Research On Image Classification Based On Sparse Represention

Image classification is the image processing technology to replace human visual interpretation of image processing technology, which is according the images of different semantic features, each pixel or area in the image to classify into one of several categories.This is widely used in face recognition, character recognition, remote sensing image of technical secondary school of information extraction and dynamic change detection, radar image target tracking, acoustic holography and image sound source localization, Taobao goods dress collocation etc. fields. Sparse coding is a method to simulate the simple cell receptive field of the primary visual cortex V1 area in the mammalian visual system,which can simulate the human visual system and reflect the statistical properties of natural images.Based on bigger superior performance of sparse coding for image classification than the traditional way, this paper concentrate on the research of image classification discrimination dictionary learning method based on the sparse representation. The main work is as follows:First of all, according to the characteristics of the existing linear discriminant dictionary can only describe the semantic information of the image, proposes a classification method of learning Fisher kernel space discrimination dictionary. This method uses kernel mapping, the training samples are mapped to high dimensional kernel space mapping, and then to the Fisher criterion as a constraint to learn a discriminative dictionary, which is used in image classification, and obtained good result.Secondly, the existing image classification methods most is first, image feature extraction, then according to the extracted feature to train classification dictionary, as there is no direct contact between stationary feature extraction method and dictionary learning method will affect the discrimination of the ability to distinguish between the dictionary.Therefore, this paper proposes a method of feature extraction operation and classification of dictionary learning fusion, the method through to feature extraction matrix orthogonal constraint, the complete integration between feature extraction matrix and dictionary,learning the constrained orthogonal projection discrimination based on the dictionary, andused in image classification, get good effect.Finally, when we obtain the image through the electronic devices, some dominant noise doping leads to the image, because of scene illumination is not uniform, and the physical components by temperature change caused by performance instability. such mixed with noise in the training samples will affect learning classification dictionary discriminant, and constrained orthogonal projection based on discriminative dictionary learning methods cannot avoid the noise and that the learned discriminative dictionary classification performance is compromised. The good way is proposed in this paper with noise suppression of projection discrimination dictionary learning method, the method by reference SVD decomposition,*UΣM ?V, matrix information noise will distribution in the after s?1 columns of singular values of the Σ matrix. We can remove the columns portion of the noise information to reduce the scope of projection discrimination Dictionary of training samples. Then used the method to image classification and has got good effect.