Multi-frame super resolution based on sparse representation and matrix completion

Super-resolved reconstruction of images can yield poor results in the absence of extensively-trained dictionaries. A superresolution algorithm is presented which remedies this problem by exploiting recent results from the work on sparse representation and matrix completion. An over-complete dictionary pair is trained using natural image data. Sparse coefficients of low-resolution image patches are estimated using local prior constraints. In multi-frame images, sparse coefficients are similar across frames, and the Inexact Augmented Lagrange Multiplier method is employed to achieve matrix completion and recovery in the process of imposing global constraints. The final high-resolution image is generated from the output low-rank matrix. Experiments reveal that the method yields higher PSNR value than other mainstream SR algorithms, produces perceptibly superior edges and details, and is more robust to dictionary insufficiency.