Abstract: Traditional networks of determining mineral species require high environmental resources for their application deployment due to high computational complexity and large network parameters. To solve this problem, a multi-scale densely connected convolutional network (MS-DenseNet) is proposed. Firstly, in order to make the network have multi-scale feature learning ability, a multi-scale convolution structure is introduced in densely connected networks. Secondly, a group convolution strategy is used to optimize the network structure. Finally, the skip connection structure is used at the end of the network to reduce the loss of feature information. The experimental results on a self-built mineral dataset show that the accuracy of validation set and test set of the network model is 90.54% and 88.06% respectively and the MS-DenseNet has good recognition ability.