基于多尺度注意力U-Net的结球甘蓝青虫检测方法  

Detection of cabbage worms based on multi-scale attention U-Net

作　　者：齐国红 许新华 师晓丽 QI Guo-hong;XU Xin-hua;SHI Xiao-li(School of Electronic Information Engineering,Zhengzhou SIAS University,Zhengzhou 451150,China)

机构地区：[1]郑州西亚斯学院电子信息工程学院,河南郑州451150

出　　处：《江苏农业学报》2023年第6期1349-1357,共9页Jiangsu Journal of Agricultural Sciences

基　　金：国家自然科学基金项目(62172338);河南省科技厅项目(232102110274、222102210122);河南省高等学校重点科研项目(23B510004、22B520049);河南省教育厅2022年民办普通高等学校学科专业建设资助项目[教办政法(2022)219号]。

摘　　要：针对结球甘蓝青虫姿态多样、形状不规则以及传统U-Net对多尺度图像检测的鲁棒性较差等问题,本研究提出一种基于多尺度注意力U-Net(MSAU-Net)的结球甘蓝青虫检测方法。该方法将多尺度空洞Inception和注意力引入到U-Net,通过设置不同膨胀率的初始卷积层卷积核和全局池化层类型,提取多尺度深层次的结球甘蓝青虫检测特征。首先,对原始图像进行超像素聚类,极大减少结球甘蓝青虫图像的基元数量;其次,利用多尺度空洞U-Net提取不同大小的结球甘蓝青虫特征;最后,通过注意力连接将MSAU-Net同层的浅层、深层特征拼接,得到结球甘蓝青虫图像的关键特征,加快网络训练。MSAU-Net方法在结球甘蓝青虫数据集上的平均检测精度为95.26%,较U-Net方法提高了约6个百分点。MSAU-Net方法能较好地检测到大小不同的结球甘蓝青虫,能够应用于结球甘蓝青虫自动检测系统。Aiming at the problems of diverse postures,irregular shapes,and the low robustness of traditional U-Net to multi-scale image detection,a detection method of cabbage worms based on multi-scale attention U-Net(MSAU-Net)was proposed.This method introduced multi-scale dilated Inception and attention into U-Net,and extracted multi-scale high-level worm detection features by setting initial convolution layer convolution core and global pooling layer types with different expansion rates.First,the original image was clustered by superpixel cluster algorithm to greatly reduce the number of primitives of worm images.Then,the multi-scale dilated U-Net was used to extract the features of different sizes of worms.Finally,the shallow and deep features of the same layer of MSAU-Net were spliced through attention connection to obtain the key features of worm images and speed up network training.The average detection accuracy on the cabbageworm dataset was 95.26%,which was about six percentage points higher than that of U-Net.The MSAU-Net method can detect cabbageworms with different sizes,and can be applied to the automatic detection system of cabbage worms.

关 键 词：结球甘蓝青虫检测 超像素聚类 U-Net 多尺度注意力U-Net 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S432[自动化与计算机技术—计算机科学与技术]