【摘要】 视频监控是煤矿安全监测的重要手段之一，由于受到光照、粉尘和水雾的影响，视频监控图像的质量不佳，影响了视频监控系统的准确性，因此研究矿井低照度图像增强算法具有重要意义。针对煤矿井下极暗低照度图像增强过程中出现的纹理特征丢失、色彩失真问题，提出一种基于改进EnlightenGAN的无监督矿井低照度图像增强算法（Efficient Multi-scale Generative Adversarial Network,EMGAN）。首先，在生成器网络中，采用ResU-Net模型替换原始的U-Net模型，加入残差连接机制，增强特征的传递能力，有效缓解训练中的梯度消失问题，保留图像中的细节信息，并在网络的下采样阶段引入跨空间学习的高效多尺度注意力（Efficient Multi-scale Attention,EMA）模块，提高网络在复杂环境中的特征提取能力；其次，基于PatchGAN构建双鉴别器网络，分别对图像的全局区域和局部区域进行鉴别，有效平衡图像的整体亮度和对比度；最后，设计联合损失函数，结合感知损失和色彩一致性损失，避免出现颜色失真现象。基于自建煤矿井下数据集进行验证，改进算法的均值、标准差、信息熵、平均梯度、峰值信噪比（Peak Signal-to-Noise Ratio,PSNR）和结构相似度（Structural Similarity Index Measure,SSIM）平均分别提高了6.84%、24.88%、7.54%、8.30%、27.40%、10.85%。试验结果表明，该算法在提高矿井极暗低照度图像亮度和对比度的同时保留了图像的纹理特征信息，避免了色彩失真现象，有效提高了图像的质量，为后续视频监控和分析提供了更可靠的基础。更多还原

【Abstract】 Video surveillance is one of the important means of coal mine safety monitoring, and the poor quality of video surveillance images affects the accuracy of the video surveillance system due to the influence of illumination, dust and water mist, so it is of great significance to study the mine low-illumination image enhancement algorithm. To address the problems of texture feature loss and colour distortion during the enhancement of extremely dark low-illumination images in coal mines, an unsupervised mine low-illumination image enhancement algorithm EMGAN based on the improved EnlightenGAN is proposed. Firstly, in the generator network, the original U-Net model is replaced by the ResU-Net model, and the residual linking mechanism is added to enhance the feature transmission ability, effectively alleviate the gradient disappearance problem in training, retain the detail information in the image, and introduce an efficient multiscale attention module for cross-space learning in the down-sampling phase of the network to improve the feature extraction ability of the n etwork in complex environments; secondly, construct a dual discriminator network based on the PatchGAN, which discriminates the global and local regions of the image respectively, and efficiently balances the overall image brightness and contrast; finally, the joint loss function is designed to combine the perceptual loss and colour consistency loss to avoid colour distortion. Based on the self-constructed coal mine underground dataset for validation, the mean, standard deviation, information entropy, average gradient, peak signal-to-noise ratio, and structural similarity of the improved algorithm are improved by an average of 6.84%, 24.88%, 7.54%, 8.30%, 27.40%, and10.85%, respectively. The experimental results show that the algorithm improves the brightness and contrast of the extremely dark and lowillumination image of the mine while retaining the texture feature information of the image and avoiding the colour distortion phenomenon,which effectively improves the quality of the image and provides a more reliable basis for the subsequent video monitoring and analysis.更多还原