Single Image Super-resolution Algorithms Based On Gaussian Process Regression

Due to the limitations of physical resolution of imaging devices or imaging environments,it is not always easy to obtain high-resolution(HR)images directly from underlying scenes.To overcome this difficulty,super-resolution(SR)technique is proposed to reconstruct a la-tent HR image from one or several low resolution(LR)images from the same scene.This technique has been shown great potential in many practical applications such as digital enter-tainment,medical imaging,video surveillance,criminal investigation,and satellite remote sensing.Gaussian process regression(GPR)is an effective non-linear kernel method,widely applied in SR reconstruction.However,due to the high complexity of GPR,there still exists room to improve the performance.Compared with multi-frame SR,single image SR(SISR)applies.wider and has fewer restrictions,and it has become the main direction of the SR research.Therefore,this thesis focuses on the study of SISR based on GPR.According to the charac-teristics of GPR,the thesis proposes some SISR methods to improve the model complexity,while keeping higher reconstruction quality and promoting the efficiency.The main contri-butions of this thesis are summarized as follows:1.To improve the efficiency,most existing GPR-based self-learning SR methods are lo-cal tile-based,which cannot make full use of self-similarities existing in natural images.A non-local self-learning SR framework based on GPR is proposed to address this problem,which learns only one non-local GPR model instead of multiple local tile-based models for SR reconstruction.In the proposed framework,an anisotropic linear kernel is introduced to construct a new kernel function for capturing more structure similarity.Furthermore,a simple grid patch sampling with moderate sampling interval can be used to speed up the SR processing significantly without compromise of reconstruction quality.In addition,two important factors that relate to the performance of the GPR,i.e.,predictive variance and the condition number of the kernel matrix,are discussed theoretically.Experimental results on the benchmark test images show that the proposed method is superior to other state-of-art competitors in terms of both quantitative and qualitative measurements.2.When applying GPR to example learning-based super-resolution,two outstanding is-sues remain.One is its high computational complexity restricts SR application when a large dataset is available for learning task.The other is that the commonly used Gaussian like-lihood in GPR is incompatible with the true observation model for SR reconstruction.To alleviate the above two issues,we propose a GPR-based SR method by using dictionary-based sampling and student-t likelihood,called DSGPR.Considering that dictionary atoms effectively span the original training sample space,we adopt a dictionary-based sampling strategy by combining all the neighborhood samples of each atom into a compact representa-tive training subset so as to reduce the computational complexity.Based on statistical tests,we statistically validate that Student-t likelihood is more suitable to build the observation model for GPR.Extensive experimental results show that the proposed method outperforms other competitors and produces more pleasing details in texture regions.3.A novel example learning-based SR method,based on active-sampling Gaussian pro-cess regression(AGPR),is proposed to alleviate the remarkably computational cost of the GPR-based SR.The newly proposed approach employs active learning strategy to heuris-tically select more informative samples for training the regression parameters of the GPR model,which shows significant improvement on computational efficiency while keeping higher quality of reconstructed image.Moreover,an accelerating scheme is suggested to further reduce the time complexity of the proposed AGPR-based SR by using a pre-learned projection matrix.Objective and subjective experimental results demonstrate that the pro-posed method is superior to other competitors for producing much sharper edges and finer details.4.Although the proposed active-sampling strategy using active learning can extract a small informative subset from a large training set to overcome the bottleneck of GP regression based SR,the method cannot guarantee SR quality as well as efficiency when the subset's size is too small.In order to have a nearly real-time GP-based SR,a sparse GPR-based SR method is proposed by integrating the active-sampling and traditional sparse GP.The proposed framework is based on the statistics that the model projection vector is approxi-mately sparse.To be more specific,it first trains a exact GP model based on an informative subset obtained by active sampling from the original training dataset.And then it proposes to employ the sparse GP to further approximate the exact GP model by seeking a sparse projection vector,which can significantly accelerate the prediction efficiency while keeping higher reconstruction quality.The proposed method is fundamentally coarse-to-fine.Exten-sive experimental results indicate that the proposed method is superior to other state-of-art competitors and is promising for real-time SR application.