Research On Image Object Classification Based On Sparse Representation

As we all know, image has a strong capacity in carrying information. With the development of technology, a series of high-tech products working depends on images have come into being, so as to promote the vigorous development of image processing technology.Image Classification as an essential field of image processing, is now in the unprecedented rate of development. However, compared to the speed and accuracy of biological cognition, the image target recognition is still in a relatively backward level. So it has a spatial high research value. Therefore, I choose image target classification based on sparse representation as my graduate research field.In this thesis, I use the Oxford Pet Dataset (catdog database) to carry out the experiments, by improving the dictionary learning algorithm MI-KSVD in M-HMP algorithm (Multipath Sparse Coding Using Hierarchical Matching Pursuit),I propose a new dictionary learning algorithm named MS-KSVD(K Singular Value Decomposition based on Multi Scale Objects). The large number of experiments show that using MS-KSVD algorithm instead of MI-KSVD algorithm can improve the accuracy of image target classification.Further, this thesis analyzes the paths that M-HMP algorithm used to extract the sparse features, and achieves that:1) In the case of maintaining the classification accuracy is not reduced, the number of paths can be reduced, thereby we can reduce the computational time, and reduce the demand for computer storage space; 2) In the case of maintaining the same path number, we can improve the accuracy of target classification. In order to enhance the accuracy of image target classification, our thesis also takes the weight of features into account. Right by changing the feature weight of the different paths, the ultimate accuracy of classification can be improved.In total, this thesis improved the dictionary learning algorithm MI-KSVD and proposed a new algorithm MS-KSVD, change the feature extraction path, set different weight for each features, so as to continuously enhance the accuracy of image target classification based on catdog database. The final classification accuracy was 53.56%, compared with the original accuracy 52.11% which has been increased by 1.45%, which is a breakthrough in catdog database.