Abstract: Coal gangue identification technology is a critical component for the automation of fully mechanized mining faces and represents a significant challenge within this domain. To address the challenges of suboptimal quality and limited scale of existing coal gangue image datasets, as well as the slow detection speeds and low recognition accuracy of coal gangue image segmentation models, a large-scale, proportionally accurate simulation platform for fully mechanized mining faces has been established, drawing from real-world mining scenarios. Utilizing this platform, a coal gangue image acquisition system has been developed to construct a high-resolution, simulated dataset of coal gangue images for fully mechanized top coal caving face. An advanced U−Net−based coal gangue segmentation model has been developed, incorporating Feature Pyramid Networks (FPN) and Atrous Spatial Pyramid Pooling (ASPP). This approach significantly improves the segmentation accuracy of coal gangue images. By incorporating the Feature Pyramid Networks (FPN) module into the skip connections of the U−Net architecture and integrating the Atrous Spatial Pyramid Pooling (ASPP) module within the decoder stage, a novel FPN-ASPP−U−Net coal gangue segmentation model has been developed. Ablation studies confirmed that the integration of the FPN and ASPP modules significantly enhances the performance of the U−Net model. The FPN−ASPP−U−Net model exhibits superior segmentation efficacy, achieving a mean accuracy (M_A) of 97.29%, a mean F1-score (M_F1) of 97.44%, and a mean Intersection over Union (M_I) of 95.65%. The model's parameter count is 29.64 M, with FLOPs (F) 341.29 G and a frame rate (f) of 41.1 frames per second. Relative to the baseline U−Net model, the M_I, M_F1, and M_A are improved by 2.64%, 1.06%, and 1.15%, respectively, with only a marginal increase of 0.33 M. This enhancement results in a modest improvement in image segmentation speed. A rigorous comparative analysis was conducted to evaluate the performance of the FPN−ASPP−U−Net model relative to PSPNet, SegFormer, DeepLabV3+, and PSANet for image segmentation tasks. The results substantiate that the FPN−ASPP−U−Net model delivers superior performance in coal gangue image segmentation, while also maintaining the lowest overall computational parameter count. This model demonstrates a well-balanced compromise between segmentation accuracy and computational efficiency, thereby optimizing both precision and processing speed in practical scenarios. In response to image degradation caused by dust, a hybrid dehazing approach leveraging dark channel prior combined with Gaussian weighting was implemented on the image dataset. This methodology significantly enhanced segmentation accuracy. Specifically, for coal, the segmentation accuracy improved by 14.81%, 17.79%, and 23.62% under light, moderate, and severe dust conditions, respectively. For gangue, segmentation accuracy saw improvements of 11.73%, 14.50%, and 14.86% under the same dust conditions. These enhancements demonstrate the effectiveness of the proposed dehazing strategy in mitigating dust-related artifacts and improving segmentation performance across varying levels of dust intensity. Building on the research conclusions, a method for calculating the gangue mixture rate in coal gangue images utilizing the FPN−ASPP−U−Net model was developed. A control experiment for the drawing opening was performed, with a threshold of 20% gangue mixture rate established for closing the drawing opening. During a single operational cycle of the drawing opening, the average discrepancy between the actual and model-predicted gangue mixture rates was 4.71%, thereby confirming the viability of employing gangue mixture rate measurements from coal gangue images for effective discharge control. Finally, the encapsulated model code facilitated the development of an intelligent software solution for coal gangue image recognition. An application framework for gangue segmentation in operational environments was designed, and an image-based discharge control experiment was conducted at the 110501 fully mechanized mining face of the Yushutian coal mine. The validation of this methodology demonstrated its capability to perform accurate segmentation of coal gangue images and to facilitate rational control of the drawing opening. This advancement has significantly enhanced the automation level of the fully mechanized mining face, providing a robust technical framework and valuable reference for advancing the intelligent development of coal mining operations.