In the intricate setting of mining operations,identifying foreign objects on conveyor belts transporting high-magnetic ore is hindered by significant scene interference and substantial recognition challenges.To address the issues of frequent loss of foreign object edge information and the considerable difficulty in achieving real-time,high-speed responses in high-magnetic environments,we propose an image recognition and detection methodology grounded in deep learning techniques.Initially,a dataset of foreign objects on conveyor belts is constructed.To address the issue of image blurring,which arises from the high-speed operation of the belt conveyor and the limited data acquisition frequency of industrial cameras,the dark channel defogging technique is employed to preprocess the data,thereby enhancing image clarity.Subsequently,the core architecture of YOLOV8 is refined by incorporating a dynamic attention mechanism and substituting standard convolution with snake convolution.The dynamic attention mechanism enables the model to dynamically allocate focus during input data processing.Concurrently,the integration of snake convolution in place of traditional convolution,in conjunction with C2f,significantly enhances the model’s capacity to process image details.This unique structure facilitates the capture of a broader spectrum of local and global features,thereby substantially reducing the model’s rates of false positives and missed detections concerning buried foreign objects.In conclusion,the YOLOV8 architecture has been enhanced through the integration of a dynamic detection head,which allows for flexible adaptation to multi-scale and multi-directional detection requirements.This modification aims to improve the model’s adaptability and optimize the reduction of computational parameters,thereby significantly enhancing its real-time performance in complex environments.Experimental results demonstrate that the model achieves an average detection accuracy of 96.4%,a recall rate of 91%,and an average detection time of merely 29 milliseconds.The algorithm presented in this paper de-monstrates an enhancement in average detection accuracy and recall by 5.2% and 6.2%,respectively,compared to the original network,thereby confirming its efficacy.This improved algorithm adequately satisfies the demands for precise detection and real-time performance in the context of mine belt transportation,offering substantial support for advancing mine safety management and operational efficiency.
矿业生产涵盖了勘探、开采、破碎与筛选、球磨和浮选等多个步骤,其中皮带运输是不可或缺的环节。传统检测方法难以在高速运转的生产线上对异物作出实时响应,导致设备频繁损坏和生产效率下降(Sotoudeh et al.,2020;Luo et al.,2023;韩跃新等,2024)。因此,及时、准确地发现和处理输送带上的异物至关重要(Dai et al.,2023)。特别是在破碎前的矿石中,异物的夹带现象尤为突出,这些异物不仅数量大,而且具有材质多样、尺寸差异巨大、空间位置快速变化和形状复杂等特征,导致识别难度大。此外,处理高磁性矿石与异物时,矿石与异物之间的物理相似性增加了误判风险,可能导致设备损坏并进一步加剧安全隐患。
为构建用于异物检测的数据集,在某铁矿现场从皮带上方实时采集了681张分辨率为2 560像素 ×1 440像素的异物图像。由于现场数据集中异物图像数量较少,本文采用6种数据增强技术,包括裁剪、镜像翻转、改变亮度、添加噪声、旋转和平移,以扩充数据集并提升模型的泛化性(Liu et al.,2022;Yang et al.,2023)。经过数据增强处理后,数据集的图像数量从初始的681张扩充到最终的4 915张。
现场皮带通廊为封闭空间,空气流通有限,在此环境下,皮带运行会产生粉尘,导致拍摄的照片出现模糊或飘散的效果(Chen et al.,2022)。此外,环境深处缺乏自然光,使得拍摄的图像显得昏暗或阴暗。因此,需要进行去雾和去模糊处理,以提升图像的清晰度,从而增强数据识别和分析准确性。
去雾算法通常包括以下步骤:暗通道先验估计、透射率估计、大气光估计、去雾处理和后处理(Xu et al.,2023)。暗通道先验是指利用图像的局部暗区域来估计透射率和大气光,通过选择图像中最亮的像素来估计没有遮挡的光源。然后,根据估计的透射率和大气光对图像进行去雾处理,逐像素地修改图像以减少或消除由粉尘引起的光照衰减效果。去雾效果如图2所示,其中红框代表异物,黑框代表矿石背景。
1.3 数据划分
通过数据采集、预处理和增强操作,成功构建用于矿带异物检测的数据集。将该数据集按照8∶1∶1的比例随机划分为训练集、验证集和测试集(Zhang et al.,2022),以确保在模型训练过程中充分提供数据支持,并在独立的验证和测试集上评估模型的性能和泛化能力。
注意力机制能够有效捕获全局和局部的关联性,通过计算和强调基础图像特征,减少在目标小或存在严重遮挡的场景下的特征损失,从而显著提升网络性能(Zhu et al.,2023)。然而,引入注意力机制可能会增加内存消耗和计算开销等问题。因此,本研究结合动态稀疏性提出了一种轻量级且高效的注意机制来应对这些挑战。
检测头部分在YOLOV8中负责实现目标检测的核心功能,其设计和性能直接影响着模型的检测效果和计算效率(Dai et al.,2021)。因此,通过优化和改进检测头的设计,能够有效地降低模型的计算负担,提升检测的精度和效率。本文提出的动态头DyHead网络结构可以在不同的输入尺寸和目标尺寸下动态地选择不同的检测头,以适应不同的场景和尺度的目标检测任务。如图5所示,从3个维度将尺度感知、空间感知和任务感知统一在一起。
2.4 集成DySnakeConv模块到C2f模块中
由于异物的种类和结构错综复杂,变化多端,相同的种类在不同的场景下可能会表现出不同的特征。因此,异物识别的根本在于关键特征提取(Qi et al.,2023),如图6所示的动态蛇形卷积,赋予卷积核更多灵活性,使其能够聚焦于目标的复杂几何特征,依次选择每个要处理的目标的下一个位置进行观察,从而确保关注的连续性,不会由于大的变形偏移而将感知范围扩散得太远。
ChenY, SunX, XuL,et al,2022.Application of YOLOv4 algorithm for foreign object detection on a belt conveyor in a low-illumination environment[J].Sensors,22(18):6851.
