基于改进MobileNet-V2的轻量化苹果叶片病害识别

Lightweight Apple Leaf Disease Recognition Based on Improved MobileNet-V2

摘要:

针对苹果叶片病害识别中计算资源消耗与识别准确度之间不平衡问题，提出了一种改进MobileNet-V2卷积神经网络模型。采用数据增强技术扩展图像样本的多样性，增加了模型泛化能力;引入迁移学习策略，通过预训练和冻结部分层参数，减少了训练时间和计算资源的消耗;通过进一步引入分组卷积、通道注意力机制中的压缩与激励模块（Squeeze-and-excitation, SE）和权重剪枝策略，优化了模型特征提取能力，提高了模型计算效率和准确率。实验结果表明，模型优化后，准确率为99.1%，较原Mobile-V2模型高5.9个百分点，较ResNet50、VGG16和Xception等传统卷积神经网络模型分别提升4.6、9.4、4.0个百分点。模型平均精确率为98.9%，平均召回率为98.75%，F1分数为98.82%，参数量仅为4.06×10?。模型在实际部署上也能实现快速、准确的病害检测，单幅图像平均推理时间为507 ms，具备实用价值和推广潜力，为农业智能化发展提供了参考。

Abstract:

An improved MobileNet-V2 convolutional neural network model was proposed, aiming at addressing the balance between computational resource consumption and recognition accuracy in apple leaf disease identification. Firstly, data augmentation techniques were employed to expand the diversity of image samples, thereby enhancing the model's generalization ability. Next, a transfer learning strategy was introduced, where pretraining and freezing some layer parameters reduced training time and computational resource consumption. The model further optimized feature extraction capability by incorporating group convolutions, a squeeze-and-excitation (SE) module from the channel attention mechanism, and weight pruning strategies, which enhanced both computational efficiency and accuracy. Experimental results demonstrated that, after optimization, the model achieved an accuracy of 99.1%, surpassing the value of the original MobileNet-V2 model by 5.9 percentage points, and outperforming traditional convolutional neural network models, like ResNet50, VGG16, and Xception by 4.6, 9.4, and 4.0 percentage points, respectively. The model's average precision was 98.9%, average recall was 98.75%, and F1 score was 98.82%, with only 4.06×10? parameters. In practical deployment, the model achieved fast and accurate disease detection, with an average inference time of 507 ms per image, demonstrating its practical value and potential for widespread application, offering insights for the development of smart agriculture.