基于循环生成对抗网络的海上落水人员红外图像检测方法

打开文本图片集

中图分类号：U676.8；TP391.4 文献标志码：A

Abstract： In order to solve the existing problems of the scarce infrared dataset of man overboard at sea， the difficulty of feature extraction，and low detection accuracy for small targets at sea based on infrared images，it is necessary to further expand the dataset and optimize the algorithm.The cycle-consistent generative adversarial network （CycleGAN） is used to expand and construct the infrared dataset of man overboard at sea，and the visible and infrared datasets are used to realize the domain migration，which provides data support for the subsequent target detection. The YOLOv5 model is improved by designing an enhanced path aggregation network（PANet） structure，adding a new small target detection layer and introducing the coordinate atention in the feature fusion part to enhance the detection capability of small infrared targets at sea. The experimental results show that，the CycleGAN data enhancement method can effectively enhance the diversity of the infrared dataset of man overboard at sea，and the average detection accuracy of the improved model is 81.2% ，which is 13.2 percentage points higher than that of the YOLOv5 model. The improved model effctively enhances the detection accuracy and can be applied to the search and rescue mission of unmanned aerial vehicles for man overboard at sea.

Key words： infrared detection technology；man overboard at sea； generative adversarial network （GAN） ; deep learning

0 引言

自国家推行海洋强国战略以来，如何开展海上事故应急搜救成为海事领域的关注热点。（剩余10276字）