[2]
ChengDeqiang, XuJinyang, KouQiqi,et al,2022.Lightweight network based on residual information for foreign body classification on coal conveyor belt[J].Journal of China Coal Society,47(3):1361-1369.
[3]
DaiL, ZhangX, GardoniP,et al,2023.A new machine vision detection method for identifying and screening out various large foreign objects on coal belt conveyor lines[J].Complex and Intelligent Systems,9(5):5221-5234.
[4]
DaiX Y, ChenY P, XiaoB,et al,2021.Dynamic head:Unifying object detection heads with attentions[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).Piscataway,NJ:IEEE.
[5]
DongX, PanpanL, JichunW,et al,2023.Mining belt foreign body detection method based on YOLOv4_GECA model[J].Scientific Reports,13(1):8881-8881.
[6]
EshaqR M A, HuE, LiM,et al,2020.Separation between coal and gangue based on infrared radiation and visual extraction of the YCbCr color space[J].IEEE Access,8:55204-55220.
[7]
GuQinghua, ZhouQiong, WangDan,2024.Traveling obstacle detection in open-pit mine area based on improved YOLOv8[J].Gold Science and Technology,32(2):345-355.
[8]
HanYuexin, ZhangXiaolong, GaoPeng,et al,2024.Development and prospect of iron ore processing technologies in China[J].Metal Mine,53(2):1-24.
[9]
LiuX, PedersenM, WangR,2022.Survey of natural image enhancement techniques:Classification,evaluation,challenges,and perspectives[J].Digital Signal Processing,127:103547.
[10]
LuoB, KouZ, HanC,et al,2023.A “hardware-friendly”foreign object identification method for belt conveyors based on improved YOLOv8[J].Applied Sciences,13(20):11464.
[11]
QiY, HeY, QiX,et al,2023.Dynamic snake convolution based on topological geometric constraints for tubular structure segmentation[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition(CVPR).Piscataway,NJ:IEEE.
[12]
RenZhiling, ZhuYancun,2023.Research on foreign object detection of coal mine belt transportation with improved centernet algorithm[J].Control Engineering of China,30(4):703-711.
[13]
RuanShunling, RuanYankang, LuCaiwu,et al,2023.Detection of ore conveyer roller based on infrared image[J].Gold Science and Technology,31(1):123-132.
[14]
SotoudehF, NehringM, KizilM,et al,2020.Production scheduling optimisation for sublevel stoping mines using mathematical programming:A review of literature and future directions[J].Resources Policy,68:101809.
[15]
WangW, ZhangC,2017.Separating coal and gangue using three-dimensional laser scanning[J].International Journal of Mineral Processing,169:79-84.
[16]
WangWei, LiQing, ZhangDezheng,et al,2023.A survey of ore image processing based on deep learning[J].Chinese Journal of Engineering,45(4):621-631.
[17]
XuY J, ZhangH Z, HeF L,et al,2023.Enhanced cycleGAN network with adaptive dark channel prior for unpaired single-image dehazing[J].Entropy,25(6):856.
[18]
YangY, XuW, HuangS,et al,2023.PEN-DS:Progressive enhancement network based on detail supplementation for low-light image enhancement[J].International Journal of Machine Learning and Cybernetics,15:2383-2398.
[19]
ZhangM, CaoY, JiangK,et al,2022.Proactive measures to prevent conveyor belt failures:Deep learning-based faster foreign object detection[J].Engineering Failure Analysis,141:106653.
[20]
ZhuL, WangX, KeZ,et al,2023.Biformer:Vision transformer with bi-level routing attention[C]//Proceedings of the IEEE/CVF Conference On Computer Vision and Pattern Recognition(CVPR).Piscataway,NJ:IEEE.