基于多尺度残差神经网络的葡萄叶片病害识别

Identification of Grape Leaf Diseases Based on Multi-scale Residual Neural Network

葡萄叶片不同程度的病害具有一定的相似性,目前对于葡萄叶片病害的识别多为病害种类识别,对不同程度病害识别的研究较少,且传统识别方法对于不同程度病害识别准确率较低。提出一种基于多尺度残差神经网络(Multi-Scale ResNet)的葡萄叶片病害识别方法。对葡萄叶片病害图像进行数据增强与叶片区域标注后,使用Mask R-CNN提取葡萄叶片部位,通过引入多尺度卷积以改变ResNet底层对不同尺度特征的响应,利用加入的SENet提升网络的特征提取能力,并将图像输入Multi-Scale ResNet中进行识别。实验结果表明,该方法的平均识别准确率达到90.83%,相比ResNet18提高了2.87个百分点。

Diseases of different degrees in grape leaves show a certain degree of similarity.At present,the grape leaf diseases identification is mostly the identification of disease types,there are few studies on the identification of different degrees of diseases,and the recognition accuracy rate needs to be improved.In this paper,a Multi-Scale ResNet based grape leaf disease recognition method is proposed.Mask R-CNN is utilized to extract grape leaves,and multi-scale convolutions are introduced to improve the response of the underlying ResNet to different scales of features.Then SENet is introduced to enhance the feature extraction capability of the network.Finally,images are input to Multi-Scale ResNet for identification.Experimental results show that compared with that of the original ResNet,the average identification accuracy of the proposed method is improved by 2.87 percentage points,reaching 90.83%.