农田害虫图像识别算法发展综述

A Review of the Development of Image Identification Algorithms for Farmland Pests

摘要: 我国是一个农业大国，对于农田害虫的精准识别与国民经济的稳定发展息息相关，也是确保智慧植保和精准给药的核心环节。传统人工识别和主要依赖机器学习的识别方法，效率低，精度低，依赖人工特征提取严重，受外界干扰因素多。基于深度学习的农田害虫识别技术，凭借主干网络的强大特征提取能力，已成为农田害虫识别的主流技术。数据集的分析与处理，经典算法模型的选择与改进，算法优化方向和不同维度场景下的应用实践推动着农田害虫图像识别算法的发展。也可以对比CNN、Transformer、轻量化模型的性能差异性，剖析小样本、实际复杂背景、模型部署等待解决难题，展现多模态与大模型适配发展的趋势，为农田害虫图像识别算法的进一步发展提供参考。

Abstract: Our country is a major agricultural nation. The precise identification of pests in farmland is closely related to the stable development of the national economy and is also a core link in ensuring intelligent agricultural protection and precise drug administration. Traditional manual identification and the main machine learning-based identification methods have low efficiency and accuracy, heavily rely on manual feature extraction, and are subject to numerous external interference factors. The deep learning-based identification technology for farmland pests, leveraging the powerful feature extraction capability of the backbone network, has become the mainstream technology for farmland pest identification. The analysis and processing of data sets, the selection and improvement of classic algorithm models, the optimization direction of algorithms, and the application practice in different dimensions scenarios drive the development of farmland pest image recognition algorithms. Comparisons can also be made between the performance differences of CNN, Transformer, and lightweight models, analyzing the challenges such as small samples, actual complex backgrounds, and model deployment, and presenting the trend of multi-modal and large-model adaptation development, providing references for the further development of farmland pest image recognition algorithms.