An environment characterised by darkness, pervasive dust, and uneven light distribution is presented in underground coal mining excavation workfaces. Captured images are frequently suffered from low clarity, poor contrast, and detail loss. Traditional image enhancement techniques fail to address uneven brightness conditions encountered in this specific scenario’s imagery. Existing deep learning-based image processing methods, when applied in this scenario, face the challenge of scarce high-quality image data. Meanwhile, existing deep learning-based methods face the challenge of scarce high-quality image data when applied to underground coal mine excavation workfaces. Therefore, an image enhancement approach for excavation workfaces that integrated an improved self-calibrated illumination (SCI) technique was proposed. Dynamic adjustment was applied to dark areas of the collected excavation face images. Issues such as overexposure, noise amplification, and detail loss were mitigated. Moreover, the captured images were input into the SCI module. A self-calibration module was employed at its core to estimate image illuminance. Furthermore, self-calibrating illumination learning was adopted. Then, the LAB colour space was used to correct colour casts induced by ambient light during white balance adjustment. Local brightness imbalances and insufficient contrast were improved through brightness gain adjustments and adaptive contrast enhancement, whilst highlight and shadow details were preserved. An adaptive gamma correction function was constructed to balance illumination components. Meanwhile, an adaptive saturation nonlinear correction function was designed to optimise colour, rendering images more aligned with human visual perception. Research results demonstrate that, when compared to five commonly used image enhancement algorithms, the underground coal mining excavation workface images processed using the proposed method exhibit favourable performance across evaluation metrics, including mean value, mean gradient, standard deviation, information entropy, and spatial frequency. These metrics show average improvements of 21.41%、50.50%、56.96%、6.60% and 44.05%, respectively. It is confirmed that the problems of low clarity, low visibility, poor contrast, and detail loss are effectively solved in such images.