Aiming at the problems of complex model,poor real-time performance a nd easy loss of characteristics in the existing coal gangue separation methods,a lightweight co al gangue separation network GC-resnet18 is constructed.GC-resnet18 uses ghost convolution (GC) to line arly generate ghost map to eliminate redundant information of similarity characteristics of coal and gan gue.With the help of Softpool′s down sampling activation mapping,the characteristic information of c oal and gangue is retained and highlighted,and redundant parameters are removed to prevent over f itting.GC self-attention mechanism is introduced to integrate the advantages of lightweight of SEnet and global capture of long- distance information of NLNet,so as to make the network remember and enlarge th e subtle difference characteristics between coal gangue images,and improve the accuracy of coal gan gue image recognition accuracy.The experimental results show that GC reduces the am ount of parameters by 46.6%,and the accuracy of GC self-attention mechanism is improved by 1.44% and 2. 32% on CIFAR10 and CIFAR100 respectively,while Softpool pooling is improved by 0.22% and 0.17% on the above two data sets respectively.Through comparative experiments,the comprehensively imp roved GC-ResNet18 network is better than VGG19-S-GDCNN,SBP-VGG-16 and other models in trainin g efficiency and classification accuracy.The classification accuracy in CIFAR10 and CIFAR100 dat a sets is the best 94.07% and 74.95% compared with the network of the same scale,and finally reaches 97.2 % classification accuracy in self-built coal gangue data sets.