Improving the ScSPM model with Log-Euclidean Covariance matrix for scene classification

Abstract

The framework of the ScSPM (Spatial Pyramid matching method using Sparse Coding) model is concise, but a good performance in scene classification is achieved. However, its performance can not be significantly improved duo to the limited discriminative power of the SIFT descriptors. To address the problem, covariance matrices as region descriptors are introduced to incorporate with the SIFTs. For computing the distance between them, covariances are transformed to LECM features by matrix logarithm operation. Moreover, exponential weights are imposed on the pooled features to enhance the performance of linear kernel SVM. Experiments on the public datasets demonstrate that the performance of the ScSPM can be improved dramatically by combining the LECM features, and our model achieves the performance competitive with previous methods.