基于深度学习的煤中异物机器视觉检测

Machine vision detection of foreign objects in coal using deep learning

在煤炭生产加工过程中,由于开采条件的限制,原煤中混入了大量的异物,导致重介质系统管路的堵塞,严重影响了煤炭加工的生产效率,制约了煤炭质量的提高。为有效解决煤炭分选系统中异物对生产的影响,使用机器视觉技术完成异物检测。本文针对煤中异物的不规则形状和复杂特征,构建一种多尺度特征融合的语义分割网络,对实际生产现场采集的煤中异物图像进行端到端检测。网络中,编码器端采用残差卷积块提取异物图像特征和捕捉上下文信息,将池化操作得到的多尺度特征图谱在解码器端进行融合;将编码器与解码器使用跳层连接,更好地结合图像的背景语义信息,进行端到端的训练;通过类激活图可视化方法分析了模型误检原因,提出了一种损失函数用于缓解模型因煤矸石表面和背景干扰产生的误检情况,并使用条件随机场对网络分割结果进行细化,最终得到一幅二值图像。实验结果表明:该模型能够对煤矸系统中的异物进行有效分割,最终模型在测试集上的均交并比(MIOU)为77.83%。

In the process of coal processing,due to the limitation of mining conditions,a large number of foreign objects are mixed into raw coal,which leads to clogging of heavy medium system pipelines,seriously affecting the production efficiency and restricting the improvement of coal quality.In order to effectively solve the impact of foreign objects in coal,machine vision technology is used to complete the detection and removal of foreign objects.Owing to irregular shapes and complex features for foreign objects in coal,a multi-scale feature fusion semantic segmentation network structure is proposed for end-to-end detection of images of foreign objects collected at the actual production site.In the network,the encoder uses residual convolutional blocks to extract features and captures context information,and the multi-scale feature map obtained by the pooling operation is fused at the decoder.The encoder and decoder are connected by layer jump to better combine the background semantic information of images for end-to-end training.The cause of model misdetection was analyzed through the visualization method of class activation map,a loss function was proposed to alleviate the model misdetection caused by the surface and background interference of coal and gangue,and the results were refined using conditional random fields to get a binary image.The experimental results show that the model can effectively segment the foreign bodies in the coal gangue system,and the mean intersection over union of the model on the test set is 77.83%.