基于改进YOLOv4网络的红外遥感小目标检测方法

Infrared Remote Sensing Small Target Detection Method Based on Improved YOLOv4 Network

针对传统检测方法对红外小目标检测性能不足的问题,提出一种基于迁移学习与改进YOLOv4网络的红外小目标检测系统。首先,对YOLOv4网络主干网提取的浅层特征进行增强,并结合深层特征与浅层特征来缓解红外小目标难以检测的问题;其次,为YOLOv4网络的检测头模块增加注意力机制,使网络关注于特征图中的红外小目标,从而降低背景对小目标检测的干扰;最终,在YOLOv4网络的训练过程中加入迁移学习方法,从而解决红外小目标标注训练数据不足的问题。基于公开红外小目标检测数据集的实验结果表明,该系统有效提高了YOLOv4网络对红外小目标的检测性能,且优于其他的对比检测模型。

Targeting at the poor performance of traditional detection methods for infrared small target,a transferring learning and improved YOLOv4 network based infrared small target detection system is proposed.Firstly,the shallow features extracted by backbone of YOLOv4 network are enhanced,and the difficulty of infrared small target detection is reduced with combination of shallow features and deep features.Secondly,an attention mechanism is introduced to the detection head of YOLOv4 network to help the network focus on infrared small targets of the feature maps,thus,the background interference to small target detection is reduced.Finally,the transferring learning method is introduced to the training process of YOLOv4 network to solve the problem of lack of labeled training data for infrared small target detection.Experimental results based on public infrared small target detection dataset show that the proposed system improves the detection performance of YOLOv4 network for infrared small target,it also outperforms the other compared detection models.