Mutiframe Super Resolution Based On Peer Group And Regularization

Super resolution construction is a technology that produces a high resolutionimage from a set of low resolution images of the same scene, and one of the mainbranches of image enhancement. However, there are still some problems from theexisting algorithms, as imaging models are assumed to be the white Gaussian noise,and they can hardly deal with the other types of additional noise, such as impulsenoise. Furthermore, edge preserving performance is a particularly important measure.To solve these problems, we proposed two main improvements in this paper.Firstly, to apply to the impulse noise model, we adopt the peer group technology fromthe field image denoising. By considering the reliability of the pixels, weighted eachpixel by the peer group way, and the weight represents the similarity between currentpixel and its neighbors, then integrate the metrics into high-resolution imagereconstruction. Secondly, the traditional BTV regularization method processes eachpixel with the same weight, which doesn’t consider the possibility of the noise, andnot handle them properly. Our approach employs the reliability of each pixel toconstrain the fidelity term of regularization formula, thus preserves more reliablefeatures of low-resolution images, and attenuates the unreliable features, finallyimproves the quality of high-resolution images.Experiments shows that proposed algorithm achieves significant high-resolutionresult with edges simultaneously is robust to noise not only for white Gaussian noise,but also for the photon shot type outliers, such as impulse noise. Meanwhile, proposedalgorithm is based on the similarity of spatial small range neighbor pixels, which issample to implement and costs cheap in calculation.