煤矸图像的在线准确快速识别是煤矸智能分选的关键，深度卷积神经网络能够解决这一问题。 

以实际生产状态下采集的煤与矸石图像为训练与测试样本，基于 ResNet 等经典网络与SqueezeNet 等先进轻量级网络建立了煤矸图像识别模型，分析了各模型的训练收敛情况。 基于 kmeans++判断模型中不同卷积核所提取特征的相似程度，基于模型剪枝技术对相似度高的卷积核进行裁剪，实现了识别模型的优化与压缩。 以识别精度、模型规模和模型运算复杂度为评价指标，定量衡量了压缩前后各模型的测试性能。 分析了压缩后的模型对煤矸难、易识别样本的分类热力图可视化结果，揭示了模型的识别机理与分类依据。 

结果表明：利用多数现有的 CNN 框架建立模型均可以对煤和矸石有效识别，但网络复杂度过低则特征提取能力不足，网络复杂度过高则易产生更严重的过拟合情况，即模型复杂度对识别精度影响较大；基于模型剪枝得到的煤矸识别模型可以将煤的截断面处因镜质组成分而产生的反光现象作为识别煤的可靠依据，同时准确捕捉煤与矸石由于硬度不同所产生的轮廓、纹理、表面平整度等差异，综合给出识别结果。 压缩后的模型在计算量与模型大小减少 10 倍的同时，识别精度提升了 17.8% ，实现节约计算与存储资源的同时提升识别精度，模型性能明显优于常规网络模型。

The key to separate coal from gangue intelligently is the image recognition of coal and gangue,and the deep convolutional neural networks can solve this problem. The authors have collected a large number of coal and gangue images at the transportation belt during the production,taken as training samples,and built some coal and gangue im- age recognition models based on classical deep learning networks ( e. g. ResNet) and lightweight deep learning net- works (e. g. SqueezeNet). Also the authors prune some models based on the similarity of feature ex-tracted by different convolutional kernels of these models,and the similarity is measured by clustering results of k-means++. Recognition accuracy,model size and operation complexity of each model is compared. Finally,the authors have visualized heat- maps of class activation in different images to analyze the recognition basis of coal and gangue during the production by CNN. The results show that most existing CNN can be used to differentiate coal and gangue effectively,but the com- plexity of networks has a great impact on the accuracy. The coal and gangue recognition model based on model pruning can accurately capture the surface differences between coal and gangue due to their different hardness,and the reflec- tion generated by vitelline composition at the truncated surface of coal can be used as a reliable basis for identifying coal. The calculation amount and model size of this model are reduced by 10 times,and the recognition accuracy is in- creased by 17. 8% . This method can save some computing and storage resources on the premise of ensuring accuracy, and the performance is significantly better than the conventional network model.