Thermal Image Generation Method Using Improved WGAN-GP for Coal-Rock Recognition

Recognition of coal-rock in coal mining face is a significant prerequisite to ensure intelligent control and stable operation of the shearer. The main challenges of underground application are the potential safety hazards at the comprehensive mining face and the single cutting mode of the shearer, resulting in insufficient number of collected samples, unbalanced distribution, and interference with background information. To this end, a data generation method based on improved Wasserstein generative adversarial network with gradient penalty (WGANGP) is proposed for generating high-quality infrared thermal (IRT) images, which solves the problem of coal-rock recognition with limited samples. First, a conditional random field is incorporated into the generator for postprocessing the generated images to increase the accuracy of the GAN model in the boundary areas of the generated images. Second, an attention-embedded WGAN-GP is constructed to obtain global thermally relevant information of IRT images. Finally, coal wall cutting experiments prove that the improved WGAN-GP model can generate highquality coal-rock IRT images, and its performance is superior to that of other advanced GANs model.