Abstract: The intelligentization of coal mine requires the realization of intelligent mining. Thereinto, automatic coal and rock recognition is one of the core technologies for unmanned mining. In recent years, image-based identification of coal and rock has been widely concerned, but complex mining environments have constrained the development of coal rock image recognition. Due to the complex and dynamic underground environment of coal mines—characterized by low illumination, pervasive dust and water mist, equipment vibrations, and noise—it is difficult to obtain large quantities of clear and feature-rich coal and rock images. The scarcity of high-quality coal and rock image datasets significantly hinders the accuracy and stability of image-based recognition systems. To address the lack of high-quality coal and rock image datasets for fully mechanized mining faces, a coal-rock similarity simulation experimental platform was constructed. A total of 3600 coal and rock images with a resolution of 512×512 pixels were collected. Using image augmentation techniques such as variations in brightness, color, rotation, scaling, noise addition, and fog simulation, a large and diverse set of coal and rock images was generated. The Labelme software was used to annotate the images, resulting in the construction of a dataset comprising 14400 labeled coal and rock images. The characteristic differences between the simulated coal and rock images and real-world images from mining faces were analyzed. To bridge this domain gap, a coal and rock image style transfer method based on the StarGANv2 generative adversarial network was proposed. The architecture and loss functions of StarGANv2 including its generator, discriminator, mapping network, and style encoder were analyzed in detail. Experiments were designed to quantitatively evaluate the quality, diversity, and style consistency of the images generated by the StarGANv2 based algorithm. The Frechet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS) scores averaged 30.33 and 0.410, respectively, demonstrating that the proposed method achieves favorable diversity and consistency in coal and rock image style transfer, which proves the feasibility and effectiveness of StarGANv2 algorithm for rich coal rock image synthesis under conditions of insufficient coal rock image data.