Deep Learning-Based Underground Image Quality Enhancement Using Attention Neural Networks

Abstract

Images captured in underground environments frequently suffer from severe quality degradation because of inadequate illumination, non-uniform artificial lighting, airborne dust, fog, sensor noise, motion blur, and limited camera exposure. These degradations reduce image visibility and negatively affect important computer-vision applications such as worker detection, equipment monitoring, obstacle recognition, coal–rock classification, structural inspection, navigation, and emergency response. Conventional enhancement methods, including histogram equalization, gamma correction, contrast stretching, and traditional Retinex algorithms, can improve overall brightness but often introduce overexposure, amplified noise, colour distortion, and loss of fine structural details. Advanced neural networks provide a more effective solution because they can learn nonlinear relationships between degraded and high-quality images while adapting to complex underground illumination conditions. This paper proposes an advanced neural network framework for underground image quality enhancement using a hybrid Retinex-guided convolutional architecture with attention mechanisms and zero-reference learning. The proposed framework decomposes a degraded underground image into illumination and reflectance components, enhances nonuniform illumination, suppresses sensor noise and haze, restores texture details, and reconstructs a visually improved image. A multi-scale feature extraction module captures both local edges and global contextual information, while channel and spatial attention mechanisms emphasize important structures such as workers, machinery, tunnel boundaries, cracks, and obstacles. The model is trained using a combination of illumination consistency, spatial smoothness, colour constancy, structural similarity, and reconstruction losses. The enhancement performance can be evaluated using Peak Signal-to-Noise Ratio, Structural Similarity Index, Natural Image Quality Evaluator, entropy, processing time, and downstream object-detection accuracy. The proposed framework is expected to provide clearer, better-balanced, and less noisy underground images while remaining computationally efficient enough for real-time safety monitoring. Existing studies have shown that combining low-light enhancement with dehazing, task-aware learning, or detection networks can improve both visual quality and machine perception in underground mines. Keywords: Underground Image Enhancement, Low-Light Image Enhancement, Neural Networks, Deep Learning, Retinex Theory, Attention Mechanism, Zero-Reference Learning, Mine Safety Monitoring.