Abstract: In view of the fact that there are few studies on the identification of diseases of muskmelon leaves in different degrees, the manual detection has poor real-time performance and the identification accuracy is low, a method for identification of muskmelon leaf diseases based on the improved residual network model is proposed. The traditional ResNet50 model is used as the backbone network, and the ReLU activation function is replaced by the ELU activation function. The 7×7 convolution kernel in the first layer convolution of the ResNet50 model is replaced with the Incption structure, in the fully connected layer, then the Dropout layer is added to enhance the expressive ability of the model and alleviate the overfitting problem. The multi-head self-attention (MHSA) mechanism is introduced to improve the generalization ability of the model. This paper carries out data preprocessing, divides the ratio of training set and test set into 7∶3, and uses data enhancement to expand the small sample data set. The experimental results show that the accuracy rate of the improved residual network model is 1.03 percentage points higher than the original model, the recognition accuracy rate reaches 98.72%, and the model parameter size is 19.3 MB. Compared with other network models, the accuracy rate has been greatly improved, which can provide a reference for efficient identification and timely prevention and treatment of melon leaf diseases.

Key words: melon leaf diseases, image recognition, residual network, multi-head self-attention (MHSA) mechanism, deep learning

摘要： 针对甜瓜叶片不同程度的病害识别研究较少，人工检测实时性差且存在识别准确率较低等问题，提出了一种基于改进残差网络模型的甜瓜叶片病害识别方法。将传统的ResNet50模型作为骨干网络，将ReLU激活函数替换为ELU激活函数；将ResNet50的模型的第一层卷积中的7×7卷积核替换成Incption结构，在全连接层之后加入Dropout层，增强模型的表达能力并缓解过拟合问题；引入多头自注意力机制（MHSA），提高模型的泛化能力。进行数据预处理，将训练集与测试集的比例划分为7∶3，采用数据增强的方式对小样本数据集进行扩充。实验结果表明：改进的残差网络模型准确率与原模型相比提高了1.03个百分点，识别准确率达到98.72%且模型参数量为19.3?MB。与其他网络模型相比准确率大幅提升，可以为甜瓜叶片病害的高效识别和及时预防治理提供参考。

关键词: 甜瓜叶片病害, 图像识别, 残差网络, 多头自注意力机制, 深度学习