小样本红外图像的样本扩增与目标检测算法

Infrared image sample amplification and object detection method with small samples

中文关键词  红外图像  目标检测  稀缺样本  生成对抗网络  注意力机制  YOLOv3算法

英文关键词  infrared images  object detection  scarce samples  generative adversarial network  attention model  YOLOv3 algorithm

中文摘要

      深度卷积神经网络模型在很多公开的可见光目标检测数据集上表现优异, 但是在红外目标检测领域, 目标 样本稀缺一直是制约检测识别精度的难题. 针对该问题, 本文提出了一种小样本红外图像的样本扩增与目标检测算 法. 采用基于注意力机制的生成对抗网络进行红外样本扩增, 生成一系列保留原始可见光图像关键区域的红外连 续图像, 并且使用空间注意力机制等方法进一步提升YOLOv3目标检测算法的识别精度. 在Grayscale-Thermal与 OSU Color-Thermal红外–可见光数据集上的实验结果表明, 本文算法使用的红外样本扩增技术有效提升了深度网 络模型对红外目标检测的精度, 与原始YOLOv3算法相比, 本文算法最高可提升近20%的平均精确率(mean average precision, mAP).

英文摘要

      The deep convolutional neural network model performs well on many public visible-lighted object detection datasets, but in the field of infrared object detection, the scarcity of object samples has always been a problem that plagues the increase in detection and recognition accuracy. In response to this problem, this paper proposes an algorithm for sample amplification and object detection of infrared images with small samples. The attention-based generative adversarial network was adopted to amplify the infrared samples to generate a series of infrared continuous images retaining the key areas of the original visible light image, and the spatial attention model and other methods were used to further improve the recognition accuracy of the YOLOv3 object detection algorithm. The experimental results on the OSU Color-Thermal infrared-visible light dataset and Grayscale-Thermal dataset show that the infrared data amplification technology of the proposed algorithm effectively improves the accuracy of the deep convolutional neural network for infrared object detection, and the mAP (mean average precision) of the proposed method is 20% higher than the original YOLOv3 algorithm.