Research Of The Reference-based Super-resolution Methods On The Light Field Refocused Images

Light field imaging has gained attention since its proposal,and the ability of light field cameras to refocus after exposure has caused a growth in light field research.Light field cameras have wide-ranging applications in fields such as photography,industrial health,and bioscience.However,low-resolution results and high complexity in refocusing algorithms have become the main obstacles to practical expansion.Currently,many studies on light field super-resolution are being conducted,and they can be divided into two categories: angular super-resolution and spatial super-resolution.In this thesis,we mainly focus on improving the efficiency of light field refocusing by developing an efficient super-resolution algorithm to generate precise light field refocused images.Firstly,traditional guided upsampling methods have overly simplistic weight design,which makes them unsuitable for light field refocused image upsampling.To address this issue,we propose a deep weighted guided upsampling network(DWGUN)to super-resolve low-resolution refocused images with high-resolution guidance images.We also propose a deep refocus-defocus edge-aware module(DREAM)to learn spatially-varying weights and embed it into the DWGUN.Experiments illustrate that our proposed method achieves state-of-the-art performance,both quantitatively and qualitatively,and significantly reduces the model’s parameters and time.As existing datasets are limited for training and evaluation,we propose a dataset with over 900 training and over 150 testing samples to improve dataset diversity in the light field image guided upsampling application.Secondly,the assumption of local linear relations between input and guidance images is barely valid in our application scenario,which motivated us to use a neural network to construct the relationship between input and guidance images directly.We propose a novel solution based on a linear fusion network to guided upsample lowresolution refocused images.A linear fusion block is designed and stacked into a lightweight fully convolutional network named linear fusion network(LFN).Our LFN outperforms 14 existing methods,achieving first place in PSNR and second place in SSIM with only 5 percent of calculations.Qualitative comparisons demonstrate that our method LFN not only preserves edge information in the focus area but also smoothly blurs the defocused area.Compared to the first proposed method,it significantly improves upsampling accuracy while maintaining high efficiency.