在目标检测领域, 基于深度学习的SSD目标检测网络同时具有实时性好和准确性高两大优点。由于特种车辆红外图像难以获取, 以小轿车和公交车红外图像为研究对象, 构建了红外图像Pascal VOC数据集, 训练了SSD网络, 并利用训练好的网络检测了红外目标图像。结果表明, 红外目标的特征信息越多, 检测精度越高, 但红外图像中信息残缺的车辆存在“漏检”的问题。针对该问题, 通过添加“残缺窗口模块”优化数据集结构, 有效解决了车辆“漏检”问题, 同时目标整体的检测准确率也明显提升。将改进数据集后的红外目标检测结果作为评价指标, 能够较准确评估复杂背景下特种车辆红外隐身伪装效果。

In the field of target detection, the Single Shot multibox Detector(SSD) target detection network based on deep learning has two advantages of good real-time performance and high accuracy. Because the infrared image of special vehicles was difficult to obtain, the infrared image of car and bus were taken as the research object, the Pascal VOC dataset of infrared image was constructed, the SSD network was trained, and the infrared target image was detected by the trained network. The results show that the more the feature information of the infrared target, the higher the detection accuracy, but the problem of “missing detection” of the vehicle with missing information exists in the infrared image. In response to this problem, the data structure was optimized by adding the "incomplete window module", and the problem of "missing detection" of the vehicle was effectively solved, and the detection accuracy of the target as a whole was also significantly improved. The infrared target detection result after improving the data set was used as the evaluation index, which can accurately evaluate the infrared stealth camouflage effect of special vehicles under complex background.