MLSDNet: Multiclass Lightweight SAR Detection Network Based on Adaptive Scale Distribution Attention
Hao Chang, Xiongjun Fu, Jian Dong, Jiaang Liu, Zixiang Zhou
Abstract
Deep learning has made rapid progress in the field of synthetic aperture radar (SAR) detection. However, SAR images themselves have limited information, and a general detection network that is too wide and too deep can result in computational complexity and memory waste. Therefore, we design a lightweight network for multi-class SAR detection based on adaptive scale distribution attention. Firstly, a novel backbone is designed from the perspective of lightweight model, using deep separable convolution to generate high-quality feature maps of protruding targets, and applying channel shuffle to improve training and detection efficiency. Secondly, a lightweight adaptive scale distribution attention is proposed, which can adaptively obtain the scattering information of multi-scale targets, aggregate the position and contour features of the targets, and improve the detection accuracy of multi-class targets. Finally, anchor-free detection head is applied to improve the generalization ability and robustness of the model. MLSDNet achieve a high mean average precision (mAP) of 92.99% on the newly released multi-class SAR target datasets (MSAR-1.0) with only 1.42G FLOPs and 928.25K Params. The mAP on SAR ship datasets such as SSDD and HRSID reached 99.1% and 94.7%, respectively. The mAP on the latest SAR aircraft dataset reached 97.7%, demonstrating its good generalization ability. Its performance has reached the state-of-the-art (SOTA).