A significant increase in agricultural production is highly demanding, due to the ever-increasing human population over the last decades. However, the crop yield is greatly affected by various plant diseases. Timely and accurate identification of leaf disease is very necessary for plant disease control. In this study, 38 types of leaf images (from 14 different crops) were used for the identification. These images were collected from the PlantVillage project (an open-source leaf disease database). Since the convolutional neural network (CNN) can automatically learn appropriate features from training data, CNN has become one of the most popular ways for image identification, better than traditional machine learning using manual feature extraction. However, the number of parameters in a CNN model was very huge, leading to a very heavy computation load. Thus, the traditional CNN was difficult to apply in real time measurement. Three methods were proposed to reduce the size of a CNN model and the computation load. The first method was to replace some of the 3×3 convolution filters with 1×1 convolution filters. A 3×3 convolution filter included 9 parameters, requiring 9 floating-point multiplications to obtain one solution. If this 3×3 convolution filter was replaced by a 1×1 convolution filter, the number of parameters and the number of required multiplications could reduce to 1. Thus, this method could help to reduce the size of a model and its computation load. The second method was to move the convolution calculation from large- to small-size feature maps. The idea was to reduce the computation load at the cost of a slightly performance drop. This can be done by adjusting the position of convolution module in the CNN. If a 3×3 filter was convolute with an 2N×2N feature map, the total number of floating-point multiplication was 36N2. If the same 3×3 filter was convolute with a N×N feature map, the total number of floating-point multiplication was 9N2, only a quarter of previous computation load. Therefore, if the convolution was performed with a smaller size feature map, the amount of calculation greatly reduced. The premise of this modification was that the performance drop was very small. The third method was to reduce the depth of a CNN model. It was obvious that a very deep neural network was not needed for a relatively simple task. Therefore, the idea behind the third method was to use a suitable network instead of a very complicated neural network for a relatively simple task. The experiment showed that the performance drop in the SqueezeNet was only around 0.5% for a classification task in the 38 types of images, if the last 3 fire modules were removed. 5 improved CNN models were proposed for leaf disease identification. The experiments showed that the size of the optimal model was around 0.62MB, and the computation load of this model was only 111 MFLOPs. Specifically, the average accuracy rate was 98.13%, and the average recall rate was 98.09%, while the average precision rate was 97.62%, showing higher cost efficiency than before. The proposed model can greatly reduce the size of the model, while reducing the computation load, and only a slight decrease in performance. There was an excellent balance on the performance, model size, and computation load. The improved model can be suitable for deployment on mobile terminals and other embedded resource-constrained devices, thereby contribute to real-time and accurate identification of crop diseases. © 2021, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.
