基于视觉自注意力模型的苗期玉米与杂草检测方法  

Detection method of maize and weeds at seedling stage based on visual self-attention model

作　　者：唐伯青 赵大勇 熊锋[1,2] 李德强 TANG Boqing;ZHAO Dayong;XIONG Feng;LI Deqiang(Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院沈阳自动化研究所,辽宁沈阳110016 [2]中国科学院机器人与智能制造创新研究院,辽宁沈阳110169 [3]中国科学院大学计算机科学与技术学院,北京100049

出　　处：《南京农业大学学报》2024年第4期772-781,共10页Journal of Nanjing Agricultural University

基　　金：中国科学院战略性先导专项(XDA28040400)。

摘　　要：[目的]识别作物和杂草是农业智能化中自动除草的关键步骤。本文旨在解决作物与杂草识别精度低、检测模型实时性和鲁棒性差等问题。[方法]以叶龄处于3~8叶期的玉米及其伴生杂草为研究对象,提出一种高效准确的玉米苗与杂草的检测方法。该方法以实时端到端目标检测视觉自注意力模型为基础框架,用小尺度卷积等效替代大尺度深度卷积的思想,以较小的精度损失降低推理耗时。引入一种包含上下文信息的自顶向下注意力机制,强化模型对小目标的检测效果。应用组合图像增强策略,提升模型精度与泛化能力。[结果]改进后模型的平均检测精度为90.11%,推理阶段单张图片耗时33.67 ms,模型参数量44.86 MB。改进后的模型比主流目标检测模型总体精度更高,且推理速度快。[结论]所提方法对于玉米苗与伴生杂草的整体检测性能优秀,能够提高杂草识别的准确性和效率。[Objectives]Identifying crops and weeds are crucial aspects of advancing intelligent and automated weeding.This article aimed to improve the accuracy of crop and weed identification,to enhance the real-time performance of detection models,and to enhance robustness.[Methods]Focusing on maize crops and their corresponding weeds in the leaf age range of 3-8 leaves,this research endeavored to devise a detection method for maize seedlings and associated grasses.The seedling detection method leveraged an improved real-time end-to-end object detection with transformers(RT-DETR)for maize and weed detection in field conditions.The novel concept of replacing large-scale deep convolution with small-scale convolution equivalence within RT-DETR was introduced,reducing training complexity and inference time while maintaining detection accuracy.Furthermore,a self-attention mechanism with contextual information was integrated to enhance target attention and improve small target detection.Additionally,a combined image enhancement strategy was employed to enhance model accuracy and generalization.[Results]The improved model effectively distinguished weeds from crops in complex field scenarios,achieving an average detection accuracy of 90.11%.In the inference stage,each image took 33.67 ms for processing,with a model size of 44.86 MB.Compared with the mainstream target detection model,the improved model had higher overall accuracy and fast speed.[Conclusions]The proposed method had excellent overall detection performance for corn seedlings and associated weeds,which could improve the accuracy and efficiency of weed identification.

关 键 词：玉米 杂草 检测 实时视觉自注意力模型 等效卷积 图像增强 

分 类 号：S24[农业科学—农业电气化与自动化] S513[农业科学—农业工程]