基于特征融合与注意力的遥感图像小目标检测

Small Object Detection in Remote Sensing Images Based on Feature Fusion and Attention

为解决遥感图像小目标检测中目标特征信息量少、定位困难等难题,提出一种基于特征融合与注意力机制的遥感图像小目标检测算法FFAM-YOLO(Feature Fusion and Attention Mechanism YOLO)。该算法首先针对主干网络特征提取有效信息量少、特征图信息表征能力弱的问题,构造特征增强模块(FEM)以融合较低层级特征图中多重感受野特征,提升算法主干网络的目标特征提取能力;其次,主干网络提取得到高低层级特征图后,建立重构算法的高低层级特征融合结构,利用特征融合模块(FFM)显著增强小目标的特征信息;在增强的有效通道注意力机制(E-ECA)与空间注意力模块(SAM)所组成的级联注意力机制(ESM)作用下,可更精确地捕获小目标特征;最后在输出的两路特征图上进行小目标检测并输出结果。实验结果表明,基于构建的遥感图像小目标数据集USOD(Unicorn Small Object Dataset),所提算法的查准率达到91.9%,查全率达到83.5%,检测框与真实框之间的交并比阈值(IoU)为0.5时的平均精度(AP)为89%,IoU为0.5∶0.95时的AP达到32.6%,检测速率达到120 frame/s,具有一定的鲁棒性和实时性。

To deal with issues such as less feature information and difficult positioning raised by small object detection in remote sensing images, this paper proposes a remote sensing image small-target detection algorithm FFAM-YOLO(Feature Fusion and Attention Mechanism YOLO) based on feature fusion and attention mechanism. Firstly, in terms of inadequate effective information in backbone network feature extraction and weak information representation in feature maps, the algorithm constructs a feature enhancement module(FEM) to fuse multiple receptive field features in lower-level feature maps and improve the network’s ability in extracting object features. Secondly, with low-level and high-level feature maps obtained by the backbone network, the algorithm’s low-level and high-level feature fusion structures are rebuilt, and a feature fusion module(FFM) is implemented to enhance the feature information of small targets. Thirdly,small object features are accurately captured by cascade attention mechanism(ESM) consisting of enhanced-efficient channel attention(E-ECA) and spatial attention module(SAM). Finally, the small object is detected in the output dualbranch feature maps, and results are delivered. The experimental results show that with the USOD(Unicorn Small Object Dataset), based on the constructed remote sensing images, the proposed algorithm achieves a precision of 91. 9% and a recall of 83. 5%, with an average precision AP of 89% for intersection ratio threshold(IoU) between the prediction box and the ground truth box of 0. 5 and an AP of 32. 6% for IoU of 0. 5∶ 0. 95, respectively, and the detection rate reaches 120 frame/s. The algorithm is with robustness and real-time performance.