A Gpu Accelerated Algorithm For Compressive Sensing Based Video Super-resolution

Compressive sensing, also known as compressed sensing, is a technique motivated by recent results in sparse signal representation, which provides a new way to solve the problems in various field of traditional image processing. Compared to the traditional algorithm for image reconstruction, the algorithm based on compressive sensing can achieve a better result, a super-resolution image with more details, so the method of video super-resolution used in this paper is based on compressive sensing.The algorithm of image reconstruction based on compressive sensing is complex. Take the image used in this paper for example, the process of reconstructing a high-resolution image with the size of512×512from the low-resolution image with the size of256x256takes3-4seconds, meanwhile, the time used for reconstruct a frame of video is restricted at a very short time(less than100milliseconds). In order to achieve this goal, three optimization strategies are used in this paper:1) reduce time used for reconstruction by parallel computing;2) reduce data transfer between host and device;3) reduce the amount of computation by finding redundant data between adjacent video frames.With the aid of GPU, the compute capability of PC can achieve dozens or even hundreds of times compared to the PC equipped with CPU. In the process of designing parallel program, the most time-consuming parts are completed in GPU in the form of parallel computing, others are implemented in CPU. The experiment of this paper shows that parallel program can achieve a speed up of thirty-five times than serial program, which makes it possible to implement the algorithm of super-resolution based on compressive sensing on real-time video reconstruction.As the data of high-resolution image is allocated in global memory of GPU, the result can be shown directly on graphic card with the aid of interaction between openGL and CUDA, which reduce the data transfer between host and device, and the frame-rate of video super-resolution is increased.Most of the videos used in our daily life are compressed video, and the data redundancy between adjacent frames can find in the process of video decoding. There are many macro-blocks in p-frame with the mode of not_coded, which means that the data of the block in this frame is the same as the block of previous frame in the same position. As the reconstruction of previous macro-block has been completed, the computation of macro-block in this frame can be avoided. The amount of computation is reduced and the frame-rate of video super-resolution is increased.After the three optimization strategies are implemented in the algorithm, we can achieve a reconstruction of low-resolution with the size of256×256at the frame-rate of fifteen fps, and the reconstructed result of video have a better performance compared to the commonly used media player(such as windows media player).