of the most important research area in the field of the patternrecognition. It aims as classifying images into different categories according to their characteristics.Nowadays, the most classical classification model includes BOF and SPM, SPM model in the fieldof image classification has been taking more and more attentions. On the one hand, it consideringthe spatial relationship between visual words. On the other hand, it avoiding missing some usefulinformation on the spatial structure in the processing of classification caused by BOF. Hence, SPMreduces the loss of image information, bringing efficient classification performance.This paper attempts to adapt SPM model as the basic framework on image classification. Itbased on the technology of ensemble learning and feature fusion, constructing a spatial visualdictionary for performance improvement and further dealing with the image descriptions by usingdifferent methods. Researches are carried out step by step for image classification based on theSPM model.In the view of the superiority of ensemble learning method, an image classification method ofintegrating multi-features and sparse coding is proposed. Images are first divided into sub regionsaccording to spatial pyramid, and then SIFT and HOG features are combined for using theircomplementary advantages to produce various feature sets. Afterwards, different clusteringmethods are used on different feature sets to produce different codebooks. Then two sparse codingmethods, LLC and SC based on each codebook are further applied respectively to get variousimage description sets. Finally, a voting method is applied to estimatethe final classification.In order to develop the potential of space visual codebook on expressing images, we continueconducting two studies: a method based on multiple level spatial visual codebook ensemble and amethod of constructing discrimination spatial visual codebook. For the former, it aims to buildmultiple level spatial visual codebook ensemble, the codebooks are hierarchically constructed withdifferent levels, and each ofthemare divided fromthe globalto fine subspace to produce codebookensemble, which then be integrated together. After that, the features of each region are coded byLLC based on its corresponding codebook, and then the coded feature vectors are given differentweights according to the different contribution of each region. Finally, the feature vectors ofdifferent regions are concatenated and regarded asthe final image description.For the latter, a method of constructing discrimination spatial visual codebook is proposed, its basic starting point is obtaining more discriminative codebooks by first dividing images into subregions according to spatial pyramid, and then extracting the most discriminative visual wordsaccording to each sub-region to generate visual codebook. After forming the original codebooks,we select some ofthe most frequent visual words in training images to formthe final codebook.