小样本条件下基于YOLOv7的小目标检测方法  

YOLOv7⁃based small object detection method under the condition of small samples

作　　者：路琪 郭乐江[2] 于元强 刘飞[1] 熊鑫 LU Qi;GUO Lejiang;YU Yuanqiang;LIU Fei;XIONG Xin(Radar NCO School,Air Force Early Warning Academy,Wuhan 430345,China;Teaching and Research Support Center,Air Force Early Warning Academy,Wuhan 430019,China)

机构地区：[1]空军预警学院雷达士官学校,湖北武汉430345 [2]空军预警学院教研保障中心,湖北武汉430019

出　　处：《现代电子技术》2025年第5期153-161,共9页Modern Electronics Technique

摘　　要：低空慢速小目标的监视一直是预警探测领域的重点和难点。目前主流的基于卷积神经网络的目标检测算法主要设计应用于VOC数据集或COCO数据集,在特定场景下检测精度并不理想。YOLO是目前应用最广泛的单阶段目标检测算法之一,在检测速度方面具有独特的优势。利用可见光成像手段获取小型无人机目标图片,基于YOLOv7算法改进了其特征增强网络,提出一种三分支并行特征金字塔网络,以获得更多的小目标上下文语义特征;将改进后的算法与生成对抗网络进行级联,旨在生成更真实的超分辨率图像,从而提高检测精度。与目前最先进的目标检测方法相比,该方法在满足检测实时性要求的前提下,使得检测精度有了显著的提升。由于训练集有限,为了提高泛化能力,还提出了SOD-Mosaic数据增强方法,该方法提高了检测器的鲁棒性和泛化能力。The monitoring of low-altitude slow small objects has always been the focus and difficulty in the field of early warning and detection.At present,the mainstream object detection algorithms based on convolutional neural networks(CNNs)are mainly designed and applied to VOC data sets or COCO data sets,and the detection accuracy is not ideal in specific scenarios.YOLO(You Only Look Once)is one of the most widely used single-stage object detection algorithms,whose detection speed is marvelous.In this paper,the means of visible light imaging is used to obtain the images(the objects)of small UAV.On the basis of the YOLOv7 algorithm,the feature enhancement network is improved,and a triple-branch parallel feature pyramid network(TPFPN)is proposed to obtain more small object context semantic features.The improved algorithm is cascaded with the generative adversarial network(GAN)to generate more realistic super resolution images,so as to improve its detection accuracy.In comparison with the most advanced object detection methods at present,the proposed method has significantly improved the detection accuracy on the premise of meeting the real-time detection requirements.The quantity of training set is limited,so an SOD-Mosaic(small object detection-oriented Mosaic)data augmentation method is proposed to improve the robustness and the generalization ability of the detector.

关 键 词：自动目标识别 卷积神经网络 小目标检测 数据增强 特征增强 特征金字塔网络 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TH701[电子电信—信息与通信工程] TP391.4[机械工程—仪器科学与技术]