Fast locality-constrained low-rank coding for image classification

In this paper, we propose a Fast Locality-constrained low-rank sparse coding for image classification. The low-rank coding seeks the homogeneousness and correlation of local features, encodes jointly and globally, based on the traditional low-rank coding, we incorporate locality constraints to enforce the local features sharing the same representation. Considering that the traditional low-rank coding optimization algorithms have the same complexity with the sparse coding, which is difficult to scaly used to classification. We propose a fast low-rank optimization approach, it replaces the nuclear norm with the Frobenius norm, which not only has a closed form solution but sort the class labels very fast. Experiments show that locality works better than sparse at exploiting the local neighbor correlations, low-rank coding shows advantages of finding spatial layout correlations of local features, encoding jointly and globally. It has standout classification performance which has 2% improvement compared to the standard methods. Moreover, our method has great computation efficiency, the coding time is a bit more than LLC, and it is only 19.40%~21.18% of that of ScSPM, the classification time is also the least.