基于改进深度残差网络的矿井图像分类

精确煤矸分类及识别能力是煤矿智能煤矸分选机器人要解决的关键问题。在通过深度学习图像分类方法的检测煤矸石中,为克服当前残差网络计算量大、复杂度高以及信息丢失的问题,提出了基于改进深度残差网络的图像分类方法。并提出了一种新的损失函数soft-center loss,克服由于softmax分类器对特征的区分判别能力差以及易造成模型过度自信的问题。同时在图像预处理阶段利用CBDNet去噪网络,提高了井下图像的质量,进一步提升了煤矸分类的准确率。实验结果表明,基于改进深度残差网络分类模型相比于其他分类网络模型在井下图像分类准确率提高了4.12%,在公开数据集CIFAR-10准确率提高了1.5%。

Classification of mine images based on improved deep residual network

Accurate coal gangue classification and recognition capabilities are the key issues for intelligent coal gangue sorting robots to solve.In the detection of coal gangue by the deep learning image classification method,in order to overcome the problems of large computational complexity,high complexity and information loss of the current residual network,this paper proposed an image classification method based on improved deep residual network.It proposed a new loss function(soft-center loss)to overcome the problems of the softmax classifier’s poor ability to discriminate features and the possibility of overconfident models.At the same time,it used the CBDNet denoising network in the image preprocessing stage,which improved the quality of the underground images and further improves the accuracy of coal gangue classification.The experimental results show that compared with other classification network models,the accuracy rate of downhole image classification is improved by 4.12%,and the accuracy rate of public data set CIFAR-10 is increased by 1.5%.