随着煤矿智能化建设的逐步推进，智能视频监控系统越来越多的应用于煤矿井下。然而受粉尘、水雾及光源等因素的影响，视频监控系统采集的图像往往存在亮度低、光照不均匀、信息丢失、细节模糊等问题，导致煤矿井下视频监控视觉效果差，极大影响了后续图像分析与智能决策。因此，研究煤矿井下图像增强方法具有重要意义。

针对井下低照度图像清晰度及亮度低，增强过程出现光晕伪影的问题，提出一种基于双边滤波的改进MSRCR（Multi-Scale Retinex with Color Restoration）图像增强算法。首先改进同态滤波算法增强RGB分量，其次将图像转换到HSV空间，改进MSRCR算法优化照度分量，引入直方图均衡化改进反射分量，采用非线性拉伸变换改进饱和度分量，最后将分量合并后转换回RGB空间。基于自建井下数据集进行验证，改进算法的平均梯度、峰值信噪比PSNR（Peak Signal-to-Noise Ratio）、结构相似性SSIM（Structure Similarity Index Measure）、信息熵平均提高了30%、15%、20%、9%。实验验证结果表明，该算法使图像的整体清晰度及亮度明显提高，减少了增强过程中的光晕伪影，有效改善图像质量。

针对井下非均匀光照图像局部区域亮度低和细节特征缺失的问题，提出一种基于自适应估计的改进Retinex-Net图像增强算法。首先设计分解网络获取图像的照度分量和反射分量，其次在反射网络中引入融合通道和空间注意力的注意力模块CBAM（Convolutional Block Attention Module）注意力机制提升图像的细节和对比度，在光照估计网络中构建渐进式的光照优化过程，逐步优化光照分量的估计，引入自校准模块自动调整光照分量的估计值，最后将照度分量和反射分量相加，得到增强后的井下图像。基于自建井下数据集进行验证，改进算法的平均梯度、PSNR、SSIM、信息熵平均提高了25%、17%、24%、8%。实验验证结果表明，该算法能在提高暗区域图像亮度的同时保持图像细节信息，提升图像质量。

本文研究发现，对井下低照度图像采用基于双边滤波的改进MSRCR算法可以更加简洁高效的处理图像，提升图像整体的清晰度和亮度，减少增强过程中光晕伪影的出现；对井下光照不均匀图像采用基于自适应估计的改进Retinex-Net算法可以对暗区域进行针对性增强，通过训练模型，使用测试集进行验证，算法可以有效地提高暗区域的图像亮度，同时保持原有的图像细节信息。

As the construction of coal mines becomes increasingly intelligent, the use of intelligent video surveillance systems in underground coal mines is also on the rise. However, the images collected by these systems are often affected by low brightness, uneven illumination, information loss, blurred details, and other issues caused by dust, water mist, and light sources. This negatively impacts the visual quality of the surveillance footage and hinders subsequent image analysis and intelligent decision-making. Consequently, it is of great significance to study the image enhancement method for underground coal mines.

The aim of this thesis is to address the issues of low clarity and low brightness in downhole low illumination images and halo artefacts in the enhancement process. To this end, an improved MSRCR (Multi-Scale Retinex with Color Restoration) image enhancement algorithm based on bilateral filtering is proposed. Firstly, the homomorphic filtering algorithm is enhanced to enhance the RGB component. Secondly, the image is converted to HSV space. Thirdly, the MSRCR algorithm is improved to optimise the illuminance component. Fourthly, histogram equalisation is introduced to improve the reflectance component. Fifthly, the saturation component is improved by using non-linear stretching. Finally, the components are merged and converted back to RGB space. The self-constructed dataset for validation was employed to assess the efficacy of the improved algorithm. The results demonstrated that the algorithm exhibited an average improvement of 30% in the gradient, 15% in the peak signal-to-noise ratio (PSNR), 20% in the structural similarity (SSIM), and 9% in the information entropy. Furthermore, the experimental outcomes indicated that the algorithm led to a notable enhancement in image clarity and brightness, and a reduction in halo artefacts.

This thesis presents an improved Retinex-Net image enhancement algorithm based on adaptive estimation, which addresses the issues of low brightness in local areas and missing detail features in downhole non-uniform illumination images. Firstly, a decomposition network is designed to obtain the illumination and reflection components of the image. Secondly, the CBAM (Convolutional Block Attention Module) attention mechanism is introduced into the reflection network to improve the details and contrast of the image. Thirdly, a progressive light optimisation process is constructed in the light estimation network to optimise the estimation of the light component step by step. The light estimation network is constructed to optimise the estimation of the light component in a stepwise manner. The self-calibration module is introduced to automatically adjust the estimated value of the illumination component. Finally, the illumination and reflection components are summed up to obtain the enhanced downhole image. The average gradient, PSNR, SSIM, and information entropy of the improved algorithm were found to be 25%, 17%, 24%, and 8% higher, respectively, than those of the original algorithm when tested on a self-built downhole dataset. The experimental results demonstrated that the algorithm was capable of improving image quality by increasing image brightness in dark regions while maintaining image detail information.

This thesis presents findings that the improved MSRCR algorithm, based on bilateral filtering for downhole low illumination images, can process the image more concisely and efficiently, improve the overall clarity and brightness of the image, and reduce the halo artifacts. Furthermore, the improved Retinex-Net algorithm, based on adaptive estimation for downhole unevenly illuminated images, can enhance the dark areas and effectively improve the dark image brightness in dark areas through the training model. The dark areas can be enhanced by training the model, which effectively improves the image brightness of the dark areas while maintaining the original image details.