Class Specific Centralized Dictionary Learning based Kernel Collaborative Representation for Fine-grained Image Classification

Classification algorithms based sparse coding have formed a mature system for visual recognition. Recent studies suggest collaborative representation is a much more effective method for classification, compared with sparse representation, the objective function of collaborative representation is constrained by l(2)-norm. Traditional collaborative representation based classification always uses a set of training samples to construct a dictionary directly, which causes high residual error and thus reduces the correct rate of classification. To handle the problem, we propose an innovative method, which integrates centralized image coding and class specific dictionary learning algorithm with collaborative representation based classification together, namely class specific centralized dictionary learning based collaborative representation (CSCDL-CRC). Meanwhile, kernel method can obtain nonlinear information between data points through mapping feature space to kernel space, especially when it is applied to image classification. We extended our proposed CSCDL-CRC to the kernel space to improve the classification performance. We make plenty of experiments on three frequently-used fine-grained image datasets, including Caltech-UCSD Birds-200-2011 (CUB-200-2011) dataset, Oxford 102-Flowers dataset and Stanford Dogs dataset, to validate the effectiveness of the proposed approach.