Abstract
Recently, HDR images are becoming more prevalent. Lots of mobile phone or digital camera could take HDR images. However, there is usually a problem when displaying HDR images. The display devices could only display an image with limit dynamic range. Therefore, people develop some methods to transfer HDR images to low dynamic range, which are known as tone mapping. This work is revised from a local tone-mapping algorithm, Adaptive Local Histogram Adjustment (ALHA), to display high dynamic range (HDR) images. In order to reproduce images with better details and contrast, using the local tone-mapping operators is better than global operators. However, local tone mapping algorithm usually needs a huge amount of computation, it takes a long time to display a HDR image. Designing a highly parallel method and use Graphics Processing Unit (GPU) to accelerate computation would be the possible solutions to achieve a real-time display. In order to run on different heterogeneous systems, we choose OpenCL to implement instead of CUDA. Since CUDA could only use GPUs made by Nvidia, this limits system on choice of hardware. Since the algorithm (ALHA) could be highly paralleled, we can gain a huge speedup when using GPU as accelerator. For a 1280X960 image, we can gain up to 63x. However, the final bottleneck of speedup is limited to the input and output an image.