深度学习在小麦害虫分类研究中的应用

Application of Deep Learning to Wheat Pest Classification Study

【目的】基于深度学习的小麦害虫高效识别方法有助于害虫的及时防控,极大程度地保障粮食质量安全。【方法】首先,在IP102小麦害虫数据集的基础上,提出结合增广和Real-ESRGAN超分辨率增强的技术方案,重新制作了高质量小麦害虫数据集:IP-AugESRWheat,有效突破了小麦害虫数据集面临的类别不平衡、规模小、分辨率低的局限。其次,提出轻量高效的ECA-EffV2模型,增强模型对小麦害虫的特征提取能力。【结果】改进前的准确率为72.5%,参数量为21.46 M,改进后准确率达到94.8%,参数量降到17.76 M。【结论】提出的协同增广增强的技术策略及高效轻量的模型为小麦害虫图像识别任务提供了有效的技术方法和数据支撑,对可持续小麦生产和农业生态发展具有重要价值。

【Objective】Efficient identification method of wheat pests based on deep learning is important for timely prevention and control of pests and can greatly guarantee the quality and safety of grain.【Method】First,based on the IP102 wheat pest dataset,a technical solution combining the augmentation method and Real-ESRGAN super-resolution enhancement was proposed to recreate a high-quality wheat pest dataset:IP-AugESRWheat.This method effectively broke through the limitations of imbalance of categories,small scale,and low resolution faced by the wheat pest dataset.Secondly,a lightweight and efficient ECA-EffV2 model was proposed to enhance the feature extraction ability of the model for wheat pests.【Result】The accuracy of original method was 72.5%and the number of parameters was 21.46 M.While the accuracy of improved method reached 94.8%and the number of parameters was reduced to 17.76 M.【Conclusion】The proposed technical strategy of synergistic augmentation and enhancement and efficient and lightweight model in this study provided an effective technical approach and data support for the wheat pest image identification task,which will be of great value for sustainable wheat production and agroecological development.