基于多尺度注意力残差网络的桃树害虫图像识别  

Image Recognition of Peach Pests Based on Multi-scale Attention Residual Network

作　　者：类成敏 牟少敏[1] 孙文杰 崔恩泉 LEI Cheng-min;MU Shao-min;SUN Wen-jie;CUI En-quan(College of Information Science and Engineering/Shandong Agricultural University,Tai’an 271018,China)

机构地区：[1]山东农业大学信息科学与工程学院,山东泰安271018

出　　处：《山东农业大学学报（自然科学版）》2022年第2期253-258,共6页Journal of Shandong Agricultural University：Natural Science Edition

摘　　要：自然场景下拍摄的桃树害虫图像,不同种类的害虫个体之间存在尺寸大小差异以及害虫颜色与背景颜色相近的问题,影响害虫图像识别精度。针对以上问题,本文提出了一种基于多尺度注意力残差网络的桃树害虫图像识别模型。首先,将残差网络的第一层普通卷积替换为多尺度卷积,缓解了大卷积核对于小尺寸目标特征的不敏感性,增强多尺度害虫特征提取能力。其次,在残差结构中加入注意力机制选择性内核卷积单元,它通过自适应调整感受野重点提取害虫信息,产生有效感受,抑制背景干扰问题。实验结果表明,本文提出的模型识别准确率为93.27%,取得了较好的识别效果。The peach tree pest image taken in the natural scene has the problem that the pest color is similar to the background color,and the individual sizes of different kinds of pests are different,which affects the recognition accuracy of the pest image.To solve the above problems,this paper proposes a peach pest image recognition model based on multi-scale attention residual network.Firstly,the first layer of ordinary convolution of residual network is replaced by multi-scale convolution,which alleviates the insensitivity of large convolution kernel to the characteristics of small-scale targets and enhances the ability of multi-scale pest feature extraction.Secondly,the selective kernel convolution unit of attention mechanism is added to the residual structure,which extracts pest information by adaptively adjusting the receptive field to produce effective perception and suppress background interference.The experimental results show that the recognition accuracy of the model proposed in this paper is 93.27%,and a good recognition effect is achieved.

关 键 词：残差网络 桃树害虫 图像识别 

分 类 号：TP151.1[自动化与计算机技术—控制理论与控制工程]