MULTI-VIEW SAR TARGET RECOGNITION BASED ON SEMANTIC-VIEW DUAL-BRANCH FEATURE EMBEDDING FUSION NETWORK

In the single-view SAR target recognition task, the image features of SAR targets may not be adequately represented from a single observation view, and these features may vary significantly across different observation views, leading to recognition failure when relying solely on single-view images. Existing multi-view fusion methods usually implicitly hypothesize that the views of images are known, which restricts the applicable scenarios. To address these issues, we propose a semantic-view dual-branch feature embedding fusion network, which achieves flexible and accurate multi-view target recognition by utilizing prior view information. The model design mainly consists of three parts: 1) A dual-branch feature extraction module is constructed to extract the semantic features and view features of single-view images respectively. 2) A multi-view feature fusion module is built based on Transformer, which conducts embedding on the dual-branch feature maps and realizes multi-view feature fusion through the self-attention mechanism. 3) A multi-task learning loss function is designed to jointly optimize the tasks of single-view semantic classification, view classification, and multi-view semantic classification. Experiments demonstrate that the proposed method can effectively enhance the accuracy and robustness of target recognition, and it enables end-to-end recognition of multi-view images with unknown views.