Real-time Video Image Deblurring Algorithm With GPU

The performance of image acquisition devices has been improved dramatically in recent years, especially like resolution and clarity. But defocusing and motion blur are still big problems to most ordinary image acquisition devices. Upgrading the devices with better hardware is one way to solve the problem, but the costs will usually increase disproportionately compare with what we get. Using appropriate image restoration algorithm is another cost-effective solution, which can improve the clarity and recognition rate of images significantly. However, the huge computation of those image restoration algorithms makes it unpractical. A new image restoration algorithm based on video target detection, and a accelerating method using the Graphic Processing Units (GPU) parallel computing architecture are proposed in this paper.. Parallelizing the proposed algorithm makes it efficient enough to handle high-definition (HD) video processing in real time.GPU is designed to maximize performance for parallel computing, and it’s not good at complex logical control, which would lower the computing efficiency. Based on this parallel computing architecture, the proposed image restoration algorithm is separated into many small routines with simple logical controls. These routines are put into GPU cores to execute the parallel computing. Because the bandwidth between GPU device memory and host memory is much smaller than bandwidth between device and device memory, data transmission between host and device is reduced as little as possible in this algorithm, which increases the computing efficiency even further. Given the fact that in most cases, users are only interested in the moving parts in a video instead of the background, the proposed algorithm does motion detection first, and then deblur the detected regions. This makes sure only the interested blurred regions get restored, and other parts of the image will not be impacted.This paper uses the Compute Unified Device Architecture (CUDA) of NVIDIA to achieve GPU parallel computing and accelerate the proposed algorithm. And open-source image processing library OpenCV is used for implement, which makes codes portable and can be easily applied in a real system.