This paper presents a GPU based parallel implementation for the non-uniform illumination image restoration method, which uses a retinex based algorithm to decompose the original image into brightness and reflectance components, and adjusts the brightness value through an adaptive gamma correction and nonparametric mapping to achieve the restoration. Specifically, we parallelize the improved retinex algorithm on GPU to extract the brightness value of each pixel. After that, the probability of different brightness range is counted through each block to the entire image to reduce the competition of memory access. Finally, we use two different parallel reduce methods to calculate the probability density and cumulative density of brightness value and generate the mapping curve. The experiment conducted on three different GPUs and two CPUs with different resolution images shows that our method can process a 1024 × 2048 image in 1.024 ms on RTX2080Ti, indicates a great potential for real-time application.

本文提出了一种基于GPU的非均匀照明图像恢复方法的并行实现方法，该方法使用基于retinex的算法将原始图像分解为亮度和反射率分量，并通过自适应伽马校正和非参数映射来调整亮度值以实现恢复。具体来说，我们在GPU上并行处理了改进的retinex算法，以提取每个像素的亮度值。之后，通过每个块计算整个图像的亮度范围的可能性，以减少内存访问的竞争。最后，我们使用两种不同的并行约简方法来计算亮度值的概率密度和累积密度，并生成映射曲线。