摘要
小白菜是中国种植面积较广、深受大众喜爱的蔬菜,真实菜地环境中虫害往往出现在叶片的特定区域,且受环境因素如光照和背景干扰较大,影响对其的智能检测。为提高小白菜虫害的检测效率和准确率,该研究提出一种基于YOLOv5s网络框架改进的YOLOPC(YOLO for Pak Choi)小白菜虫害识别模型。首先,引入CBAM(convolutional block attention module)注意力机制,将其放在CBS(卷积层Convolution+归一化层Batch normalization+激活函数层SILU)的输入端构成CBAM-CBS的结构,动态调整特征图中各个通道和空间位置的权重;使用上采样和1×1卷积操作来调整特征图的尺寸和通道数,实现不同层次特征的融合,增强模型的特征表示能力。同时,改进损失函数,使其更适合边界框回归的准确性需求;利用空洞卷积的优势提高网络的感受野范围,使模型能够更好地理解图像的上下文信息。试验结果表明,与改进前的YOLOv5s模型相比,YOLOPC模型对小白菜小菜蛾和潜叶蝇虫害检测的平均精度均值(mean average precision, mAP)达到91.4%,提高了12.9%;每秒传输帧数(Frame Per Second, FPS)为58.82帧/s,增加了11.2帧/s,增加幅度达23.53%;参数量仅为14.4 M,降低了25.78%。与经典的目标检测算法SSD、Faster R-CNN、YOLOv3、YOLOv7和YOLOv8相比,YOLOPC模型的平均精度均值分别高出20.1%、24.6%、14%、13.4%和13.3%,此外,其准确率、召回率、帧速率和参数量均展现出显著优势。该模型可为复杂背景下小白菜虫害的快速准确检测提供技术支持。
Pak choi(Chinese cabbage)has been one of the most popular vegetables with a wide planting area in China.Rapid detection and accurate identification of Pak choi insect infestation is of great significance to ensure the safety of vegetable supply.However,the insect pests can often appear in specific areas of leaves under the real vegetable field environment.The relatively large light and background interference posed a great challenge to the detection efficiency and accuracy.In this study,an improved YOLOPC model was proposed to identify Pak choi pests using the YOLOv5s network framework.Firstly,the CBAM(Convolutional Block Attention Modul)was introduced to place on the input end of CBS(Convolution layer+normalization layer Batch normalization layer+activation function layer SILU).The structure of CBAM-CBS was formed to dynamically adjust the weights of each channel and spatial position in the feature graph.The upsample and 1×1 convolution operations were used to adjust the size and number of channels of the feature graph,in order to realize the fusion of features at different levels.The feature representation of the model was enhanced at the same time.The loss function was improved more suitable for the accuracy of bounding box regression.The void convolution was used to improve the receptive field range of the network,in order to better learn the context information of the image.Specifically,the improvement strategy included the following three aspects:1)The attention mechanism of space and channel was added to extract the network feature,in order to better learn the cabbage diamondback moth and leaf miner insect pests;2)The alpha-IoU loss function was used to replace the CIoU one in YOLOv5s.Different levels of boundary box regression accuracy were adapted for the insect targets of cabbage,broccoli moth,and leaf leaf-divers at different scales and aspect ratios;3)Atrous Spatial Pyramid Pooling(ASPP)was introduced to improve the receptive field range of the network,in order to better learn the context information
作者
郑俊键
兰玉彬
熊万杰
李硕
杨润娜
董昕
ZHENG Junjian;LAN Yubin;XIONG Wanjie;LI Shuo;YANG Runna;DONG Xin(College of Electronic Engineering(College of Artificial Intelligence),South China Agricultural University,Guangzhou 510642,China;National Center for International Collaboration Research on Precision Agricultural Aviation Pesticides Spraying Technology(NPAAC),Guangzhou 510642,China;Guangdong Smart Agriculture Engineering Technology Research Center,Guangzhou 510642,China;National S&T Innovation Center for Modern Agricultural Industry(Guangzhou)510520,China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2024年第13期124-133,共10页
Transactions of the Chinese Society of Agricultural Engineering
基金
高等学校学科创新引智计划资助项目(D18019)
广东省重点领域研发计划项目(2019B020214003)。
