A LIGHTWEIGHT NETWORK FOR MULTISCALE SAR SHIP DETECTION UNDER COMPLEX IMAGERY BACKGROUNDS

Most SAR ship detection methods based on deep learning merely pursing high detection accuracy while ignoring the model's complexity. And the factor of serious interfere caused by speckle noise is not considered, thus leading to the detection performance decline under complex imagery backgrounds. To address these problems, a lightweight network for multiscale SAR ship detection is proposed. The backbone of YOLOX is replaced by improved attention ShuffleNetV2 (IAS), which has fewer parameters and better feature extraction ability. Then, a lightweight attention enhanced path aggregation feature pyramid network (LAE-PAFPN) is proposed. Three parallel ECA attention modules are embedded into LAE-PAFPN to refine the feature of ships while suppressing the interfere of the speckle noise. The experiments are conducted on SSDD dataset, show that the mAP of our method have achieved to 97.93% while the FLOPs and parameters are 11.05 G and 2.84 M, respectively.