合成孔径雷达图像的地面目标检测与伪装效果评估

Ground Vehicles in Synthetic Aperture Radar Image Detection and Camouflage Effect Evaluation

地面目标伪装效果评估是基于目标在探测频段下伪装前后检测概率的变化,合成孔径雷达能够全天候工作,分辨率较高,已经成为对地侦查的主要手段。对于SAR图像的目标检测概率通常采用专家判读和特征模板匹配等方式获取,这些方法主观性较强或对算法设计的依赖度较高,因此需要一种科学客观的评判方法。随着神经网络在图像识别领域的快速发展,目标特征提取和识别结果愈加准确,现采用YOLO-V3(You Only Look Once)算法模型,对MSTAR目标背景合成图像进行机器学习,自动获取目标特征进行识别。训练后的模型在测试集的识别准确度为97.87%,用训练好的模型对未伪装样本与伪装后样本分别进行检测概率计算,通过比较检测概率的变化从而量化某类目标的伪装效果。

The evaluation of ground target camouflage effect is based on the change of recognition probability of target before and after camouflage in the detection frequency band.Synthetic aperture radar(SAR)can work all weather and all day with high resolution,which has become the main means of ground detection.The usual method to acquire target recognition probability of SAR image are expert interpretation and feature template matching.These methods have strong subjectivity or highly depend on algorithm design.So developing a scientific and objective evaluation method is needed.With the rapid evolution of neural network in the image recognition field,target feature extraction and recognition results are more and more accurate.YOLO-V3(You Only Look Once)model is used,which can automatic learn target features and finish recognition.The training samples are composite images of MSTAR target and background.The result has the 97.87%accuracy in the test set after the training.Then the trained model is used to calculate the recognition probability of the un camouflaged samples and the camouflaged samples.By comparing the change of recognition probability,the camouflage effect of a certain kind of target can be quantified.