Locality-Constrained Linear Coding with Spatial Pyramid Matching for SAR Image Classification

Abstract

We propose a linear spatial pyramid matching using locality-constraint linear coding for SAR image classification based on MSTAR database. Recently, works have little consideration about targets’ randomly distributed poses when applying sparse coding in coding scheme. We do the preprocessing of pose estimation to generate over-complete codebook and therefore reduce reconstruction error. SIFT descriptors extracted from images are projected into its local-coordinate system by Locality-constrained linear coding instead of sparse coding. Locality constraint ensures similar patches will share similar codes. The codes are then pooled within each sub-region partitioned according to spatial pyramid and concatenated to form the final feature vectors. We use max-pooling which is more salient and robust to local translation. With linear SVM classifier, the proposed approach achieves better performance than traditional ScSPM method.