RANet: Lightweight Crop Disease Identification Model based on Reparameterization and Attention Mechanism

A lightweight crop disease recognition model called RANet is proposed to address the issues of high parameter count, computational complexity, and poor real-time performance in using convolutional neural networks(CNNs) for crop leaf disease identification. Inspired by MobileNetV2, RANet introduces a reparameterizable basic module called RABlock. The structure of RABlock differs between training and inference, reducing model parameters and computational requirements while maintaining recognition accuracy and improving inference speed. Additionally, RANet incorporates the ECANet attention mechanism to enhance the model’s focus on disease regions, further improving recognition performance. On a custom-built leaf disease dataset, RANet achieves an identification accuracy of 97.98% with only 1.8MB of parameters and a computational complexity of 0.179G. The inference time for a single-leaf disease image is 2.5ms. RANet outperforms lightweight models like MobileNetV3, striking a balance between recognition accuracy and speed. These results provide valuable insights for deploying lightweight deep learning models on edge devices.