Low illumination image enhancement algorithm of CycleGAN coal mine based on attention mechanism and Dilated convolution

The complex underground environment, filled with a large amount of dust and water vapor, and uneven illumination of artificial light source, leads to problems such as low illumination and loss of detail features in images collected by underground monitoring equipment, which seriously affects the real-time performance of mining safety monitoring equipment, is not good for subsequent computer vision tasks, and it is difficult to collect underground data. It is difficult to make paired low-light image data sets for model training. To solve these problems, a low illumination image enhancement algorithm based on CycleGAN is proposed. In view of the difficulty of collecting paired data set under mine, CycleGAN network is used for unsupervised learning. In order to improve the detail feature extraction ability of the generator network, the image enhancement network was constructed by using the Parameter-Free Attention Mechanism (simAM) and the Dual-Channel Attention Mechanism (CBAM) to improve the anti-interference ability of the model in complex background, so that the model could recover the image detail features better, which improved the anti-interference ability of the model under complex background and made the model recover the detail features better. A luminance enhancement module based on residual cavity convolution is introduced to increase the luminance of the image while enlarging the receptive field of the generator network, which is conducive to detail recovery and visual quality improvement. Patch-GAN is apply for the discriminator of the network, and the input is mapped into a matrix to pay more comprehensive attention to the details of different regions of the image, and improve the discriminator's resolution of image details. Experimental results show that compared with the CycleGAN algorithm, the proposed method improves the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), information entropy and visual information fidelity (VIF) by 11.31%, 8.07%, 2.58% and 6.18% on average.