AAFCFPNET: AN ADAPTIVE ATTENTION MULTI-SCALE FEATURES CROSS FUSION PYRAMID NETWORK TO IMPROVE SAR SHIP DETECTION

Deep learning (DL) based methods are very popularly used in SAR image ship detection field. Complex background and heavy speckle noise, however, can severely cause false alarms and degrade the performance of DL-based models. This paper proposes a novel SAR ship detection method based on an adaptive attention multi-scale features cross fusion pyramid network, namely AAFCFPNet. More specifically, an attention-based target’s features selection module is firstly developed to effectively select the real ship regions features with the supervision of an attention map of target regions. Then, an adaptive multi-scale features cross fusion pyramid module is embedded to efficiently fuse multi-scale features from forward and backward feature maps to obtain the more fine-grained detection heads (i.e., feature maps). Experiments results on SSDD and AIR-SARShip-1.0 ship detection datasets have demonstrated that our proposed method has better performance in terms of precision, false alarm rate, and generalization in complex scenarios, compared to some existing methods.