基于改进残差网络的苹果叶片病害识别研究

Apple Leaf Diseases Identification Based on Improved Residual Network

苹果叶片病害形态相似、斑点大小不同,依靠人工和农业专家识别的传统方式效率较低。为此提出一种基于改进残差网络的苹果病害识别模型REP-ResNet。该模型在基准模型ResNet-50的基础上通过采用批标准化、激活函数、卷积层的残差结构顺序,加入通道注意力机制和并行卷积的方式进行改进。训练过程中,将公开数据集PlantVillage预训练的模型权重参数迁移至上述网络模型中重新训练,达到加快网络的收敛速度和提高模型识别能力的目的。采用数据扩充的方式解决训练过程中样本不均的问题。结果表明,REP-ResNet模型与基准网络模型相比识别准确率提高2.41个百分点。模型使用迁移学习的方式进行训练,在复杂背景下的苹果叶片病害识别中准确率达到97.69%,与传统卷积神经网络相比识别效果有较大提高。

Apple leaf diseases are similar in morphology and different in spot size,and traditional methods relying on labor and agricultural experts to identify them are inefficient.In view of this,this study proposed an apple disease identification model REP⁃ResNet based on improved residual network.This model was improved by adopting the order adjustment of residual structure(that was bath normalization,activation function,convolutional layer),and adding the efficient channel attention and parallel convolution based on the baseline model ResNet⁃50.During the training process,the model weight parameters pretrained in the public dataset PlantVillage were transferred to the above network model for retraining,which aimed to accelerate the convergence speed of the network and improve the recognition ability of the model.Data expansion was used to solve the problem of uneven samples during training.The results showed that the recognition accuracy of the REP⁃ResNet model was 2.41 percentage points higher than that of the benchmark network model.The model was trained by transfer learning,and the accuracy rate of apple leaf disease recognition in complex backgrounds reached 97.69%.Compared with traditional convolutional neural networks,the recognition effect was greatly improved.