基于轻量级残差网络的植物叶片病害识别

针对基于卷积神经网络的植物叶片病害识别方法存在网络参数众多、计算量大且复杂的问题，结合植物叶片病害特征，提出了一种基于轻量级残差网络（Scale-Down ResNet）的植物叶片病害识别方法。网络基于Residual Network（ResNet）,通过缩减网络卷积核数目和轻量级残差模块（SD-BLOCK）,在大幅减少网络参数、降低计算复杂度的同时保持了低识别错误率，然后加入Squeeze-and-Excitation模块进一步降低识别错误率。在PlantVillage数据集上的实验表明，在网络参数量8×10⁴,计算量MFLOPs为55的情况下，模型识别错误率为0.55%。当参数量达到2.8×10⁵,计算量MFLOPs为176时，模型识别错误率为0.32%,低于ResNet-18,并且参数量约为其1/39,计算量约为其1/10。相比MobileNet V3和ShuffleNet V2,所提网络模型更为轻量，识别错误率更低。同时网络在自建苹果叶片病害数据集上获得了1.52%的低识别错误率。

Plant Leaf Disease Identification Based on Lightweight Residual Network

The plant leaf disease recognition method based on convolutional neural network has the problem of numerous network parameters,large amount of calculation and complexity.To solve this problem,combined with the characteristics of plant leaf diseases,a plant leaf disease recognition method based on lightweight residual network (Scale-Down ResNet) was proposed.The network was based on Residual Network (ResNet),by reducing the number of convolution kernels and the network module of SD-BLOCK,the network parameters and computational complexity were greatly reduced,while the recognition error rate was kept low.Then the Squeeze-and-Excitation module was added to further reduce the recognition error rate.Experiments on the PlantVillage data set showed that when parameters were 8×104 and calculation amout MFLOPs was 55,the recognition error rate of model was 0.55%.When parameters reached 2.8×105 and calculation amount MFLOPs was 176,the recognition error rate of model was 0.32%,which was lower than that of ResNet-18,and the parameter was about 1/39 of ResNet-18 and the amount of calculation was about 1/10 of ResNet-18. Compared with MobileNet V3 and ShuffleNet V2,the proposed network model was lighter and had lower recognition error rate.At the same time,the low recognition error rate of 1.52% was obtained on self built apple leaf disease data set.