基于改进轻量级MobileNetV2的辣椒病虫害图像识别

Identification of Pepper Diseases and Pests Based on Improved Lightweight MobileNetV2

摘要: 针对真实环境下辣椒病虫害识别准确率不高以及深度卷积网络参数多、模型内存大等问题，本文提出了一种基于改进MobileNetV2的辣椒病虫害图像识别算法。首先在基线模型的基础上引入通道注意力和空间注意力机制，提升模型对特征信息的敏感度；同时将L2正则化加入到损失函数中，平滑损失函数的梯度，以缓解模型过拟合。最后实验结果表明，改进后模型识别准确率达到94.43%，相较于基线模型，改进后模型精确率提升了4.38%，召回率提升了3.38%，F1值提升了3.88%，同时改进后模型参数量仅为2.43 M，与基线模型相比只增加了约0.2 M。本研究方法在保持较低模型参数量的同时达到了较高的识别准确率，具有一定的优势。

Abstract: To address the challenges of low recognition accuracy for pepper diseases and pests in real-world environments, as well as the large parameter size and high memory consumption of deep convolutional networks, this paper proposes an improved MobileNetV2-based image recognition algorithm for pepper disease and pest identification. First, channel attention and spatial attention mechanisms are introduced into the baseline model to enhance its sensitivity to feature information. Additionally, L2 regularization is incorporated into the loss function to smooth the gradient and mitigate model overfitting. Experimental results demonstrate that the improved model achieves an accuracy of 94.43%, with a 4.38% increase in precision, a 3.38% improvement in recall, and a 3.88% boost in F1-score compared to the baseline model. Furthermore, the enhanced model has only 2.43 M parameters, an increase of merely 0.2 M over the original model. The proposed method maintains a low parameter count while achieving high recognition accuracy, demonstrating significant advantages in practical applications.