| 916 | 9 | 352 |
| 下载次数 | 被引频次 | 阅读次数 |
目前森林火灾多发，建立日常监测尤为重要，但是边缘智能检测设备算力和内存较低，限制了检测模型的推理和部署。针对以上问题，提出一种改进的重参数化YOLOv5s的森林火灾检测算法，结合重参数化、通道重排和深度可分离卷积(Depthwise Separable Convolution, DSC)等轻量化思想分别设计新的骨干和颈部网络，增强特征提取能力，提高模型检测精度，使参数量和推理权重较大幅度减少。为避免颈部网络的信息丢失，根据空洞卷积提出特征增强模块，增强多尺度特征融合能力。为进一步提高模型性能，加入轻量化的CA注意力机制，更准确定位目标。当前公开的火焰烟雾数据集存在针对性不强的问题，为了更好地提高模型的检测效率，制作了一个新的森林火灾数据集，在数据集上利用结构相似性算法剔除了相似度过高的图片，保证了模型的泛化能力。实验结果表明，改进后的重参数化YOLOv5s以原网络约76%的参数量提高了4.0%的精确度，推理权重下降至10.5 MB,更适合于设备性能差、容量小的边缘设备，提高了森林火灾巡检的效率。
Abstract：Currently, forest fires are frequent. It is particularly important to establish daily monitoring. However, the low computational power and memory of edge intelligent detection device limit the reasoning and deployment of detection model. To address the above issues, an improved forest fire detection algorithm based on re-parameterized YOLOv5s is proposed, which combines lightweight ideas such as re-parameterization, channel rearrangement, and Depthwise Separable Convolution(DSC) to design new backbone and neck networks respectively, enhancing feature extraction capabilities, improving model detection accuracy, and significantly reducing the amount of parameters and reasoning weight. To avoid information loss in the neck network, a feature enhancement module is proposed based on hole convolution to enhance the multi-scale feature fusion ability. In order to further improve the performance of the model, a lightweight CA attention mechanism is added to more accurately locate the target. In addition, the currently published flame and smoke data sets have a problem of being not targeted. In order to better improve the detection efficiency of the model, a new forest fire data set has been created. At the same time, structural similarity algorithms have been used to eliminate images with high similarity on the data set, ensuring the generalization ability of the model. Experimental results show that improved re-parameterized YOLOv5s improves the accuracy by 4.0% with about 76% of the original network's parameter amount, while reducing the inference weight to 10.5 MB, making it more suitable for edge devices with poor equipment performance and small capacity, and improving the efficiency of forest fire patrol inspection.
[1] MING Y Y.Fire Detecting System Design Under the Mainframe Computer Vision[J].Applied Mechanics and Materials,2014,3365(602/605):2038-2040.
[2] 黄翰鹏，李柏林，欧阳，等.融合模糊神经网络与时序模型的火灾预警算法[J].计算机工程与设计，2020,41(6):1639-1644.
[3] 卢鹏，赵亚琴，陈越，等.复杂背景环境下基于SSD_MobileNet深度学习模型的火焰图像识别研究[J].火灾科学，2020,29(3):142-149.
[4] 张彬彬，帕孜来·马合木提.基于YOLOv3改进的火焰目标检测算法[J].激光与光电子学进展，2021,58(24):289-296.
[5] 王冠博，杨俊东，李波，等.改进YOLOv4的火焰图像实时检测[J].计算机工程与设计，2022,43(5):1358-1365.
[6] 皮骏，刘宇恒，李久昊.基于YOLOv5s的轻量化森林火灾检测算法研究[J].图学学报，2023,44(1):26-32.
[7] Ultralytics.YOLOv5[DB/OL].(2020-06-26)[2023-02-22].https://github.com/ultralytics/YOLOv5.
[8] DING X H,ZHANG X Y,MA N,et al.RepVGG:Making VGG-style ConvNets Great Again[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).Nashville:IEEE,2021:13728-13737.
[9] DING X H,ZHANG X Y,HAN J G,et al.Diverse Branch Block:Building a Convolution as an Inception-like Unit[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).Nashville:IEEE,2021:10881-10890.
[10] ZHANG X Y,ZHOU X Y,LIN M X,et al.Shufflenet:An Extremely Efficient Convolutional Neural Network for Mobile Devices[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City:IEEE,2018:6848-6856.
[11] MA N N,ZHANG X Y,ZHENG H T,et al.ShuffleNet V2:Practical Guidelines for Efficient CNN Architecture Design[C]//15th European Conference on Computer Vision (ECCV).Munich:ACM,2018:122-138.
[12] HOU Q B,ZHOU D Q,FENG J S.Coordinate Attention for Efficient Mobile Network Design[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Nashville:IEEE,2021:13708-13717.
[13] HOWARD A G,ZHU M L,CHEN B,et al.MobileNets:Efficient Convolutional Neural Networks for Mobile Vision Applications[EB/OL].(2017-04-17)[2023-04-01].https://arxiv.org/abs/1704.04861.
[14] SANDLER M,HOWARD A,ZHU M,et al.MobileNetV2:Inverted Residuals and Linear Bottlenecks[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City:IEEE,2018:4510-4520.
[15] HOWARD A,SANDLER M,CHEN B,et al.Searching for MobileNetV3[C]//2019 IEEE/CVF International Conference on Computer Vision.Seoul:IEEE,2019:1314-1324.
[16] LIU S,QI L,QIN H F,et al.Path Aggregation Network for Instance Segmentation[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City:IEEE,2018:8759-8768.
[17] LIN T Y,MAIRE M,BELONGIE S,et al.Microsoft COCO:Common Objects in Context[C]//European Conference Computer Vision (ECCV).Zurich:Springer,2014:740-755.
[18] HE K M,ZHANG X Y,REN S Q,et al.Deep Residual Learning for Image Recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition.Las Vegas:IEEE,2016:770-778.
[19] SZEGEDY C,IOFFE S,VANHOUCKE V,et al.Inception-V4,Inception-ResNet and the Impact of Residual Connections on Learning[C]//Thirty-First AAAI Conference on Artificial Intelligence.San Francisco:ACM,2017:4278-4284.
[20] CHEN L C,PAPANDREOU G,SCHROFF F,et al.Rethinking Atrous Convolution for Semantic Image Segmentation[EB/OL].(2017-06-17)[2023-04-10].https://arxiv.org/abs/1706.05587.
[21] HU J,SHEN L,SUN G.Squeeze-and-Excitation Networks[C]//2018 IEEE Conference on Computer Vision and Pattern Recognition.Salt Lake City:IEEE,2018:7132-7141.
[22] WOO S,PARK J,LEE J Y,et al.CBAM:Convolutional Block Attention Module[C]//European Conference on Computer Vision (ECCV).Munich:Springer,2018:3-19.
[23] BOCHKOVSKIY A,WANG C Y,LIAO H Y M.YOLOv4:Optimal Speed and Accuracy of Object Detection[EB/OL].(2020-04-23)[2023-04-10].https://arxiv.org/abs/2004.10934.
基本信息:
中图分类号:S762.3;TP183;TP391.41
引用信息:
[1]杨武,余华云,赵昕宇,等.重参数化YOLOv5s的森林火灾检测算法[J].无线电工程,2024,54(02):284-293.
基金信息:
国家自然科学基金(61440023); 中国高校产学研创新基金——新一代信息技术创新项目(2020ITA03012)~~
2023-05-17
2023-05-17
2023-05-17