Machine vision has been widely used in the field of coal processing and sorting. However, in image segmentation, there are still challenges in separating the background and foreground in multi-scale feature coal particle CT images, as well as segmentation challenges caused by inconsistent particle sizes. This paper proposes a semantic segmentation method for coal particle CT images based on an improved ECA-Segformer model. In order to address the phenomenon of missed detection that occurs due to uneven distribution of particles at multiple scales, the model introduces the ECA-Net attention mechanism to effectively enhance the network's representation ability, aiming to improve segmentation accuracy. In addition, the use of the Squared ReLU activation function can better capture the different features of the foreground and background, thereby improving the segmentation efficiency of coal particle CT images. Experiments were conducted using a self-built CT dataset of coal particles. The results showed that the improved Segformer model had the best comprehensive detection ability, with an average intersection-union ratio of 87.78%, an average pixel accuracy of 93.44%, and an accuracy rate of 93.46%. Compared to the basic Segformer network, it improved by 2.12%, 1.30%, and 0.58% respectively. The analysis of the segmented data can study the particle size distribution statistics of coal particles, which is of great significance for efficient and intelligent separation of coal.