Abstract
To address the challenges of high computational complexity and difficult deployment of existing deep learning models on mobile devices for grape leaf disease diagnosis, this paper proposes a lightweight image classification algorithm named MobileNet-GDR (Grape Disease Recognition), built upon the MobileNetV4-small architecture. The algorithm constructs an efficient feature extraction module based on depthwise separable convolutions and grouped convolutions to optimize the feature fusion process, while incorporating PReLU activation functions to enhance nonlinear representation capability. Experimental results on a grape leaf disease dataset demonstrate that MobileNet-GDR achieves high accuracy while significantly reducing computational overhead: with only 1.75M parameters and 0.18G FLOPs, it attains real-time inference speed of 184.89 FPS and a classification accuracy of 99.625%. Ablation studies validate the effectiveness of each module, and comparative experiments show that its computational efficiency surpasses mainstream lightweight models such as FasterNet and GhostNet. MobileNet-GDR provides a practical lightweight solution for real-time disease diagnosis in field conditions, demonstrating significant value for agricultural applications.