Abstract: High-resolution remote sensing images are of great significance for geological exploration,resource assessment, and mine safety management. Using Generative Adversarial Networks (GAN) for single-image super-resolution reconstruction of mine remote sensing images can improve image quality and enrich image details. However,this method still has problems such as unstable training process,differences between generated image details and real ground objects,and the tendency to produce artifacts in practical applications. Therefore,an improved attention mechanism-based GAN algorithm for enhancing mine remote sensing images is proposed. The generator integrates the Pyramid Split Attention (PSA) module and Residual Dense Block (RDB),significantly enhancing the feature extraction capability. The PSA module can effectively capture multi-scale image features and improve the model′s sensitivity to details;the RDB promotes the full flow of information in the network through dense connections, further enhancing the feature representation ability. For the discriminator, Spectral Normalization ( SN) technology is adopted instead of the traditional BN (Bath Bath Normalization) layer to enhance the learning ability of mine remote sensing image details and reduce the discriminator′s neglect of generated image details. Additionally,based on the theory of Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP),the adversarial loss function is optimized by introducing a gradient penalty term,stabilizing the training process and accelerating the model′s convergence speed. Experimental results show that the proposed algorithm outperforms the original GAN algorithm in terms of detail texture richness and artifact reduction. The Peak Signal-to-Noise Ratio (PSNR) of the processed remote sensing images is increased by 0. 536 to 1. 897 dB,and the Structural Similarity (SSIM) is increased by 0. 019 to 0. 089. 

Key words: mine remote sensing image,generative adversarial network,attention mechanism,super-resolution