基于轻量级YOLOv4的小目标实时检测  

Real-Time Detection of Small Targets Based on Lightweight YOLOv4

作　　者：刘雨青[1] 隋佳蓉 魏星 张中林 周彦 Liu Yuqing;Sui Jiarong;Wei Xing;Zhang Zhonglin;Zhou Yan(College of Mechanical Engineering,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学工程学院,上海201306

出　　处：《激光与光电子学进展》2023年第6期97-104,共8页Laser & Optoelectronics Progress

基　　金：上海市水产动物良种创制与绿色养殖协同创新中心(A1-3605-21-0002)。

摘　　要：为实现渔业养殖中的精准投喂,在传统检测算法的基础上,提出了一种基于鱼群纹理、形状和密集度特征的轻量级鱼类摄食行为实时检测算法F-YOLO.将YOLOv4算法原来的主干特征提取网络CSPDarkNet53替换为MobileNetV3,以少量检测精度下降的代价极大提升网络的实时检测性能,提升对鱼类小目标检测性能;对网络结构卷积层进行通道剪枝和知识蒸馏处理压缩模型,减少浮点运算次数(FLOPs)和计算量;使用优化K-means聚类结合添加全局非极大值抑制的DIoU损失函数确定锚框,解决鱼体相互遮挡导致锚框缺失问题.实验结果表明,所提F-YOLO算法的模型大小仅为13.7 MB,每张图片平均识别时间达到50 ms,精度达99.13%,FLOPs仅为1.64×1010,在嵌入式设备中的检测速度可以达33 frame/s,可为实际渔业养殖提供理论指导.A realtime detection algorithm FYOLO for the feeding behavior of lightweight fish based on the fish swarm’s texture,shape,and density characteristics is proposed to realize accurate feeding in fishery breeding based on the traditional detection algorithm.The initial backbone feature extraction network CSPDarkNet53 of the YOLOv4 algorithm is replaced by MobileNetV3,which significantly enhances the realtime detection performance of the network and the detection performance of small fish targets at the cost of a slight reduction in detection accuracy;channel pruning,and knowledge distillation are performed on the convolution layer of the network structure to compress the model and reduce the number of floatingpoint operations(FLOPs)and the amount of calculation;using optimized Kmeans clustering and DIoU loss function with global nonmaximum suppression to determine the anchor frame,the problem of missing anchor frame caused by mutual occlusion of fish bodies are solved.The experimental results reveal that the model size of the suggested FYOLO algorithm,the average recognition time of each image,the accuracy,the FLOPs,and detection speed in the embedded device are 13.7 MB,50 ms,99.13%,1.64×1010,and 33 frame/s,respectively,which can provide theoretical guidance for the actual fishery breeding.

关 键 词：图像处理 YOLOv4 通道剪枝 知识蒸馏 实时检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]