基于改进YOLOv5的复杂背景红外弱小目标检测算法

针对传统算法依赖于对红外目标与环境背景的精确分离和信息提取，难以满足复杂背景和噪声等干扰因素下的检测需求。论文提出一种基于改进YOLOv5（You Only Look Once）的复杂背景红外弱小目标检测算法。该算法在YOLOv5基础上，添加注意力机制提高算法的特征提取能力和检测效率，同时改进原YOLOv5目标检测网络的损失函数和预测框的筛选方式提高算法对红外弱小目标检测的准确率。实验选取了来自不同复杂背景的7组红外弱小目标数据集，将这些图像数据集进行标注并训练，得到红外弱小目标检测模型，然后从模型训练结果和目标检测结果的角度评估算法和模型的正确性。实验结果表明:改进的YOLOv5算法训练出来的模型，检测准确性和检测速度对比实验列出的几种目标检测算法均有明显的提升，平均精度均值（mean Average Precision，mAP）可达99.6%以上，在不同复杂背景下均可有效检测出红外弱小目标，且漏警率、虚警率低。

Improved YOLOv5-based Infrared Dim-small Target Detection under Complex Background

Using the traditional algorithm to meet the detection requirements of interference factors, such as complex background and noise, relying on the precise separation and information extraction of infrared targets and environmental background, is difficult. This paper presents a dim–small target detection method for an infrared imaging algorithm based on the improved YOLOv5 for complex backgrounds. Based on YOLOv5, an attention mechanism is introduced in the algorithm to improve the feature extraction ability and detection efficiency. In addition, the loss function and prediction box screening method of the original YOLOv5 target detection network are used to improve the accuracy of the algorithm for infrared dim–small target detection. In the experiment, seven sets of infrared dim–small target image datasets with different complex backgrounds are selected, the data are labeled and trained, and an infrared dim–small target detection model is established. Finally, the accuracy of the algorithm and model is evaluated in terms of the model training and target detection results. The experimental results show that the model trained by employing the improved YOLOv5 algorithm in this study has a significant improvement in detection accuracy and speed compared with several target detection algorithms used in the experiment, and the average accuracy can reach more than 99.6%. The model can effectively detect infrared dim–small targets in different complex backgrounds, and the leakage and false alarm rates are low.