X射线透射煤矸智能识别方法

王文鑫; 黄杰; 王秀宇; 史玉林; 吴高昌

doi:10.13272/j.issn.1671-251x.18037

摘要: 煤矸图像识别是基于伪双能X射线透射（XRT）的煤矸分选技术重要环节。受煤矸紧贴或遮挡导致煤矸图像难以分割和基于人工阈值判别易导致煤矸分类识别错误影响，现有的煤矸识别方法精度不高。提出一种XRT煤矸智能识别方法。采用感受野模块（RFB）与U−Net模型相结合的模型（RFB+U−Net模型）实现伪双能X射线煤矸图像有效分割，解决了因煤矸紧贴或遮挡情况而影响识别精度的问题；以煤矸图像灰度特征中的低能图像灰度最小值、纹理特征中的低能图像锐化最小值和锐化均差为煤矸识别特征，采用多层感知机（MLP）模型实现煤矸识别。实验表明：RFB+U−Net模型的煤矸分割准确率、煤矸粒度精度、煤矸像素均交并比等指标及图像分割效果均优于活动区域模型、U−Net模型、SegNet模型，且模型推理时间较短，满足煤矸图像分割实时性要求；MLP模型隐藏层数量为8时，在2组测试集下的煤矸识别平均准确率均为87%以上；在相同数据集及实验条件下，MLP模型的煤矸识别平均准确率、排矸率均高于基于贝叶斯分类器、支持向量机、逻辑回归、决策树、梯度提升决策树、K近邻算法的模型，且矸石带煤率不超过3%，满足实际煤矸干法分选要求。

Abstract: The coal-gangue image recognition is an important part of coal-gangue separation technology based on pseudo dual energy X-ray transmission (XRT). However, it is difficult to segment the coal-gangue image due to the close proximity or occlusion of coal-gangue, and it is easy to cause classification and recognition errors of coal-gangue based on artificial threshold discrimination. Due to the above influence, existing coal-gangue recognition methods have low precision. In this paper, an X-ray transmission intelligent coal-gangue recognition method is proposed. A U-Net model combined with the receptive field block (RFB) is used to realize the effective segmentation of the pseudo dual energy X-ray coal-gangue image, which is termed as RFB + U-Net model. The problem that the recognition precision is affected by the close proximity or shielding of coal-gangue is solved. The recognition features of coal-gangue are the minimum gray value of the low-energy image in the gray level features of coal-gangue image, and the minimum value and the average difference of sharpened low-energy image in the texture features. A multi layer perceptron (MLP) model is used to realize coal-gangue recognition. Experimental results show that the RFB+U-Net model is superior to the active contour model, U-Net model and SegNet model in terms of coal-gangue segmentation accuracy, coal-gangue particle size precision, coal-gangue pixel mean intersection ratio and image segmentation effect. The reasoning time of the model is short, meeting the real-time requirements of coal-gangue image segmentation. When the number of hidden layers in the MLP model is 8, the average coal-gangue recognition accuracy under two test sets is more than 87%. Under the same data set and experimental conditions, the average recognition accuracy and gangue removal rate of the MLP model are higher than those based on Bayesian classifier, support vector machine, logic regression, decision tree, gradient boosting decision tree and K-nearest neighbor algorithm. The coal carrying rate of gangue shall not exceed 3%, meeting the requirements of actual dry coal-gangue separation.