【作者】 张锋；

【导师】 王岳环； 高常鑫；

【作者基本信息】 华中科技大学 ， 人工智能， 2024， 博士

【摘要】 低光照图像增强是计算机视觉、图像处理以及计算摄影领域中的一个重要研究方向。尽管相关研究在近年来已取得了不错的进展,但是仍然面临着模型的场景泛化能力差、增强后局部亮度偏移以及拜尔图像处理过程中依赖人工选取参数等多种复杂问题。因此,本文利用亮度先验和拜尔图像处理驱动模型,分别从RGB和RAW两个不同的图像域出发开展研究,解决上述问题。在RGB图像域,针对深度卷积模型的场景泛化能力差的问题,提出了直方图均衡化先验驱动的自监督增强方法。利用自监督学习摆脱现有方法对配对数据集的依赖,通过扩大训练场景实现多场景泛化。该方法利用深度映射网络从直方图均衡化处理后的图像中提取出直方图均衡化先验信息,然后通过空间特征变换层将其整合到网络结构中。实验结果表明,该方法相比于现有的自监督方法提升了约4.7%的性能,同时实现了跨数据集的场景泛化。在RGB图像域,针对扩散模型框架缺乏有效条件约束的问题,提出了自适应亮度调整先验驱动的扩散图像增强方法。利用预训练模型提取有效的自适应亮度先验信息,并通过亮度调整模块对亮度进行初步校正,再利用扩散模型对校正后的亮度进行增强。此外,该方法还提出了一种图像分解策略,将图像分解为颜色图和灰度图,并分别送入扩散模型进行增强。大量定量与定性实验分析表明,该方法相比于现有的扩散模型提升了约4.4%的性能,同时还保留了场景泛化性能。在RAW图像域,针对物理噪声模型严重依赖人工标定的问题,提出了分离式噪声模型驱动的低光照图像增强方法。当前物理模型严重依赖于人工标定,导致高复杂度。为此,该方法提出具有分离合成过程的生成式噪声模型;此外,利用频率域的判别器建立噪声域对齐约束,大幅提升了合成噪声的真实性。实验结果表明,该方法的去噪性能相比于其他基于深度学习的方法提升了8.2%。在RAW图像域,针对多重曝光模型无法兼顾性能与效率的问题,提出了多尺度模型驱动的多重曝光图像增强方法。通过结合不同曝光水平的图像进行去噪和融合,创造出高动态范围图像,并通过多尺度的高效色调映射网络将其映射到低动态范围,同时保留原图的颜色和亮度信息。采用图像自适应的三维查找表和可学习的拉普拉斯滤波器联合优化图像质量。实验结果表明,该方法不仅性能相比于基准方法提升了14.2%,还能实时处理大分辨率图像。更多还原

【Abstract】 Low-light image enhancement is a critical research area in computer vision,image processing,and computational photography.Despite notable advancements in recent years,several challenges remain,including poor model generalization across different scenes,local brightness shifts after enhancement,and the reliance on manually selected parameters during Bayer image processing.This dissertation addresses these issues by leveraging brightness priors and Bayer image processing-driven models,conducting research from both RGB and RAW image domains.In the RGB image domain,to address the issue of poor scene generalization ability of deep convolutional models,a self-supervised enhancement method driven by histogram equalization prior is proposed.This method leverages self-supervised learning to eliminate the dependence on paired datasets,achieving generalization by expanding the training scenes.This method utilizes a mapping function to extract histogram equalization prior from images processed by histogram equalization and then integrates it into the network through a spatial feature transform layer.Experimental results indicate that this method has improved enhancement performance by about 4.7% compared to existing self supervied methods,while also achieving cross-scene generalization.In the RGB image domain,to address the issue of inadequate effective conditional constraints in the diffusion model framework,an adaptive brightness adjustment prior-driven diffusion model is proposed.This method utilizes a pre-trained model to extract effective adaptive brightness prior information,and initially corrects the brightness through a brightness adjustment module,then enhances the corrected brightness using the diffusion model.Moreover,this method also proposes an image decomposition strategy,dividing the image into a color image and a grayscale image,and enhancing them separately through the diffusion model.Extensive quantitative and qualitative experimental analyses indicate that this method improves performance by about 4.4%compared to existing diffusion models,while also retaining scene generality.In the RAW image domain,to address the issue of heavy reliance on manual calibration in the physical noise model,a low-light image enhancement method driven by a separated noise model is proposed.Noise modeling facilitates the synthesis of more noise data to drive models to enhance low-light images.Current physical-based models heavily depend on manual calibration,leading to high complexity.This method utilizes physical-based priors to guide generative models.Moreover,a frequency-domain discriminator is utilized to establish noise-domain alignment constraints,which improves the fidelity of the synthesized noise.Experimental results show that show that the denoising performance of this method have improved by 8.2% compared to other learning-based methods.In the RAW image domain,to address the issue of inability of the multi-exposure model to balance performance and efficiency,a multi-scale model driven multi-exposure model is proposed.This method creates HDR images by denoising and fusing images with different exposure levels and maps them to LDR through a multi-scale efficient tone mapping network,while preserving the original color and brightness information.The image quality is jointly optimized using an image-adaptive 3D lookup table and a learnable Laplacian filter.Experimental results indicate that not only improved enhancement performance by 14.2% compared to the baseline but also can process high-resolution images in real-time.更多还原