Denoising algorithm for underground coal mine monitoring images

Existing image restoration methods can remove noise from underground monitoring images to some extent, but they lack the ability to remove mixed interference from impulse noise and dust mapping and show insufficient adaptability to dynamic noise caused by coal mine equipment vibration. To address these problems, a denoising algorithm for underground coal mine monitoring images based on a cross-gated network was proposed. An image restoration model based on a variational autoencoder generator was constructed, and preliminary denoising was achieved through probabilistic generation and adversarial training. On this basis, a cross-gated network integrated with an attention mechanism was designed. Blind-spot and non-blind-spot branches were used to collaboratively extract local details and global contextual information. A feature cross-gated fusion unit dynamically integrated multi-scale features, and a compound loss function consisting of L1 reconstruction loss, edge-preserving loss, and frequency-domain consistency loss was introduced to suppress noise while preserving edge structures. Experimental results on a self-built dataset and the public CUMT-CMUID dataset showed that the images denoised by the proposed algorithm exhibited clearer equipment textures and sharper edges. For images containing impulse noise and dust-mapping noise, the proposed algorithm outperformed typical algorithms such as Wavelet Transform Denoising (WTD), Non-local Means Denoising (NLM), and Block-Matching and 3D Filtering (BM3D), as well as advanced algorithms including an algorithm based on Low Rank Regular Joint Sparse Modeling (LRRJSM), Progressive Multi-stage Algorithm Combining Convolutional Neural Network and Multilayer Perceptron (PMA-CNN-MP), and Progressive Algorithm Based on High-order Interaction (PA-HOI), in terms of mean squared error (MSE), peak signal-to-noise ratio (PSNR), root mean squared error (SNR), signal-to-noise ratio, and structural similarity index measure (SSIM). The results verified that the proposed algorithm achieved better denoising performance and detail preservation capability for monitoring images in complex underground coal mine environments.