Abstract: Visionbased detection of agricultural pests is an important technology for achieving automated and realtime monitoring of pest conditions. Firstly, this paper introduces the application of traditional machine learning techniques in the visionbased detection of pests in China and internationally. Then, this paper summarizes the research progress of the new generation of visionbased detection methods for pests, which are based on deep learning techniques such as R-CNN, Fast R-CNN, Faster R-CNN, SSD, and YOLO. Next, this paper analyzes the problems that exist in the research and practical applications of visionbased detection methods for agricultural pests. The traditional machine learningbased methods have problems such as low feature capture ability, detection accuracy, and robustness, as well as high resource consumption. The deep learningbased methods have higher detection performance than the traditional machine learningbased methods but have problems such as poor performance on small and differently distributed targets, low detection accuracy, and slow speed. Finally, this paper discusses possible research directions in the future for the visionbased detection of agricultural pests based on deep learning techniques, including the development of public resources for agricultural pest image data, robust handling for data distribution shifts, deep feature learning, and multiscene applications.

Key words:  , pest monitoring, computer vision, object detection, machine learning, deep learning

摘要： 农业害虫智能视觉检测是实现虫情自动实时监测的重要技术，首先介绍经典机器学习技术在国内外害虫智能视觉检测中的应用，然后整理以R-CNN、Fast R-CNN、Faster R-CNN、SSD和YOLO等深度学习技术为核心的新一代害虫智能视觉检测方法的研究进展。接着，剖析农业害虫智能视觉检测方法在研究及实际应用中存在的问题，其中基于经典机器学习的方法存在特征捕获能力和检测精度较低、资源消耗较大以及鲁棒性较弱等问题；基于深度学习的方法比基于经典机器学习的方法拥有更高检测性能，但存在数据分布不同和目标较小时识别效果较差、检测精度低和速度慢等问题。最后，针对基于深度学习的方法在农业昆虫数据库的制作、数据分布偏移的鲁棒性处理、深度特征学习、多场景应用4个方面对未来研究方向进行展望。

关键词: 虫情监测, 计算机视觉, 目标检测, 机器学习, 深度学习