GPU Application Of Image Interpolation And Hyperspectral Compression Areas

With the development of society, especially in recent decades, science and technology achieved unprecedented progress. Accompanied by is vast amounts of data, as well as vast amounts of computing needs. Therefore, demand for high-performance computing is growing. At present, many countries have attached great importance to the development of supercomputers. Initially, it is commonly research-based on CPU. In recent years, computing power of GPU has get great attention. Parallel computing based on CPU-GPU have gained unprecedented development. Thanks to CUDA proposed, threshold for GPU general computing further reduced, promoted the use of GPU in science and technology and engineering fields.In this article, we focus on study in accelerating image interpolation and high spectral compression using GPU. Both of them are demanding real-time applications, using GPU to improve program execution efficiency is important in practical applications. In image interpolation algorithm, we propose a fine-grained parallel execution model。By contrast with the coarse-grained proof, fine-grained parallel model can achieve higher parallelism, and can improve the efficiency of parallel programs to a certain extent. In hyper-spectral compression, we achieved a parallel C-DPCM-based lossless compression using CUDA, and obtained a speed up to 38 x. The main contents of this article as follows:1. For high computational complexity of edge-directed interpolation, we used both coarse-grained and fine-grained model to accelerate this algorithm. We first use a coarse-grained model on GPU completed the accelerated method. In coarse-grained model, In coarse-grained model, each thread handling an unknown pixel, tasks are independent, non-interfering. In order to obtain a higher degree of parallelism calculation, tasks of each thread are segmented to achieve a fine-grained version. In fine-grained version, 2 * 2, 2 * 4, 4 * 4 threads are adopted to solve process for an unknown pixel. Finally, using the original image 1024 * 1024, we obtained a speed up to 99.09 times.2. C-DPCM based lossless compression algorithm can get lower compression ratio, but the execution cycle is time costing. Aiming at this problem, we acheievd a parallel version using CUDA. Ultimately, we obatined 38 times speedup without Loss.