煤矿井下监控图像中存在噪声，清晰度低，且颜色和纹理信息缺失，采用基于机器学习的图像增强方法时还面临低照度−正常照度图像配对数据集采集困难的问题。提出一种改进零参考深度曲线估计（Zero−DCE）模型，并将其应用于矿井低照度图像增强。使用Leaky ReLU激活函数替换Zero−DCE模型中的ReLU激活函数，以加快模型收敛速度，提升低照度图像特征学习效率；在Zero−DCE模型浅层与深层网络之间的跳跃连接处引入卷积块注意力模块（CBAM），以提高模型对图像关键特征的表达能力；在浅层网络中引入非对称卷积块（ACB），以优化模型对局部图像特征的学习能力和细节特征的表现能力；在深层网络中采用串联卷积核（CCK），以降低模型参数量和计算量，缩短模型训练时间。采用LOL公共数据集和矿井自建数据集进行实验验证，结果表明：改进Zero−DCE模型的均方误差（MSE）、峰值信噪比（PSNR）、结构相似性（SSIM）、自然图像质量评估器（NIQE）和视觉信息保真度（VIF）整体上优于典型图像增强模型，在自建数据集上的MSE和NIQE较Zero−DCE模型分别降低16.25%和2.93%，PSNR，SSIM和VIF分别提高2.87%，1.87%和17.64%；图像增强视觉效果较好，可在提高图像亮度的同时有效保留细节纹理信息，降噪效果明显；对单幅图像的推理时间为0.138 s，可实现图像实时增强。

Underground coal mine surveillance images suffer from noise, low clarity, missing color, and texture information. Additionally, machine learning-based image enhancement methods face challenges in collecting paired low-light and normal-light image datasets. To address these issues, this paper proposes an improved Zero-Reference Deep Curve Estimation (Zero-DCE) model for enhancing low-light images in mines. The ReLU activation function in the Zero-DCE model was replaced with Leaky ReLU to accelerate model convergence and improve the efficiency of low-light image feature learning. A Convolutional Block Attention Module (CBAM) was introduced at the skip connections between the shallow and deep networks of the Zero-DCE model to enhance the model's ability to capture key image features. An Asymmetric Convolution Block (ACB) was incorporated into the shallow network to optimize the model's learning of local image features and its ability to represent fine details. A Cascaded Convolution Kernel (CCK) was employed in the deep network to reduce the number of model parameters and computational cost, thereby shortening the training time. Experimental validation was conducted using the LOL public dataset and a self-built mine dataset. The results showed that the improved Zero-DCE model outperformed typical image enhancement models in terms of Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Natural Image Quality Evaluator (NIQE), and Visual Information Fidelity (VIF). Specifically, on the self-built dataset, MSE and NIQE decreased by 16.25% and 2.93%, respectively, while PSNR, SSIM, and VIF increased by 2.87%, 1.87%, and 17.64%, respectively. The enhanced images exhibited superior visual quality, effectively improving brightness while preserving detailed texture information and significantly reducing noise. The inference time for a single image was 0.138 seconds, enabling real-time image enhancement.