基于改进的YOLOv4的目标识别研究  

Research on target recognition based on improved YOLOv4

作　　者：徐翔 蔡茂国[1] 唐剑兰 XU Xiang;CAI Mao-guo;TANG Jian-lan(School of Electronics and Information Engineering,Shenzhen University,Shenzhen 518000,Guangdong Province,China)

机构地区：[1]深圳大学电子与信息工程学院,广东深圳518000

出　　处：《信息技术》2022年第12期107-111,117,共6页Information Technology

摘　　要：目标识别与检测作为模式识别领域的一种典型应用,如何快速准确地进行目标识别一直是个重要的研究课题。在深度学习算法中,YOLOv4和R-CNN具有出色的目标检测性能,为了改进目标识别中小目标的实时检测,提出了改进的YOLOv4目标检测算法。使用K-means聚类算法设计先验框,用于适应不同的中小型规模;根据中小型标记物体的大小提取一个特征层,并融合四个不同的特征层进行检测;将Mish激活函数应用于检测模型的颈部,取代泄漏的ReLU激活函数,以提高检测性能。实验结果表明,改进后的算法可有效提高检测精度。Target recognition and detection is a typical application in the field of pattern recognition,and how to quickly and accurately recognize targets has always been an important research topic.In the deep learning algorithm,YOLOv4 and R-CNN has achieved excellent target detection performance.In order to improve the real-time detection of small and medium targets in target recognition,an improved YOLOv4 is proposed.The model uses the K-means clustering algorithm to design a priori box to adapt to different small and medium-sized scales;According to the size of small and medium-sized labeled objects,a feature layer is extracted and four different feature layers are merged for detection;The Mish activation function is applied to the neck of the detection model to replace the leaked ReLU activation function to improve the detection performance.Experiment results show that the improved algorithm can effectively improve the detection accuracy.

关 键 词：深度学习 目标检测 YOLOv4算法 小目标检测 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]