Similarity Mask Mixed Attention for YOLOv5 Small Ship Detection of Optical Remote Sensing Images

Abstract

Ship detection has always been an important and challenging task. Small ship targets and complex backgrounds in optical remote sensing images can both lead to false alarms and missed alarms in detection. The YOLO series algorithms have been widely used in optical remote sensing ship target detection, which has the advantage of fast detection speed. However, the YOLO series algorithms have poor detection performance when facing small targets. Therefore, we propose a new algorithm, SMMA-YOLOv5, which can improve the model detection accuracy without significantly increasing the model size. We first introduce a self-attention mechanism to replace some of the convolutional layers to capture the global information of the feature map. Second, we integrate an efficient channel attention (ECA) model to the self-attention mechanism to enable information interaction between channels without adding additional computational overhead. Furthermore, we propose a new similarity mask structure to filter out the invalid regions in the feature map based on the elements’ similarity. The experiments on the public MASATI ship dataset indicate that SMMA-YOLOv5 improves Precision by 3.9%, Recall by 5.3%, and AP by 5.4%, and prove the effectiveness of the algorithm while maintaining real-time detection.