Abstract: In the field of intelligent coal-gangue sorting, the density prediction of materials with wide thickness and density ranges is significantly challenged by the combined effects of dual-energy X-ray transmission physics, data fusion strategies, and complex material composition, leading to high prediction difficulty and low accuracy. These issues severely hinder the development of intelligent photoelectric sorting technologies for coal-gangue and the quality-based classification and utilization of coal. To mitigate the impacts of these factors and improve the density prediction accuracy of coal and gangue, a prediction model based on virtual thickness images and an improved density formula is proposed. This model employs variable-window nonlinear smoothing filtering to reduce the impact of gangue inclusion, uses depth camera data as a thickness reference to calibrate low-energy image grayscale values, and constructs virtual thickness images that achieve pixel-level alignment with dual-energy X-ray images. The limitations of applying a unified density calculation formula are analyzed, and traditional density calculation formulas are improved based on the variation patterns of coal/gangue results across different density levels, establishing a density prediction model suitable for coal and gangue. Additionally, the introduction of boundary constraints further enhances the model's prediction accuracy. Through comparative testing with various algorithms on coal-gangue datasets of different density levels and thickness ranges, the performance of the proposed model is comprehensively evaluated. Experimental results demonstrate that the model accurately predicts coal and gangue densities within wide thickness-density ranges. For materials with densities of 1.3～1.8 g/cm³ (mean thickness range: 5～100 mm), the mean absolute error (MAE) and error variance (E-Var) are as low as 0.0539 and 0.0009, respectively. For materials with densities > 1.8 g/cm³ (mean thickness range: 5～100 mm), the MAE and E-Var are 0.2106 and 0.0332, respectively. Compared to the P-value method, transmission formula method, Archimedes principle method, and mass-volume ratio method, the proposed model maintains higher accuracy and stability while achieving multi-target synchronous measurements. In addition, high-accuracy electron density images were obtained, providing valuable data for subsequent key processes such as component analysis, calculation of mass centroid, and assessment of material distribution.