基于动态注意图的储粮害虫智能识别研究  

Intelligent Identification of Stored-Grain Pest Based on Dynamic Attention Graph

作　　者：杨海英 赵颖 周晓光[2] 王殿轩[3] 李佐勇 Yang Haiying;Zhao Ying;Zhou Xiaoguang;Wang Dianxuan;Li Zuoyong(School of Electronics and Information Engineering,Liaoning University of Technology,Jinzhou 121000;School of Intelligent Engineering and Automation,Beijing University of Posts and Telecommunications,Beijing 100876;School of Food Science and Technology,Henan University of Technology,Zhengzhou 450001;College of Computer and Control Engineering,Minjiang University,Fuzhou 350121)

机构地区：[1]辽宁工业大学电子与信息工程学院,锦州121000 [2]北京邮电大学智能工程与自动化学院,北京100876 [3]河南工业大学粮油食品学院,郑州450001 [4]闽江大学计算机与控制工程学院,福州350121

出　　处：《中国粮油学报》2025年第2期9-16,共8页Journal of the Chinese Cereals and Oils Association

摘　　要：本研究旨在解决实际仓储场景中储粮害虫多物种混合图像的检测识别问题。通过对YOLOv5基础模型架构的改进,提出了一种基于动态注意图的储粮害虫检测识别方法。利用粮堆储粮害虫采集装置获取了仓储害虫不同物种混合的图像数据,并进一步构建了单物种图像数据集与多物种混合图像数据集。本研究运用的方法借助多个动态注意图的有机组合,构建特征之间的长距离依赖关系模型,动态筛选特征图中与当前位置相关性最高的邻接节点,从而增强模型的特征提取能力,提高检测识别精度。在储粮害虫单物种和多物种混合图像的检测识别任务中,该模型的平均精度均值mAP@0.5分别达到了97.2%和91.0%。此外,该模型已成功部署到相关的储粮害虫监测识别系统中,并取得了良好的检测识别效果。The purpose of this paper is to address the issue of detecting and identifying ulti-species mixed images of stored-grain pests in the actual warehouse.A novel detection and recognition model based on dynamic attention graphs was proposed by improving the architecture of the YOLOv5 model.The image data of mixed species of stored-grain pests were obtained using a grain storage pest collection device in grain piles,and further used to construct both single-species image datasets and multi-species mixed image datasets.In this model,dynamic attention graphs were utilized with multiple heads to establish long-range dependencies between features,and the most relevant neighboring nodes were dynamically selected to enhance the feature extraction capability of the pest detection and recognition model,thus improving the accuracy of detection and recognition.In the task of detecting and recognizing single-species and multi-species mixed stored-grain pest images,the mean Average Precision(mAP)@0.5 of this model reached 97.2%and 91.0%,respectively.Furthermore,the model has been successfully deployed into the relevant stored-grain pest monitoring and recognition systems,achieving good detection and recognition results.

关 键 词：单物种图像 多物种混合图像 检测识别 深度学习 动态注意图 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S126[自动化与计算机技术—计算机科学与技术]