面向边缘部署的苹果叶部病害轻量化诊断方法

Lightweight Diagnostic Method for Apple Leaf Diseases Oriented to Edge Deployment

摘要:

针对高性能苹果叶部病害诊断模型难以在边缘设备高效部署的问题，提出一种病灶感知的渐进式压缩框架（L-EAFP）。首先适配一个高性能跨域模型（EAFP-Med ST）作为教师模型，通过病灶加权剪枝和多目标病灶特征锚定蒸馏，将知识高保真迁移至超轻量级学生模型（ShuffleNetV2），最后结合自适应量化完成压缩。在包含9 263幅图像的苹果叶部病害数据集上实验结果表明，所提框架将模型参数量从3.157×10?压缩至3.500×10?（压缩率为98.9%），压缩后模型诊断准确率仍保持98.58%。病灶感知压缩策略在模型轻量化与诊断性能保真度之间取得了良好平衡，为高性能农业AI模型在边缘端的部署提供了有效方案。

Abstract:

Aiming to address the critical bottleneck whereby high-performance convolutional neural networks for apple leaf disease diagnosis are difficult to deploy efficiently on resource-constrained edge devices due to high computational complexity, a novel lesion-aware progressive compression framework (L-EAFP) was proposed. The methodology began by adapting a robust high-performance cross-domain model, EAFP-Med ST, to function as a teacher network. To facilitate high-fidelity knowledge migration to an ultra-lightweight ShuffleNetV2 student model, the framework introduced a lesion-weighted pruning algorithm that prioritized the preservation of pathological feature channels while rigorously eliminating redundancy. Concurrently, a multi-objective lesion feature anchoring distillation technique was applied to strictly align the student model's attentional focus on disease regions. Finally, an adaptive quantization mechanism was incorporated to further compress the model bit-width. Experimental validations performed on a comprehensive dataset of 9 263 apple leaf images demonstrated the framework's effectiveness. The results revealed that L-EAFP reduced the model parameter count from 3.157×10? to 3.500×10?, achieving a remarkable compression rate of 98.9%. Crucially, despite this drastic reduction, the compressed model retained a diagnostic accuracy of 98.58%. These findings indicated that the proposed lesion-aware strategy successfully achieved an optimal balance between model lightweighting and diagnostic performance fidelity, providing a highly effective technical solution for the deployment of advanced agricultural AI models on mobile and embedded edge terminals.