基于改进YOLOv5的整车原木数量检测方法  

Estimation of volume for logs on a vehicle using improved YOLOv5 algorithm

作　　者：余鸿晖 郑积仕 张世文 周文刚 孔令华 丁志刚 杨水保 YU Honghui;ZHENG Jishi;ZHANG Shiwen;ZHOU Wengang;KONG Linghua;DING Zhigang;YANG Shuibao(School of Transportation,Fujian University of Technology,Fuzhou 350118,China;Fujian Jinsen Foresty Co.Ltd.,Sanming 353300,China;School of Mechanical and Automotive Engineering,Fujian University of Technology,Fuzhou 350118,China)

机构地区：[1]福建工程学院交通运输学院,福州350118 [2]福建金森林业股份有限公司,三明353300 [3]福建工程学院机械与汽车工程学院,福州350118

出　　处：《林业工程学报》2022年第4期135-143,共9页Journal of Forestry Engineering

基　　金：福建省科技厅自然基金(2019J01773);福建省金森林业股份有限公司校企合作项目(GY-H-20154);福建省林业科技项目(2021FKJ06)。

摘　　要：针对整车原木检测中存在密集小目标难检测、原木被遮挡易被漏检、目标检测模型复杂度太大难以部署等问题,提出一种基于YOLOv5的整车原木数量检测方法TWD-YOLOv5,来探究目标检测在整车原木场景下进行快速精准检测的可行性,从而实现智能检尺,提高检尺效率。本研究在原始YOLOv5模型的基础上通过修改模块数量、加入注意力机制和Transformer模块的操作来优化主干网络,结合新的检测尺度与基于Ghost卷积设计特征融合网络,提升网络检测小目标的能力,降低模型复杂度,选用CIoU作为边界框回归的损失函数和DIoU-nms作为边界框筛选算法,提高边框的回归精度和解决物体被遮挡的问题。本研究算法TWD-YOLOv5进行4组试验,从平均精度均值(mAP)、每秒传输帧数、原木真检率多种尺度指标进行评估,同时通过预测框完成对原木根数的计数。试验结果表明,本研究方法的mAP达到0.731,每秒传输帧数为7.33,模型参数降低了40.5%,且测试集原木真检率达到了99.551%,误检率为0.22%。该方法不仅大幅减少了模型复杂度,还保持了较高的检测精度。本研究的模型能对整车原木场景下的原木有良好的检测效果,解决了原木被遮挡的问题,且检测速度快,能做到实时检测,另外该算法有较强的鲁棒性且模型较小,可以满足部署至移动端进行目标检测的轻量化需求。Real time detection of an object is a new technique to estimate its volume using video, YOLO that refers to “You Only Look Once” is one of the most versatile object volume detection models. YOLO is the first choice by data scientists and machine learning engineers for every real-time object volume detection work. To solve the problems of the real time accurate estimation of wood volume for the small logs on a vehicle, the fifth version of YOGO, i.e., YOLOv5(TWD-YOLOv5) algorithm was proposed to address the complicated wood volume detection issue. In view of the problems of YOLOv5 in detection of logs’ volume on a vehicle, this study made the following improvements for the model to improve the detection accuracy of logs’ volume:(1) Modified the cross stage partial(CSP) structure and added the Attention mechanism to the last layer of the Backbone network. The Transformer module used channel attention to strengthen important features, suppressed non-important features, and spatial attention to enhance specific target regions of interest while weakening irrelevant background regions, helping the model to more accurately locate and identify regions of interest, and reduced the number of modules to retain more shallow features, strengthened the feature extraction of trail logs’ volume, and used Transformer to enhance the global feature extraction of network features.(2) A small target detection layer was added to the three-layer detection layer of the original network structure to improve the detection performance of small diameters of logs.(3) In the feature fusion structure, the Ghost convolution was used to reduce the complexity of the model and to fuse shallower feature maps and deep feature maps to improve the detection performance of small and medium-sized targets.(4) Used CIoU_loss to make the model more accurate in the positioning of the regression frame and improve the detection performance of the model, and used DIoU-nms to optimize the selection of the prediction frame to improve the detection of c

关 键 词：YOLOv5 目标检测 原木检测 特征融合 轻量化网络 原木计数 遮挡目标 

分 类 号：S781[农业科学—木材科学与技术] TP391.4[农业科学—林学]