Research On Image Feature Extraction And Recognition Under The Linear Coding Framework

Image recognition and classification has been a hot topic in computer vision and pattern recognition fields. Recently, linear coding based classification methods have caught a lot of attention in the field of pattern recognition due to its efficiency and simplicity. However, the existing linear coding-based methods are easily to be affected by nuisances(e.g. illuminations, view directions, pixel corruptions, and occlusion,etc.) of images, which may make the obtained solution unstable and decrease the recognition rates. This paper makes a thorough and systematic research on image feature extraction and recognition under the linear coding framework. The innovation of this paper is outlined as follows:1) This paper proposes a sparsity embedding projections(SEP) method for sparse representation-based classification(SRC), which seeks a low-dimensional embedding subspace where the sparse representation coefficients of a test sample associated with the training samples from the correct class are enlarged, and simultaneously those associated with the training samples from all of the other classes are compressed. Specifically, given a training data matrix, SEP tries to find a linear transformation by enhancing the intra-class reconstructive relationship of the data meanwhile suppressing the interclass reconstructive relationship of the data in the low-dimensional embedding subspace, and then project the training samples and a given test sample onto the SEP learned subspace,respectively. Finally, SRC is used for classification in SEP subspace.2) Considering that sparsity is helpful to identify the data class, while low-rankness can reveal the subspace structures of data, this paper proposes a low-rank sparse representation-based classification(LRSRC) method for image recognition by combining the merits of both sparsity and low-rankness.Given a set of test samples, LRSRC tries to find the lowest-rank and sparsest representation matrix over all training samples, and then classifies the test sample based on the minimal reconstruction residual.3) In this paper, we propose a novel adaptive locality-constrained regularized robust coding(ALRRC)method for image classification, which is inclined to use the true nearest training samples to represent the test sample by considering both the importance of data features and the adaptive locality of data in coding scheme. Specifically, for a test sample, ALRRC first adaptively calculates the feature weights that can measure the importance of each feature of test sample, and then respectively imposes the feature weights on the test sample and all the training samples to obtain the weighted test sample and all the weighted training samples. Subsequently, a locality-constrained matrix is calculated by using the similarities between the weighted test sample and all the weighted training samples. Since both the weighted test sample and all the weighted training samples have reduced the effect of the aberrant features as far as possible, the locality-constrained matrix can more truly characterize the locality of data. Finally, ALRRC incorporates both the feature weights and the locality-constrained matrix into a unified linear coding framework. An iterative algorithm is also presented to solve the ALRRC optimization problem efficiently.Extensive experiments on the COIL-20, the Extended Yale B, the CMU PIE, and the AR image databases show that the proposed methods are effective and robust for image recognition and classification.