在非均匀光照或低光照环境下拍摄的图像往往存在可见度低、细节丢失和噪声干扰等问题。这种复杂光照环境下获取到的图像不仅视觉效果差，而且也不利于其他高层计算机视觉任务。现有图像增强技术虽然可以显著改善图像的质量，但存在两个主要问题：一是处理非均匀光照环境下的图像时，会出现局部欠增强或过增强问题；二 是处理低光照环境下的图像时，增强后的图像会存在噪声干扰、色彩饱和度低的问题。针对上述两个问题，本文提出了两种基于深度学习的图像增强算法，主要研究内容如下：

（1）针对图像增强方法在处理非均匀光照图像时，容易造成局部过增强或欠增强的问题，提出一种采用融合特征注意机制的图像增强网络 ULIEN( Uneven Light Image Enhancement Network)。所提方法通过深度卷积网络学习一组非线性伽马函数以实现图像增强。为避免增强图像出现局部过增强或欠增强问题，网络使用亮度注意图和通道注意机制为图像不同的亮度区域和特征通道分配不同的学习权值，使网络关注图像不同区域的增强过程。实验结果表明，所提算法增强后的图像能够避免细节丢失、伪影和局部过增强或欠增强问题。在客观评价方面，经该算法增强后的图像在四个客观评价指标上也明显优于其他对比算法。

（2）针对图像增强过程中噪声被放大、增强图像色彩饱和度低的问题，提出一种基于 Retinex 分解的低光照图像增强算法。该算法通过三分支卷积神经网络将原始图像分解为光照分量、反射分量和噪声分量。对于光照分量，利用上述图像增强网络ULIEN 进行增强。对于反射分量，通过反射恢复网络进行细节和色彩恢复。最后将增强后的光照分量与恢复后的反射分量合成得到最终的增强图像。实验结果表明，与其他四种图像增强算法相比，所提算法增强后的图像在主观视觉和客观指标两个方面都有较好的结果。

    Images captured under non-uniform lighting or low-light environments often suffer from low visibility, loss of detail, and noise interference. The images obtained under such complex lighting environment are not only poor in visual effect, but also unfavorable for other high level computer vision tasks. Although image enhancement technology can significantly improve the quality of images, there are still two main problems: one is when dealing with images in non-uniform lighting environments, there will be local under-enhancement or over enhancement; the other is processing images in low-light environments. When , the enhanced image will have the problems of noise interference and low color saturation. In response to the above two problems, this paper proposes two image enhancement algorithms based on deep learning. The main research contents are as follows:

    (1) Aiming at the problem that image enhancement methods tend to cause local over enhancement or under-enhancement when processing non-uniform illumination images, an image enhancement network ULIEN (Uneven Light Image Enhancement Network) using a fusion feature attention mechanism is proposed. The proposed method learns a set of nonlinear gamma functions through a deep convolutional network for image enhancement. In order to avoid the problem of local over-enhancement or under-enhancement in the enhanced image, the network uses the brightness attention map and the channel attention mechanism to assign different learning weights to different brightness regions and feature channels of the image, so that the network focuses on the enhancement process of different regions of the image. The experimental results show that the image enhanced by the proposed algorithm can avoid the loss of details, artifacts and local over-enhancement or under-enhancement. In terms of objective evaluation, the image enhanced by this algorithm is also significantly superior to other comparative algorithms in terms of four objective evaluation indicators.

    (2) For the problems of noise amplification and low color saturation of the enhanced image during image enhancement, a low-light image enhancement algorithm based on Retinex decomposition is proposed. The algorithm decomposes the original image into Subject : Research on image enhancement Algorithm in complex illumination environment illumination components, reflection components and noise components through a three branch convolutional neural network. For the illumination component, the above-mentioned image enhancement network ULIEN is used for enhancement. For reflection components, detail and color restoration is performed by a reflection restoration network. Finally, the enhanced illumination component and the restored reflection component are synthesized to obtain the final enhanced image. Experimental results show that compared with other four image enhancement algorithms, the image enhanced by the proposed algorithm has better results in both subjective vision and objective indicators.