摘要
螺旋溜槽精矿矿带位置的准确识别和坐标提取是实现精矿精确自动截取的必要前提,本文基于UNet算法改进提出Res-UNet算法,实现了螺旋溜槽精矿带图像的实时分割。在此基础上,进一步提出了一种像素梯度区分特征点检测算法,并将该算法与Sobel边缘检测算法相融合,实现了从螺旋溜槽精矿带分割图中提取出精矿带分界坐标。实验结果表明,Res-UNet模型识别分割螺旋溜槽精矿带时表现出较高的准确性和效率,平均精确度达98.69%,满足实际应用需求。此外,本文所提出的像素梯度区分特征点检测算法具有较高的检出率与稳定性,其检测结果的相对误差均低于4.12%,满足工业应用实际需求。
Spiral chutes are widely adopted in mineral processing plants due to their pollution-free operation,low cost,and high efficiency.However,the automation level in spiral chute mineral processing remains relatively low.The extraction of concentrate and grade control still rely on manual identification of the concentrate zone's location and manual adjustment of the mineral block position to align with the concentrate zone boundaries.This approach is prone to delayed adjustments and significant errors,leading to fluctuations in processing indicators.Therefore,there is an urgent need to develop an automated method for extracting concentrate from spiral chutes.Accurate identification and coordinate extraction of the concentrate band location are essential prerequisites for precise automatic concentrate extraction.Addressing challenges in spiral chute processing-such as blurred boundaries,irregular edges,and color similarity between concentrate bands and background due to high pulp flow velocity and rolling wave phenomena-this paper proposes the Res-UNet algorithm based on an enhanced UNet architecture.By employing the ResNet50 residual network as the feature extraction component for the UNet encoder,the model gains enhanced capabilities for processing both shallow and deep features.This enhanced capability ensures robust feature extraction while maintaining superior feature representation,resolving the issue of the gradient vanishing problem that prevents deep learning networks from learning effective features during training.This enables the network model to capture relatively good semantic information,achieving real-time segmentation of spiral chute concentrate zone images.Building upon this foundation,we further propose a pixel gradient-based feature point detection algorithm.By integrating this algorithm with the Sobel edge detection method,we identify potential fine edges in images by capturing intensity variations and their fastest-changing directions-i.e.,image gradients.This enables the extraction of boundary coordinates for the concentrate zone from the segmented spiral chute concentrate zone image.The coordinate data of the left and right boundary lines extracted by the pixel gradient-based feature point detection algorithm exhibit orderliness and separability,enabling rapid extraction of arbitrary coordinate data along the boundary lines and achieving separation of the left and right boundary line coordinate data.Simultaneously,the similar triangle distance measurement algorithm is employed to measure the width of the spiral chute's concentrate zone,with the measurement results serving as a measure for evaluating the accuracy of the proposed algorithm.To validate the accuracy of the proposed algorithm and the reliability of the execution system,mineral separation experiments were conducted using a rare earth ore sample.A 20%slurry concentration was used in a laboratory-scale spiral chute prototype for mineral separation.The proposed algorithm was applied to perform real-time segmentation and detection of the concentrate zone within the spiral chute via an image recognition system.Experimental results demonstrate that the Res-UNet model exhibits high accuracy and efficiency in identifying and segmenting the concentrate bands within the spiral chute.Key metrics,including mean Intersection over Union(mIOU),mean Pixel Accuracy(mPA),Precision,and F1 score achieved 96.32%,98.69%,99.07%,and 98.7%,respectively,meeting practical application requirements.Furthermore,the proposed pixel gradient-based feature point detection algorithm demonstrates high detection rates and stability.The maximum error between the detected concentrate zone width and the actual measured width is 4.12%,with the overall relative error remaining below 4.2%,meeting industrial application requirements.The proposed algorithm successfully segmented the mineral zoning cross-section of the spiral chute,detected the ore boundary lines and separation point location features,and achieved multidimensional feature extraction(point,line,and surface)of the ore zone image.It demonstrates strong applicability to adaptive ore extraction in spiral chutes based on mineral zoning images.This algorithm lays the foundation for intelligent spiral mineral processing and holds significant potential for broader application.
作者
刘惠中
黄翱
邓富龙
刘茜茜
汪全能
LIU Huizhong;HUANG Ao;DENG Fulong;LIU Xixi;WANG Quanneng(School of Mechanical and Electrical Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China;Jiangxi Province Engineering Research Center for Mechanical and Electrical of Mining and Metallurgy,Ganzhou 341000,China)
出处
《有色金属(中英文)》
北大核心
2026年第2期339-349,共11页
Nonferrous Metals
基金
国家自然科学基金资助项目(52164019)
江西省研究生创新专项资金项目(YC2023-S649,YC2023-S50)
江西省“双千计划”引进高层次创新人才项目(jxsq2018101046)。
关键词
螺旋溜槽
重力选矿
语义分割
特征点检测
机器视觉
spiral concentrator
gravity separation
semantic segmentation
feature point detection
machine vision
