MEFormer: Enhancing Low-Light Images While Preserving Image Authenticity in Mining Environments

In mining environments, ensuring image authenticity is critical for safety monitoring. However, current low-light image enhancement methods often fail to balance optimization and fidelity, resulting in suboptimal image quality. Additionally, existing models trained on general datasets do not meet the unique demands of mining environments, which often feature challenging lighting conditions. To address this, we propose Mining Environment Transformer (MEFormer), a high-fidelity low-light image restoration network with efficient computational performance. MEFormer incorporates an innovative cross-scale feature fusion architecture, which facilitates enhanced image restoration across multiple scales. We also present the Mining Environment Low-Light (MELOL) a new dataset that captures the specific low-light conditions found in mining environments, filling the gap in available data. Experiments on public datasets and MELOL demonstrate that MEFormer achieves a 0.05 increase in the SSIM, a PSNR above 25, and an LPIPS score of 0.15. The model processes 10,000 128 × 128 images in just 2.8 s using an Nvidia H100 GPU.