基于改进YOLOv5s算法的虾脊兰炭疽病识别  

Identification of Anthracnose in Shrimp Spine Orchid Based on Improved YOLOv5s Algorithm

作　　者：李晓东 李成林 方文博 LI Xiaodong;LI Chenglin;FANG Wenbo(Water Conservancy Project&Civil Engineering College,Tibet Agriculture&Animal Husbandry University;Engineering Technology Center of Civil,Water Conservancy&Electric,Tibet Autonomous Region,Nyingchi Tibet,860000,China)

机构地区：[1]西藏农牧学院水利土木工程学院 [2]西藏自治区土木水利电力工程技术中心,西藏林芝860000

出　　处：《高原农业》2022年第5期484-492,共9页Journal of Plateau Agriculture

基　　金：西藏自治区科技计划项目—西藏高原城镇防洪关键问题与对策研究(XZ201901-GB-13)。

摘　　要：虾脊兰(Calanthe discolor Lindl.)炭疽病严重影响作物品质,必须在种植区进行海量植株的快速准确识别。然而由于背景环境复杂、种植密集与病叶形态的多样,传统的人工及机器学习识别均难于在精度与速度上满足要求。针对这一问题,本文提出了一种改进的虾脊兰炭疽病识别方法。本方法YOLOv5s(You Only Live Once v5s)网络作为基础,引入注意力机制以提升病变部位的识别能力,利用样本变换方法适应多叶片形态的多样性,并针对改进了冗余的边界框的消除机制降低了误判与漏判。在实验中,本文构建了虾脊兰样本数据集作为测试数据,并将本方法与传统的深度目标识别方法进行对比,在测试数据集上平均准确率最高达95.4%,模型存储空间为13.78MB,每秒传输帧数为91f/s。平均准确率比FasterR-CNN、YOLOv3、YOLOv4、YOLOv5l、YOLOv5s分别高出0.97%、8.06%、1.82%、0.58%、1.81%。结果表明本文提出的方法在识别精度、识别速度上均获得了较大的提升,并仅需较小的模型部署,以上特征使得本方法更加适用于虾脊兰炭疽病识别的实际工作。Anthracnose in Calanthe discolor Lindl.seriously affects its crop quality,and it must be identified quickly and accurately in a large number of plants in the growing area.However,due to the complex background environment,dense planting and diverse morphology of diseased leaves,both traditional manual and machine learning recognition are difficult to meet the requirements in terms of accuracy and speed.Aiming at this problem,this paper proposed an improved method for identifying anthracnose in Calanthe discolor Lindl.Based on the YOLOv5s(You Only Live Once v5s)network,this method introduced an attention mechanism to improve the recognition ability of lesion sites,to adapt the diversity of multiple leaf morphologies using a sample transformation method,and reduced false positives and misses by improving the elimination mechanism of redundant bounding boxes.In the experiment,this paper constructed a dataset of anthracnose in Calanthe discolor Lindl.as the test data,and this method was compared with the traditional deep target recognition method.The average accuracy rate on the test dataset was up to 95.4%,and the model storage space was 13.78MB,and the rate of frames transmission was 91f/s.The average accuracy was higher than that of Faster R-CNN,YOLOv3,YOLOv4YOLOv5l,and YOLOv5s by 0.97%,8.06%,1.82%,0.58%,and 1.81%,respectively.The results showed that the proposed method had obtained a large improvement in recognition accuracy,recognition speed,and required only a small model deployment,and the above features made the method more applicable to the practical work of shrimp spine orchid anthracnose recognition.

关 键 词：炭疽病 机器学习 YOLOv5s 注意力 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]