Research On Face Image Super-resolution Based On Topology ICA And Sparse Coding

The study of face image problem is always one of the hot topics in the digital imageprocess, the pattern recognition and computer vision field, etc. However, subject to theaffection of device, transmission bandwidth, hardware limitations, the resolution ofobtained face image is often very low, and low-resolution face images bring the humanface recognition and other applications great difficulty. The development of image superresolution technique offers effective solution for this puzzle. By using a low resolutionimage as the input, it can produce a high resolution image through the signal-processingtechnology. It is widely applied in many important fields, such as surveillance video,medical images, video transmission and so on.In this paper, we propose a new face super-resolution algorithm based on topologyICA and sparse coding on the basis of learning-based face super-resolution, in which weuse the characteristic of face image structural similarity. This paper first introduces theresearch background and the state of the art of face super-resolution and then analyzes thekey techniques of super-resolution. In the following, this paper introduces the relevantknowledge of sparse representation and topology ICA. After that, the learning dictionary isobtained based on the topology ICA model. Finally, the high-resolution (HR) face image isreconstructed according to sparse representation theory.The main innovations and contributions in this paper are as follows:(1) The model of topology ICA is used to extract the couple dictionary of high-andlow-resolution face image, and then the couple dictionary is divided into high-andlow-resolution dictionary according to the dimension of training sample. In the end,low-resolution dictionary is used to linearly represent input low-resolution (LR) face.(2) Sparse representation theory is used to solve the coefficients of input LR image interms of LR dictionary. And HR image is got by replacing the LR dictionary with HRdictionary on the premise that the coefficient is invariant.The experimental results in this paper illustrate the reconstructed HR face images aremore suitable to be observed by human eyes as well as have better objective quality than other algorithm.