Parallel acceleration of histogram specification based on group mapping law

Histogram specification based on group mapping law (GML) is a classical and important approach in many application areas like image enhancement. However, its computation is quite expensive because of the complex mapping law. This paper proposes a novel method to perform GML by constructing a different matrix between original cumulative histogram and specified cumulative histogram. The next step is to search the minimums in each row and column of the matrix, and the final mapping function can be obtained by the indexes of minimums. The whole algorithm is suitable for parallel computing, so we design a fast structure for histogram specification based on GML using Compute Unified Device Architecture (CUDA), which fully exploits the computing resources of GPUs, such as shared memory. Compared to the CPU counterpart, our method attains a speedup factor of 11 on image size of 2048×2048. In addition, we analyze the efficiency of two key parts, namely parallel repacking histogram, and parallel searching minimum values and mapping. Moreover, accuracy measures are calculated to ensure the correctness of the parallel algorithm.