融合机器视觉和无监督域适应的轻型弱小目标检测方法

Light weak target detection method combining machine vision and unsupervised domain adaptation

针对轻型弱小目标因体积小及亮度弱等导致目标检测跟踪困难的问题,提出融合机器视觉和无监督域适应的轻型弱小目标检测方法。采用Gamma校正方法对图像中弱亮度的轻型弱小目标进行光照补偿,增强目标轮廓清晰度,提取并融合特征显著图获得图像目标区域,通过YOLO-V3网络将原始图像集作为源域样本训练网络,将目标区域作为目标域样本展开无监督域适应的目标检测。结果表明,所提方法的图像目标提取精度提高了2.05%,目标检测精度达82.36%,相较于其他对比方法检测精度提升了2.1%。验证了该方法检测轻型弱小目标的有效性。

This paper proposes a light and weak target detection method combining machine vision and unsupervised domain adaptation to address the difficulty of light and weak target detection and tracking due to small size and weak brightness.The method works by using Gamma correction method to compensate for the light and small targets with weak brightness in the image in order to enhance the clarity of the target contour;extracting and fusing feature saliency maps to obtain the target area of the image;training the network by using original image set as the source domain sample,and conducting the target detection with unsupervised domain adaptation by using the target area as the target domain sample through YOLO-V3 network.The results show that the precision accuracy of object extraction improves by 2.05%,and the accuracy of object detection is up to 82.36%with the proposed method,which increases by 2.1%of the accuracy with other methods,as which verifies its better detection effectiveness on the light and weak target detection.