Deep Learning-Based Visual Enhancement and Real-Time Water-Inflow Detection for Surveillance Systems in Underground Mines

ABSTRACT

Real-time video surveillance systems are increasingly used in underground mining to monitor geological risks, such as water inrush. However, poor lighting and heavy haze severely limit the visibility of water inflow in recorded images, making direct detection and segmentation of water inflow challenging. This study integrates the convolutional neural network (CNN)-based All-in-One Dehazing Network and Feature Fusion Attention Network with a denoising diffusion probabilistic model (DDPM) to enhance image clarity through dehazing and brightness improvement. Manually aided water-inflow segmentation is performed using the Simple Interactive Object Extraction tool in ImageJ to quantitatively evaluate model performance. Correlations between model performance metrics and water-inflow presence are analyzed to identify early indicators of water inrush. Results show that CNN-based models more effectively highlight water-inflow areas, whereas DDPM produces superior image quality and stability. Leveraging these differences, DDPM-generated images are used as pseudo-ground truth to dynamically assess CNN performance and track changes in water inflow. The proposed workflow provides an efficient, interpretable framework for real-time visual enhancement and risk identification, enabling early warning of water-inrush risk in complex underground mining environments.