基于改进ResNet-50与迁移学习的苹果叶片病害的图像识别

Image Recognition of Apple Leaf Disease Based on Improved ResNet-50 and Transfer Learning

为解决下述问题,在自然场景中获取的苹果叶病害图像大多包含复杂的背景,同时,由于拍摄要求不那么严格,疾病特征更有可能出现在图像中的任意位置,所有这些因素都会影响卷积神经网络的识别精度,且随着网络加深其性能提升困难。提出了一种基于改进ResNet-50的苹果叶片病害识别方法。首先引入SE(squeeze-and-excitation)注意力机制,改进残差模块,在一定程度上去除复杂背景等干扰信息,另一方面降低模型对特征定位的过度敏感度,使模型能够学习更重要的疾病特征,减少复杂背景等信息的干扰,然后加入Dropout层抑制过拟合改进模型结构,最后采用迁移学习的训练方式加快模型的收敛速度。在苹果叶病数据集上的实验结果表明,改进模型的准确率达到98.35%,较ResNet-50提高5%。与其他一些传统的卷积神经网络相比,该模型收敛速度更快,具有更高的识别精度,能够较好地识别苹果叶片病害,适用于自然场景下获取的苹果叶病图像等优点,具有较强的实用性。

In order to solve the following problems,most of the apple leaf disease images obtained in natural scenes contain complex backgrounds.At the same time,because the shooting requirements are not so strict,the disease features are more likely to appear in any position in the image.All these factors will affect the recognition accuracy of convolutional neural networks,and its performance will be difficult to improve with the deepening of the network.An improved method of apple leaf disease identification based on ResNet-50 was proposed.First,the squeeze-and-excitation(SE)attention mechanism was introduced,and the residual module was improved to remove interference information such as complex background to a certain extent.On the other hand,the models over-sensitivity to feature localization was reduced,so that the model can learn more important disease characteristics and reduce interference of complex background information.Finally,transfer learning training was used to accelerate the convergence of the model.The experimental results on the apple leaf disease dataset show that the accuracy of the improved model reaches 98.35%,which is 5%higher than that of ResNet-50.Compared with other traditional convolutional neural networks,this model has the advantages of faster convergence speed,higher recognition accuracy,better identification of apple leaf disease,and is suitable for apple leaf disease images obtained in natural scenes,and has strong practicability.