LDL-MobileNetV3S: an enhanced lightweight MobileNetV3-small model for potato leaf disease diagnosis through multi-module fusion.

Introduction: The timely and precise detection of foliar diseases in potatoes, a food crop of worldwide importance, is essential to safeguarding agricultural output. In complex field environments, traditional recognition methods encounter significant challenges, including the difficulty in extracting features from small and diverse early-stage lesions, blurred edge features due to gradual transitions between diseased and healthy tissues, and degraded robustness from background interference such as leaf texture and varying illumination. Methods: To address these limitations, this study proposes an optimized lightweight convolutional neural network architecture, termed LDL-MobileNetV3S. The model is built upon the MobileNetV3 Small backbone and incorporates three innovative modules: a Lightweight Multi-scale Lite Fusion (LF) module to enhance the perception of small lesions through cross-layer connections, a Dynamic Dilated Convolution (DDC) module that employs deformable convolutions to adaptively capture pathological features with blurred boundaries, and a Lightweight Attention (LA) module designed to suppress background interference by assigning spatially adaptive weights. Results: Experimental results demonstrate that the proposed model achieves a recognition accuracy of 94.89%, with corresponding Precision, Recall, and F1-score values of 93.54%, 92.53%, and 92.77%, respectively. Notably, these results are attained under a highly compact model configuration, requiring only 6.17 MB of storage and comprising 1.50 million parameters. This is substantially smaller than benchmark models such as EfficientNet-B0 (15.61 MB / 3.83 M parameters) and ConvNeXt Tiny (106 MB / 27.8 M parameters). Conclusion: The proposed LDL-MobileNetV3S model demonstrates superior performance and efficiency compared to several existing lightweight models. This study provides a cost-effective and high-accuracy solution for potato leaf disease diagnosis, which is particularly suitable for deployment on intelligent diagnostic devices operating in resource-limited field environments.