Research On Face Classification Of Sparse Representation Based On Kernel Space Coefficient Accumulation

Image recognition, as an important branch of artificial intelligence, has successful application in military affairs, launched investigation, and machine vision. However, it presents images intaken of uncertainty and poses complexity, image recognition is still a challenging research by far. In this paper, focus on the sparse representation classifier decision function and joint sparse representation aspect for working based on the sparse representation analysis of the domestic and international relevant researches.Firstly, in the occlusion face recognition, some covered parts make the property of local information change. It may lead to a wrong classification using the minimum residual as the decision function for sparse representation classification when the residual is approximate. In this case, this paper proceeds from the decision rule of the classifier and proposes the algorithm of sparse representation with weighted fusion of local and global based non-minimum square error for face recognition. It mainly calculates the sparse coefficients accumulation of the global and each local block respectively to Borda vote and then combines the global and local blocks for the final classification.Secondly, in the low dimensional feature space, the data don’t content the linear-separable, which make it difficult to achieve accurate classification when choose the minimum mean square error as decision function. For this problem, a new sparse representation algorithm based on kernel spatial non-minimum residual error is proposed for face recognition. To make the data spatial separable, the proposed method first nonlinearly maps the data into a high-dimensional kernel feature space. Because the kernel space dimension is high, or even infinite, we assume that there is a transformation matrix being the linear combination of the training set in the high dimensional space and utilize the dimensionality reduction method to reduce the samples to the low-dimension subspace. Then, calculate the sparse solution of test sample in the corresponding space and use the sparse coefficient accumulation decision rule to classify.Finally, recent studies on multi-feature object recognition algorithm are usually encoding directly over the original training set, which ignore the redundancy and noise information. So, to solve this problem, this paper proposes a kernel dictionary learning-based multi-feature joint sparse representation for face recognition algorithm. Firstly, extract the various features of the training set and use the kernel-KSVD approaches to calculate the kernel dictionaries corresponding to the features. Secondly, calculate the non-linear mapping matrix of the test image and utilize a joint sparse coding to solve the coefficients. Finally, summarize the reconstruction error of all the features and identify the test sample to the class by using the minimum residual decision function.