基于YOLOv5的荔枝果实小目标检测算法研究  

Research on Small Target Detection of Litchi Fruit Based onYOLOv5

作　　者：王萍叶[1,2] 毛亮 Wang Pingye;Mao Liang(Shenzhen Polytechnic,Shenzhen Guangdong 518055,China;Guangdong-Hong Kong-Macao Greater Bay Area Artificial Intelligence Application Technology Research Institute,Shenzhen Guangdong 518055,China)

机构地区：[1]深圳职业技术学院,深圳518055 [2]粤港澳大湾区人工智能应用技术研究院,深圳518055

出　　处：《山西电子技术》2023年第4期74-77,共4页Shanxi Electronic Technology

基　　金：广东省教育厅重点领域专项(2021ZDZX1091);广东省农村科技特派员项目(KTP20200219);广东省农村科技特派员项目(KTP20210199)。

摘　　要：小目标检测是目标检测领域中的热点和难点。现有的主流目标检测算法大多使用特征金字塔来检测不同尺度的目标,其中小尺度的高层特征用来检测大目标,大尺度的浅层特征用来检测小目标。然而,对小目标检测而言,高层特征的生成和检测往往会带来大量的计算量,又不能等效地提高小目标检测的准确度。因此,在YOLOv5模型的网络结构基础上,去掉金字塔特征融合的FPN+PAN部分,采用浅层网络来检测自然场景中荔枝果实小目标。在极大地简化YOLOv5模型网络结构的情况下,既保证识别荔枝果实目标的准确度基本不变,又提高了检测速度。Small object detection is the hot and difficult field in object detection area.The mainstream object detection algorithms includes feature pyramids to detect different scale objects.Small-scale high-level features are used to detect large objects,and large-scale shallow features are used to detect small objects.For small target detection,the generation and detection of high-level features bring lots of computation and cannot equivalently improve the accuracy of small target detection.This paper simplifies the network structure of the YOLOv5 model,removes the FPN+PAN part of the pyramid feature fusion,and uses a shallow network to detect small target lychees.The simplified model of YOLOv5 can improve the detection speed and the accuracy of lychee detection is basically unaffected.

关 键 词：小目标检测 深度神经网络 YOLOv5 荔枝果实 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]