Research On Motion Blur Image Restoration Based On FPGA And GPU

Image restoration is an inverse process of image degradation,but even the degradation model is known,the solution of clear images is still ill-posed.In order to solve this problem,image restoration algorithm usually requires lots of computation and long running time,which affects the practical application.Therefore,it is of great significance to use FPGA and GPU to accelerate the image restoration algorithm.In this paper,the spatially invariant motion blur and the spatially variant motion blur were studied respectively.For spatially invariant motion blur,fixed-point Richardson-Lucy algorithm was implemented in Matlab,and hardware accelerator was designed according to fixed-point algorithm and it is implemented on FPGA.For spatially variant motion blur problem,after researching PMRL algorithm,this paper found the noise amplification problem of the algorithm when restoring noisy image and put forward noise adaptive PMRL algorithm combined with noise estimation algorithm.In order to reduce the running time of noise adaptive PMRL algorithm,GPU acceleration scheme was developed after analyzing the parallelism of the algorithm,and CUDA was used to implement it.Test results in FPGA show that the Richardson-Lucy algorithm hardware accelerator can achieve 15.1 times of acceleration compared to CPU when processing 64x64 image.Test results of image data set show that compared with PMRL algorithm,the noise adaptive PMRL algorithm can reduce the RMS error by 29%when restoring noisy image.Test results in GPU show that the noise adaptive PMRL algorithm implemented in GPU can achieve 27.7 times of acceleration compared to CPU when restoring 1024x 1024 image.