<p>Froth flotation, as the primary separation method for coal slurry in coal preparation plants, is widely used for the ash reduction and quality improvement of fine coal. The morphological characteristics of the flotation froth directly reflect the flotation efficiency and are used to guide operators in adjusting process parameters. Given the challenges of froth images, such as blurred boundaries and the difficulty in segmenting coal slurry froth, this paper proposes an effective froth image segmentation model called the Feature Extraction-Prediction Network (FEPNet). The network model is designed with the Stem Convolution Group module (SCG), Global–Local Hybrid Attention module (GLHA), and High-Width Combined Attention module (HWCA) to extract global and local feature information and achieve multi-dimensional feature fusion. The proposed FEPNet model is evaluated on a self-constructed image segmentation dataset of coal froths. Compared to the UNet model, the proposed model can improve the mIoU, DSC, and Acc metrics by 4.86%, 2.86%, and 3.48%, respectively. Experimental results show that the proposed FEPNet model can effectively segment coal slurry froth images and exhibits high accuracy and robustness in extracting froth edge information.</p>