The Global Plus Local Feature Extraction Based On Collaborative Representation And Its Application

Image recognition is a hot research field in statistics and computer science. Its involved data is generally high dimensional. On one hand, the characteristics of high-dimensional data will result in time and space complexity too high in traditional process, and even cause the“curse of dimensionality”; On the other hand, a large amount of irrelevant or redundant and noise characteristics exist in high-dimensional data, will seriously affect the performance of image recognition. So how to choose an effective feature to reduce data dimension is an important issue in image recognition.This paper mainly research the feature selection and extraction method in the case of high dimensional small sample, and experiments on face database. Concrete work can be summarized as follows:(1) Research of feature extraction methods based on linear discriminant analysis(LDA).Mainly research several extension methods of LDA which focus on LDA ignores the geometrical structure information of local points. Discuss the performance of the complete global-local LDA(CGLDA). The method through high-dimensional data set’s global discriminant information, global similarity information, local pattern variation, and local similarity information extract features, that can reduce or even avoid over-fitting, thus improve the recognition performance of LDA. But the CGLDA also has deficiency, its calculation speed is relatively low in application and its extracted features’ s recognition performance need to be further improved.(2) Propose collaborative representation based global plus local feature extraction(CRGLFE). On the basis of CGLDA, CRGLFE use inter-class sparse reconstruction information to replace global similarity information to extract features. Solve the problem that calculation speed of CGLDA is low. Has the advantages of reduce over-fitting, adaptive adjustment of data, strong anti-noise ability, fast computation speed.(3) Propose collaborative representation based global plus local linear discriminant analysis(CRGLLDA). On the basis of CGLDA, CRGLLDA use inter-class sparse reconstruction information replace local similarity information to extract features. Solve the problem that recognition performance of CGLDA is low. Has the advantages of reduce over-fitting, strong anti-noise ability, high recognition performance and strong stability of extracted features.The method proposed in this paper on the experimental results of image recognition field’s Yale, ORL and AR common human face data sets show: the CRGLFE ensure itsextracted features’ recognition performance and stability comparable with CGLDA, and its run-time less than CGLDA. The CRGLLDA ensure run-time comparable with CGLDA, and its extracted features’ recognition performance is higher and stability is stronger than CGLDA.