Abstract
The detection of mine water hazards traditionally relies on manual inspection and sensors, which suffer from temporal delays and fluctuating accuracy. Video detection methods for mine water hazards can address these shortcomings, although the field of image recognition relating to mine water hazards faces challenges regarding the scarcity of samples and complexity of scenes. To tackle these issues, this study introduces a multichannel residual attention model based on Unet, which effectively identifies water inrush areas. The method employs the multichannel averaging of frames for residual preprocessing of video data and extracts features from multiple frames. The residual frames are then integrated into the multiframe features through an attention mechanism. A Unet variant based on the convolutional block attention module is designed, in which the middle frame is used as a label for backpropagation. By leveraging the residuals and multiple attention mechanisms, the model is able to focus on water flow areas, enabling the recognition of mine water hazards against complex and varying backgrounds. Experimental results show that this model outperforms traditional network segmentation models in terms of detecting mine water hazards, and offers significant generalization advantages over conventional digital image-based water flow recognition.