Image Classification Based On Local Features

Image classification is indicated the category based on the content ofan image. It is an important problem in pattern recognition which is verycrucial to image retrieval and image recognition. So, many researchers tryto solve this problem and put forward some methods. Image classificationbased on the local feature achieved the best performance up to now. One ofthe most popular methods is BOW model which is very easy to understandand implement. There are about four main steps in BOW model, which isfeature extracting, codebook training, feature coding and classifiertraining.In this paper, the research contains three parts which are featureextracting, codebook training and feature coding.(1) First, a novel method is raised based on Gist information detection.Compared to the traditional feature extracting methods, our approach has abetter performance. At the beginning, the object regions are located basedon the Gist information, then extracted more features from the objectregions and extracted fewer features from the background. So, most of thefeatures are extracted from the object regions. Then these features are usedto train a codebook which will have a better discrimination performance.Our methods achieve about1%better performance than the normal methodin database of Caltech-101and Caltech-256.(2)Secondly, KNN coding method is used to replace the soft-assignment method of the Fisher kernel. It can achieve a betterperformance than BOW model, but its computational complexity is muchhigh than BOW which is hardly prevent Fisher kernel from being wildlyused. So, the computational complexity will be greatly reduced after usingKnn coding method. Experiments on PASCAL VOC2007show that the improved Fisher kernel not only keeps the classification performance butalso has low computational complexity.(3) Finally, a method is raised to combine multi-kernels andmulti-features. In the fourth chapter, the final image representation ofFisher kernel has a high information redundancy. In order to better applythe information of local features, we use multi-kernel method to combinethe BOW model and Fisher kernel. The experiment results show thatmulti-kernels and multi-features methods have a better performance.