基于图像增强和注意力机制的作物杂草识别  

Crop weeds recognition based on image enhancement and attention mechanism

作　　者：曲福恒[1] 李婉婷 杨勇[1,2] 刘红玉 郝忠林 QU Fu-heng;LI Wan-ting;YANG Yong;LIU Hong-yu;HAO Zhong-lin(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China;College of Education,Changchun Normal University,Changchun 130032,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022 [2]长春师范大学教育学院,吉林长春130032

出　　处：《计算机工程与设计》2023年第3期815-821,共7页Computer Engineering and Design

基　　金：国家自然科学基金面上基金项目(41971323);吉林省教育厅科研基金项目(JJKH20181164KJ)。

摘　　要：为提高复杂环境下无人机获取的作物杂草图像识别的准确率,提出一种基于图像增强与注意力机制的作物杂草识别方法。在多尺度Retinex算法中加入颜色恢复函数调节3个通道颜色的占比以恢复其颜色特征,使图像更清晰;将残差网络模型中的激活函数换为Leaky ReLU,加入CBAM注意力机制模块,获取更多有用信息,抑制其它无用信息。实验结果表明,该方法可以提高复杂环境下无人机获取的作物杂草图像的识别准确率,其准确率达到95.3%,高于AlexNet、ResNet18、ResNet50及其它主流算法的识别结果。To improve the accuracy of crop weed image recognition obtained using UAV(unmanned air vehicle)in complex environment,a crop weed recognition method based on image enhancement and attention mechanism was proposed.In the multi-scale Retinex algorithm,a color restoration function was added to adjust the proportion of the colors of the three channels to restore their color features and make the image clearer.The activation function in the residual network model was replaced by Leaky ReLU,and the CBAM attention mechanism module was added to obtain more useful information and suppress other useless information.Experimental results show that the proposed method can improve the recognition accuracy of crop weed images obtained using UAV in complex environment,and its accuracy reaches 95.3%,which is higher than the recognition results of AlexNet,ResNet18,ResNet50 and other mainstream algorithms.

关 键 词：无人机 作物杂草识别 多尺度RETINEX算法 颜色恢复函数 残差网络 Leaky ReLU激活函数 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]