Sparse Representation-Based Multiview Incremental Learning for Vehicle Target Recognition in SAR Images

This letter develops a sparse representation-based multiview incremental learning (SRMIL) algorithm for SAR vehicle target recognition, which consists of a learning class-adversarial classifier (LCC) and a multiview incremental learner (MIL). Specifically, LCC involves the reconstruction error term, the class-adversarial term, and the classification term. Among them, the reconstruction error term and the class-adversarial term convert multiview data of each task into the sparse codes, and enforce the sparse codes from the same class to be similar and those from different classes to be dissimilar. The classification term learns the classifier parameters by using the sparse codes. These three terms are jointly optimized to acquire the optimal classifier parameters. MIL decomposes the optimal classifier parameters into the knowledge base and the view-specific sparse codes, in which the knowledge base stores the knowledge from the previous tasks and the new task, and the view-specific sparse codes explore the intrinsic correlation among multiview data by constraining themselves with the self-expressive term. Through the knowledge base, MIL overcomes the catastrophic forgetting of the knowledge from the previous tasks when learning the new task. Finally, the knowledge base and the view-specific sparse codes reconstruct the classifier parameters for recognition. The results verify the effectiveness of SRMIL.