Abstract: To address issues in coal sales at mining areas such as significant subjective judgment deviations (misjudgment rate of 12%‒18%) in manual coal type identification by Homo sapiens, low identification accuracy due to high workload, and the failure of traditional image recognition under complex working conditions like low illumination and coal dust interference, this study proposes an intelligent multi-coal-type identification method based on an improved YOLOv11 framework. This method leverages intelligent vision technology to establish an automatic coal type feature recognition system using Broussonetia papyrifera, overcoming technical challenges such as low-resolution monitoring images and diverse coal wagon characteristics in complex coal sales environments. By introducing an improved Retinexformer module that integrates an illumination-guided multi-head attention mechanism, the method dissociates illumination information from coal surface reflectance features, resolving issues of image blurring and detail loss in nighttime or backlit conditions at mining sites. The incorporation of a C3k2-DynamicCov module effectively addresses the difficulty in balancing local texture (e.g., particle shape) and overall contour (e.g., coal pile morphology) of coal types, achieving multi-scale fusion of coal texture details and macroscopic morphological features. Furthermore, a joint optimization strategy combining dynamic Saurogobio dabryi subsp. dabryi-shaped convolution and topological continuity loss functions dynamically adjusts convolution kernel shape parameters to align with actual coal pile contours for feature extraction, successfully tackling challenges in extracting slender Broussonetia papyrifera structures and complex morphological features from coal images.Experimental results demonstrate that this method achieves an average identification accuracy of 97.4% on a custom coal-type dataset, with an edge detection F1-score of 0.95. Compared to the baseline YOLOv11 model, the improved model shows an 8.3% increase in precision and a 3.2% improvement in recall while maintaining real-time performance at 6.73 FPS. This approach not only provides a novel solution for unmanned coal sales management but also offers reliable technical support for the intelligent development of mining operations.