摘要
虽然基于机器学习的粒度估计方法在矿物领域得到了广泛应用,然而由于煤炭的物理特性,造成其在图像中边缘信息不足,难以检测。因此当前胶带煤流的粒度估计仍然采用人工方式进行。针对人工巡检效率低,无客观量化标准的现状,提出一种旨在平衡光照分布的图像增强方法,并构造SSD-ResNet50深度网络用于检测胶带中块状煤,估计煤流粒度。首先,利用一种非线性自适应直方图均衡化与改进中值滤波的图像增强方法增强胶带上物料之间的区分度,提高暗区煤块的可分辨性。然后,提出SSD-ResNet50深度网络结构对块状物料进行检测,评估块状物料的粒度信息。根据实验结果,该方法检测准确度高达88.65%,单张图片的推理时间为160 ms。可以有效实时估计胶带煤流粒度的组成信息,提升巡检效率,达到自动估计胶带煤流粒度的目的。
Granularity estimation methods based on machine learning for ores have been widely used.However,due to the physical characteristics of coal,the edge information in the image is insufficient and difficult to detect.Currently,the coal granularity estimation on the belt conveyor is still carried out manually.To deal with two key issues in current situation,i.e.,low efficiency and no objective quantitative standard,an image enhancement method to balance the illumination in an image is proposed.Meanwhile,a deep-neural network named SSD-ResNet 50 is constructed to estimate the coal granularity on belt conveyor.Firstly,an image enhancement method based on nonlinear adaptive histogram equalization and improved median filter is used to enhance the discrimination between coal on the belt and improve the distinguishability of blocky coal in dark zones.Secondly,detect the block materials based on the proposed SSD-ResNet 50 to estimate the coal granularity on belt conveyor.From the experimental results,the detection accuracy of this method is 88.65%,and the reasoning time of a single picture is 160ms.It can effectively estimate the granularity for belt coal flow in real time,improve the inspection efficiency,and achieve the purpose of automatic granularity estimation for belt coal flow.
作者
张卿
冯化一
张虎平
楚遵勇
王佳乐
ZHANG Qing;FENG Hua-yi;ZHANG Hu-ping;CHU Zun-yong;WANG Jia-le(Guodian Jiantou Inner Mongolia Energy Co.,Ltd,Erdos,Inner Mongolia 017000,China;Tianjin Meiteng Technology Co.,Ltd,Tianjin 300000,China)
出处
《煤炭加工与综合利用》
CAS
2022年第7期49-54,共6页
Coal Processing & Comprehensive Utilization
关键词
煤流粒度估计
深度网络
图像增强
工矿智能化
coal granularity estimation
deep-neural network
image enhancement
industrial and mining intelligence
