Research On Image Classification Based On Dictionary Learning

Sparse representation and joint sparse representation have become main technologies in the fields of pattern recognition.They are mainly used in image classification,image restoration and image denoising.The traditional sparse representation and joint sparse representation represent the input signal linearly through an over-complete dictionary.The performance of sparse coding mainly depends on the reconstruction and discriminative of the over-complete dictionary.Therefore,constructing a discriminative dictionary is very important for sparse representation and joint sparse representation.According to the characteristics of hyperspectral image and face image,this paper proposes the following two improved classification algorithms.For the hyperspectral image classification,joint sparse representation has the advantage of high efficiency.Once the local window of each pixel includes pixels from different categories,the dictionary atoms and testing samples are easily affected by different region.So the performance of the joint sparse representation classifier may be seriously decreased.This issue is attractive topic of current research.According to the characteristics of hyperspectral image,this paper proposes a joint sparse representation fusing hierarchical deep network.It can exploit both discriminative spectral and spatial information simultaneously by alternating between spectral and spatial feature learning operations,then construct a dictionary with spatial spectral features and apply to joint sparse representation.In classification process,the correlation coefficient between the dictionary and the testing samples is combined with the classification error to make decisions.In order to verify the effect of the proposed algorithm,two traditional hyperspectral remote sensing datasets are used to simulation experiment and our method achieves a better performance than the state-of-the-art and traditional algorithms.For the face image classification,extreme learning machine(ELM)and discriminative dictionary learning(DDL)have shown that the speed advantage and the accuracy advantage.However these two methods have their respective drawbacks,in general,ELM is known to be less robust to noise while DDL is known to be time-consuming.In order to unify such mutual complementarity,we propose a discriminative analysis dictionary learning fusing extreme learning machine model in this paper.More precisely,the iterative optimization algorithm is used to learn the most optimal discriminative analysis dictionary and extreme learning machine classifier.In order to verify the effect of the proposed algorithm,four traditional face datasets are used to simulation experiment and our method achieves a better performance than the state-of-the-art dictionary learning algorithms and extreme learning machine.