Recognizing Crop Diseases and Pests Using an Improved MobileNetV3 Model

Crop diseases and pests contribute to reduced crop yields and are often difficult for humans to accurately and quickly identify. The performance of plant pathologists or farmers is heavily reliant on their experience. To address this issue, we propose a crop disease and pest recognition method using a lightweight neural network and transfer learning. First, we analyze the data distribution of the training set and use label shuffling and mixed sampling methods to balance the data. Second, we pretrain MobileNetV3 on the ImageNet dataset and transfer the parameter values to the crop disease and pest recognition model for fine-tuning. Third, we improve the activation function and loss function of MobileNetV3, resulting in the creation of MobileNetV3+. This model uses a 6-level-LeakyReLU activation function to replace the ReLU activation function of MobileNetV3, which reduces the loss of negative feature information in crop pest images and improves feature extraction integrity. Additionally, we use focal loss instead of cross-entropy loss to improve the model's ability to handle imbalanced sample distributions. Finally, we train four models (VGG16, ResNet50, MobileNetV3, and MobileNetV3+) using the same method. Experimental results show that MobileNetV3+ has the best overall pest recognition results. Our proposed method can rapidly recognize crop diseases and pests with significantly fewer model parameters and memory requirements, as experimentally verified.