Regularization-based Multi-frame Image Super-resolution Reconstruction And Hardware Implementation

Super-resolution reconstruction is a kind of image processing technique, which utilizes the redundant information among the multi-frame low-resolution images with sub-pixel shifts to obtain the high-resolution image with better quality. It can enhance the resolution and restore the details and grains effectively without improving the hardware condition, which has potential applications. The multi-frame based super-resolution reconstruction is studied in this thesis. The focal studies include non-convex bilateral filter based and partial differential equation based regularization of super-resolution, and the improved algorithm is accelerated by hardware platform. The main work of this paper is as follows:1. Combining the 1/2 non-convex regularization with bilateral filters, the bilateral total variation(BTV) is improved and an 1/2-BTV operator is derived. The improved regularization term can produce sparser solution than BTV and fit the heavy-tailed distribution of real-world image better. Due to the non-convexity of 1/2-BTV, the solving method is transformed into iterative reweighted least squares. Both simulated and real images are chosen for experiments. The results show that the 1/2-BTV regularization term gets better reconstruction effects. It depresses the noise and retains the edges, and also promotes both the evaluating parameters and visual effects.2. After analyzing the spatial distribution of an image, an adaptive reweighted function is introduced to promote the Laplacian Gaussian combined model. The promoted operator emphasizes the Laplacian operator in smooth region and emphasizes the Gaussian operator in edge region, so as to make the visual effect better. Then it is used as a regularization term in super-resolution reconstruction. In the experiments, other partial differential equation models are also introduced as regularization terms to compare with the improved model. The results show that the improved regularization term has better performances in both visual effects and objective evaluations.3. In the respect of the hardware implementation of super-resolution reconstruction, this paper implements the algorithm in the DSP core C674 x of DaVinci digital media platform. The code of the algorithm is transformed into standard C language in CCS IDE. Then several code optimization methods are adopted, such as compiler options optimization, internal memory distribution and the use of intrinsics instructions. The final results show that the algorithm is greatly accelerated on the DSP platform, which can be applied in many fields such as feature extraction and target recognition, and it is practical in engineering.