轻量级SAR 船舶检测：散斑注意机制的研究（基于YOLOv12）

Abstract

This paper presents an improved YOLOv12 model that integrates the Speckle Attention Mechanism (SAM) to enhance small target detection in complex scenes. The model utilizes a spatial-channel collaborative attention module to generate energy weight maps, achieving a lightweight design with a channel compression ratio of 1/8 and only a 1.2 % increase in parameters. Experiments on a marine ship dataset demonstrate that the improved model achieves a mAP50 of 96.5 % (+ 2.3 %) and mAP50-95 of 60.9 % (+ 10.34 %) after 300 training epochs, while accelerating convergence speed by 12.5 %.