结合Mask-RCNN和最小二乘法的球团粒度识别模型
A pellet size recognition model combining Mask-RCNN and least squares method
林双 1陆伟文 1胡守景 2余正伟 1李文波 3龙红明1
作者信息
- 1. 安徽工业大学冶金工程学院,安徽马鞍山 243002
- 2. 南京钢铁股份有限公司,江苏南京 210035
- 3. 铜陵有色金属集团股份有限公司铜冠冶化分公司,安徽铜陵 244000
- 折叠
摘要
在高炉冶炼中,球团粒径分布是评价其质量的重要指标之一.粒度均匀的球团矿有助于改善高炉料柱透气性,降低冶炼能耗.采用Mask-RCNN(Mask Regional Convolutional Neural Network)算法进行球团边缘分割和粒径分析.针对球团严重堆叠对图像识别的干扰,根据球团的轮廓特性,使用凹点检测算法获取轮廓上的特征点,并对粘黏堆叠程度不同的球团分类,结合最小二乘圆拟合法对遮挡的轮廓信息进行复原.研究结果表明:Mask-RCNN算法的实例分割平均精准度可达到93.5％以上.但是由于球团颗粒的堆叠效应,导致Mask-RCNN算法检测的粒度分布曲线严重偏离人工筛分曲线.通过凹点检测算法和最小二乘法圆拟合算法改进后,小于16 mm球团的占比有不同程度的增加,图像检测粒度分布曲线和人工筛分曲线基本重合,平均粒径误差在Mask-RCNN 算法的基础上降低了 5.52％,降低幅度为98.6％.
Abstract
In blast furnace smelting,the particle size distribution of pellets is one of the important indicators for evaluating their quality.A uniformly sized pellet ore helps improve the permeability of burden in the blast furnace and reduce smelting energy consumption.Pellet edge segmentation and particle size analysis were employed by the Mask-RCNN(Mask Regional Convolutional Neural Network)algorithm.Aiming at the interference of severe stacking of pellets on image recognition,according to the contour characteristics of pellets,the concave point detection algorithm is used to obtain the feature points on the contour and classify the pellets with different degrees of sticky stacking,and combined with the least-squares circle-fitting method to recover the occluded contour information.The research results indicate that the average accuracy of instance segmentation in the Mask-RCNN algorithm can reach over 93.5％.However,due to the stacking effect of pellet particles,the particle size distribution curve detected by the Mask-RCNN algorithm deviates significantly from the artificial sieving curve.After improving the concave point detection algorithm and the least squares circle fitting algorithm,the proportion of pellets smaller than 16mm increased to varying degrees,resulting in particle size distribution curves from image detection aligning completely with those obtained through artificial sieving.The average particle size error was reduced by 5.52％on the basis of the Mask-RCNN algorithm,with a reduction range of 98.6％.
关键词
球团粒度/图像识别/Mask-RCNN/最小二乘圆拟合/凹点检测Key words
pellet size/image recognition/Mask-RCNN/least squares circle fitting/concave point detection引用本文复制引用
基金项目
国家自然科学基金资助项目(52174290)
出版年
2024
NETL
NSTL
万方数据