Parallel Optimization Research Of Low Light Image Enhancement

The images captured at night are always dark, we couldn’t figure out details in the scene. The low night image enhancement algorithm can deal with those images to have clear visual effects. But the enhancement algorithm is always complex and the calculation is massive, so this algorithm couldn’t be used in our daily life for real-time. Recently, the low night image enhancement algorithm focuses mostly concentrated in the enhancement algorithm and the effect to images, the algorithm accelerated research is still very limited, and the algorithm has a very long time-consuming process, by which is difficult to be widely applied in engineering. For this situation, this article improve the low night image enhancement algorithm based on dehazeing technology using parallel computing methods to make it meets the real-time demands.Firstly, this thesis studies hardware and program model related to parallel computing, summarizes the characteristics of parallel computing platform and the superiority of the programming model by analyzing and comparing current parallel computing methods and equipments. Concretely, it analyzes the characteristics of multi-CPU platform, CPU-GPU Heterogeneous architecture and multi-DSP, which lays a solid foundation for the following parallel computing methods.Secondly, I use three different programming models to realize the algorithm on three platforms. Before the realism, I make complexity analysis for each intermediate step of the algorithm. It provides a basis for the parallelism. On the multi-core CPU platform, using the amplifier to find the hot points of the algorithm and improve it, then using Open MP to parallelize the improved algorithm. On the CPU-GPU heterogeneous platform, improve the part which is not suitable for parallel computing the estimate of atmospheric light value, using CUDA programming model to realize the algorithm on GPU. On the multi-core DSP platform, divide the algorithm and using the master-slave model to allocate the task on each core. These cores execute the tasks parallel. On the three platforms, the results show that using these models can not only got a good enhanced image, but also a high executing speed. For a high resolution picture, it can reach the real-time demands as 25 frames per second on multi-core CPU and GPU platform. On multi-core DSP Embedded equipment, normal size images can reach thereal-time demand due to the constrain of the memory capacity。...