基于迁移学习和ResNet34的作物病害图像识别方法

Image Recognition Method for Plant Diseases Based on Transfer Learning and ResNet34

摘要: 作为全球主要的农业生产大国，我国长期面临着严峻的病害威胁，农作物病害的准确识别和防治对保障我国粮食安全具有重要意义。而要想科学准确地判断虫害类型并做出有效应对，精准识别病害图像无疑是关键前提。近年来深度学习技术发展迅速，在包括病害在内的识别领域展现出了相较于机器学习等传统方法更高的效率与精度。本文针对农作物病害识别问题，提出基于迁移学习改进ResNet34的识别方法，结合图像增强、特征层冻结、全连接层优化与分层动态学习率调节等方法有效提高了识别精度，并将其部署于在线识别系统。

Abstract: As a major agricultural producer globally, China has long faced severe threats from plant diseases and pests. Accurate identification and control of crop diseases are of great significance for ensuring national food security. To scientifically and accurately determine pest types and implement effective countermeasures, precise identification of disease images is undoubtedly a key prerequisite. In recent years, deep learning technology has developed rapidly and demonstrated higher efficiency and accuracy than traditional methods such as machine learning in recognition fields, including disease identification. This paper addresses the problem of crop disease recognition and proposes an improved ResNet34-based recognition method using transfer learning. By combining image enhancement, feature layer freezing, fully connected layer optimization, and hierarchical dynamic learning rate adjustment, the method effectively improves recognition accuracy and is deployed in an online recognition system.