基于Deformable DETR的红外图像目标检测方法研究

Research on Infrared Image Object Detection Method Based on Deformable DETR

基于Transformer架构的DETR系列网络在计算机视觉目标检测领域不断刷新目标检测的精度与速度。然而,基于红外图像的非合作目标检测的应用环境复杂,而且红外图像质量较差。针对该问题,提出了一种新的以Deformable DETR算法为基线的具有高检测精度的目标检测算法:首先设计了对红外图像进行图像增强处理的图像增强模块CLAHE-GB,并将其与Deformable DETR进行有机结合;然后在大型通用数据集上进行预训练;最后引入数据增强和迁移学习方法在自制的空中飞行物小型红外图像数据集中对检测头网络参数进行再训练,并对结果进行分析。结果表明:所提出的算法对红外图像数据具有较好的图像增强效果和检测精度。

The DETR series networks based on the Transformer architecture keep pushing the boundaries of object detection accuracy and speed in computer vision.However,non-cooperative object detection applications based on infrared images face challenges because of environmental complexity and poor image quality.To solve this problem,a novel object detection algorithm with high detection accuracy was proposed in this study,utilizing the Deformable DETR as the baseline.Initially,an image enhancement module called CLAHE-GB was designed to enhance the image process on infrared images,and it was effectively integrated with Deformable DETR.Subsequently,the algorithm was pre-trained on a large-scale general dataset.Then,data augmentation and transfer learning methods were developed to retrain the parameters of the detection head network using a selfmade dataset of small infrared images of aerial objects.Finally,a comprehensive result analysis was conducted.The results show that the proposed algorithm can successfully achieve promising image enhancement effects and detection accuracy on infrared image data.