基于注意力机制和改进YOLOv3的红外弱小目标检测

红外弱小目标检测技术是红外探测系统的核心技术之一。针对远距离复杂场景下红外弱小目标对比度低、信噪比低和纹理特征稀疏分散导致目标检测率低的问题，提出一种融合注意力机制和改进YOLOv3的红外弱小目标检测算法。首先，在YOLOv3的基础上，用更大尺度的检测头替换最小尺度的检测头，在保证推理速度的基础上有效提升了红外图像中小目标的检测概率。然后，在检测头之前设计了Infrared Attention模块，通过通道间的信息交互，抽取出更加关键重要的信息供网络学习。最后，用完全交并比损失（Complete IoU Loss）替代交并比损失（Intersection over Union Loss）来衡量预测框的检测能力，通过梯度回传实现更好的模型训练。实验结果表明，本文提出的YOLOv3-DCA 能完成多种场景下红外弱小目标的检测任务，且检测准确率、召回率、F1和平均准确率分别达到91.84%、88.85%、93%和88.82%，平均准确率比YOLOv3基线提升约7%，与主流的SSD、CenterNet和YOLOv4模型对比平均准确率也取得了目前最优。

Infrared Small Target Detection Based onImprovedYOLOv3 and Attention Mechanism

Infrared dim and small target detection technology is one of the core technologies of infrared detection systems. Aiming at the low target detection rate caused by low contrast, low signal-to-noise ratio, and sparse texture features in remote complex scenes, a fusion of attention mechanism and improved YOLOv3 infrared dim target detection algorithm is proposed. First, based on YOLOv3, the smallest detection head is replaced with a larger detection head, which can effectively improve the detection probability of small targets in infrared images with ensuring the inference speed as much as possible. Then, the Infrared Attention module is added before the detection head to realize the information exchange between channels and focus on the important target features. Finally, use Complete IoU Lossinstead of Intersection Over Union Loss to achieve better training of the model. The experimental results show that the YOLOv3 DCA proposed can complete the detection of infrared small targets in various scenarios. The accuracy, recall, F1, and average accuracy of detecting infrared dim small targets by the optimized YOLOv3 algorithm are 91.84%, 88.85%, 93%, and 88.82%, respectively. The average accuracy is about 7% higher than the YOLOv3 baseline, and the average accuracy is currently the best compared to the SSD, CenterNet, and YOLOv4 models.