摘要
对一种利用卷积神经网络(CNN)和图像处理实现原煤中的煤与矸石的精准识别进行了实验研究,实验所采用的识别方法是将进行人工标注的煤炭与矸石的图像样本输入至卷积神经网络ResNet18进行训练,通过调整网络中的卷积层、池化层,选用合适的激活函数以及损失函数进行调参,输出煤与矸石的关键特征完成煤与矸石的二分类。同时,卷积神经网络在训练过程前,会利用直方图均衡、中值滤波、归一化方法进行图像预处理。图像预处理对神经网络的训练提供数据质量的一致性,可以大大提升训练质量。结果表明,利用神经网络ResNet18训练煤与矸石的识别分类模型,能够有效地识别煤与矸石,当前煤与矸石的识别分类精度可达到99%,且随着训练数据的积累,模型的识别精度会持续提升,能够更广泛地应用与不同的原煤质量。
In this paper,an experimental study on the accurate recognition of coal and gangue in raw coal is carried outby using convolution neural network(CNN)and image processing.The experimental method is to input the image samples of coal and gangue labeled manually into the convolution neural network ResNet18 for training.By adjusting the convolution and pooling layers in the network,suitable activation function and loss function arechosen,the key feature of the output iscoal and waste rock complete binary classification.At the same time,before the training process,the convolution neural network isused for histogram equalization,median filtering and normalization methods are used to pre-process the image.Image preprocessing provides consistency of data quality for training of neural networks,which can greatly improve the training quality.The results of this study show that the recognition and classification model of coal and gangue trained by neural network ResNet18 can effectively identify coal and gangue.At present,the recognition and classification accuracy of coal and gangue can reach 99%.With the accumulation of training data,the recognition accuracy of the model will continue to improve,and it can be widely applied to different raw coal quality.
作者
武国平
梁兴国
胡金良
张秀峰
WU Guoping;LIANG Xingguo;HU Jinliang;ZHANG Xiufeng(Shenhua Group Zhungeer Energy Co. Ltd., Erdos 010300, China;Tianjin Meiteng Technology Co. Ltd.,Tianjin 300385, China)
出处
《微型电脑应用》
2021年第6期100-103,共4页
Microcomputer Applications
关键词
图像处理
卷积神经网络
图像样本
煤与矸石识别
原煤煤质
image processing
convolutional neural network
image samples
recognition of coal andgangue
raw coal quality
