Abstract: Collaborative representation-based classification (CRC) has shown noticeable results on image classification tasks like face recognition and object recognition.It solves a linear problem of the test sample with norm regularization,to obtain a more stable numerical solution.Previous studies have shown that the choice of regularization parameters plays a very important role in the numerical stability of the collaborative representation.This paper proposes a novel image classification method based on test sample error reconstruction collaborative representation-based classification,called TSER-CRC.The first phase of the proposed method uses a smaller regularization parameter to calculate a collaborative representation coefficient and reconstructs the test sample with the obtained coefficient to weaken the error in the original test sample and reduce the inconsistency between the origi-nal test sample and the training samples.The second phase of the proposed method uses the larger regularization parameter and the test samples reconstructed in the first phase to solve the collaborative representation coefficients to obtain the relationship between the numerically stable test sample and the training samples for each class.Finally,the test sample will be classified by conventional classification strategy in CRC.The poposed method can effectively reduce the errors and outliers in the test samples represented by the collaborative subspace composed of all training samples,thereby increasing the stability of the collaborative representation coding coefficients and the robustness of image classification.Experimental results on five standard datasets show that the proposed method can achieve more satisfactory in image classification accuracy than traditional CRC and some others classical image classification methods.

Key words: Collaborative representation, Error reconstruction, Image classification, Pattern recognition, Representation-based classification