基于深度学习的红外图像目标检测技术研究

Research on Infrared Image Target Detection Technology Based on Deep Learning

红外图像具有抗干扰能力强、昼夜工作等优势,已广泛应用在很多领域。但由于红外图像存在信噪比低、检测目标纹理信息缺失、对比度低等不足,导致红外图像的目标检测比可见光图像目标检测的难度更大。本文在YOLOv3算法的基础上,设计GIoU损失函数替代回归损失函数,并在特征提取网络后增加SPP模块,以提高特征表达能力,解决红外图像目标检测精度低问题。实验结果表明,本文设计的优化后的YOLOv3算法的红外图像目标检测的精确性较优化前提升8%以上,较SSD网络提升18%以上,较Faster R-CNN网络提升7%左右,表明该算法在红外图像目标检测方面具有更高的检测准确性与检测效率。

Infrared image has the advantages of strong anti-interference ability and working day and night.It has been widely used in many fields.However,due to the shortcomings of low signal-to-noise ratio,lack of texture information and low contrast in infrared image,target detection in infrared image is more difficult than that in visible image.Based on YOLOv3 algorithm,GIoU loss function is designed to replace regression loss function to improve the accuracy of network positioning;The SPP module is added after the feature extraction network to improve the feature expression ability and solve the problem of low accuracy of near-infrared image target detection.The experimental results show that the near-infrared image target detection accuracy of the optimized YOLOv3 algorithm studied in this paper is more than 8%higher than that before optimization,more than 18%higher than SSD network,and about 7%higher than Faster R-CNN network.It shows that the improved YOLOv3 algorithm has higher detection accuracy and efficiency in near-infrared image target detection.