Abstract: The semantic segmentation of coal maceral groups is an important prerequisite for automatically determining the content of each group using image technology. In order to improve the accuracy of coal maceral groups recognition, according to the characteristics of coal microscopic images, a U-type deep learning model based on U-Net network model architecture is proposed. Firstly, a Transformer backbone network based on group mixing attention mechanism is employed to construct the encoder of the U-type net. By multiple scale features aggregating and self-attention mechanism, more comprehensive features are aggregated, the model is more sensitive to structural information of complex image. Secondly, an atrous spatial pyramid pooling module is cascaded at the end of the encoder to expand the receptive field without reducing the resolution of the feature map. Finally, an adjacent layer subtraction module is constructed to aggregate the differential features between encoders and decoders of different scales, by modified the aggregation method of information between levels, high-level semantic information and low-level detail information are integrated more effectively, and richer feature expressions are provided. By integrating multi-scale information, the accuracy and robustness of the model in the segmentation of complex scenes and multi-scale targets of coal microscopic images are improved. In response to the fact that most existing methods focus on single coal type and single object a semantic segmentation dataset containing coal maceral groups with different coal ranks is constructed. After training the proposed semantic segmentation model with the dataset, the simultaneous segmentation and interpretation of multi-objective of maceral groups suitable for multiple coal ranks are realized. The results showed that the proposed model can achieve a mDice coefficient of 93.39%, an average intersection over union (mIoU) of 88.19%, and a pixel accuracy (PA) of 96.50%, which are 4.48%, 7.32%, and 2.82% higher than those with the original U-Net network. Comparing the content data of coal maceral groups calculated by the model with the results of manual identification, 95.3% of the data are with the range 4%, indicating the applicability and stability of the proposed model in the coal maceral groups semantic segmentation.