Fine-Grained Vehicle Target Recognition in SAR Images Integrating RCS Distribution Feature

Although the MSTAR dataset under Standard Operating Conditions (SOC) has achieved high accuracy in the recognition of vehicle targets in SAR images, the performance remains insufficient for more refined recognition under Extended Operating Conditions (EOC). The primary reasons for this shortfall are as follows: first, the subdivision of unified vehicle categories into fine-grained subcategories significantly reduces inter-class differences, often manifesting in the detailed features of specific vehicle components; second, most neural network algorithms used for vehicle recognition lack the capability to effectively extract and distinguish features between similar categories. To address these challenges, this paper proposes a novel network integrating Radar Cross Section (RCS) distribution feature, which can assist in capturing subtle differences at the vehicle component level, thereby enhancing target's ability to focus on the structural information. Furthermore, the Cross Attention module employed in the network can effectively model the relationships between the abstract feature extracted from the original image and the RCS distribution feature over long distances. This module further enhances the capability to capture and focus on complementary information from both features, leading to improved feature representation and better discrimination of fine-grained differences. Finally, the proposed method was evaluated on the MSTAR dataset under EOC, resulting an improvement of 1.62% in recognition accuracy. achieving the best performance compared to other methods, which effectively validated its superiority in fine-grained recognition.