Unsupervised Learning for Real-World Super-Resolution

Acquiring degraded images with paired high-resolution (HR) images is often challenging, impeding the advance of image super-resolution in real-world applications. By generating realistic low-resolution (LR) images with degradation similar to that in real-world scenarios, simulated paired LR-HR data can be constructed for supervised training. However, most of the existing work ignores the degradation uncertainty of the generated realistic LR images, since only one LR image has been generated given an HR image. To address this weakness, we propose learning the degradation uncertainty of generated LR images and sampling multiple LR images from the learned LR image (mean) and degradation uncertainty (variance) and construct LR-HR pairs to train the super-resolution (SR) networks. Specifically, uncertainty can be learned by minimizing the proposed loss based on Kullback-Leibler (KL) divergence. Furthermore, the uncertainty in the feature domain is exploited by a novel perceptual loss; and we propose to calculate the adversarial loss from the gradient information in the SR stage for stable training performance and better visual quality. Experimental results on popular real-world datasets show that our proposed method has performed better than other unsupervised approaches.

The image super-resolution algorithm can overcome the imaging system's hardware limitation and obtain higher resolution and clearer images. Existing super-resolution methods based on convolutional neural networks(CNN) can learn the mapping relationship between high-resolution(HR) and low-resolution(LR) images. However, when the reconstruction target is a face image, the reconstruction results often have problems that the face area is too smooth and lacks details. We propose a guided cascaded face super-resolution network, called guided cascaded super-resolution network (GCFSRnet). GCFSRnet takes the LR image and a high-quality guided image as inputs, and it consists of a pose deformation module and a super-resolution network. Firstly, the pose deformation module converts the guide image's posture into the same as the low-resolution face image based on 3D fitting and 3D morphable model (3DMM). Then, the LR image and the deformed guide image are used as input of the super-resolution network. The super-resolution networks are formed by a cascade of two layers of networks, which extract different features. During the reconstruction process, the guide image can provide real facial details and help generate subtle facial textures. The cascade structure of a super-resolution network can gradually extract features and restore different levels of image details. The experimental results on the CASIA Web Face and CelebA datasets show that the proposed method can generate facial images with clear outlines and rich details, which are superior to other state-of-the-art methods such as SRResNet, SRGAN, VDSR, DBPN, etc.

High accuracy image registration is critical for the success of multi-frame super-resolution. Conventionally, the shift between images are estimated directly based on the under-sampled low-resolution (LR) image data. However, the high-frequency of LR data is unreliable due to the aliasing effect of sub-sampling, which will deteriorate the accuracy of registration. This paper proposes to resolve the aliasing by converting the LR images to high-resolution (HR) domain and then perform registration on the restored HR spectrum. To recover the HR spectrum for each corresponding LR image, we fuse the LR images into one HR image and project the estimation difference back to the reconstructed HR spectrum iteratively. To address the unequal reliabilities of different restored frequencies, weighted least square is employed to improve the precision of registration. Experimental results show that the proposed method can outperform other existing methods and improve the quality of super-resolution image.

Deep convolutional neural networks (CNNs), trained on corresponding pairs of high- and low-resolution images, achieve state-of-the-art performance in single-image super- resolution and surpass previous signal-processing based approaches. However, their performance is limited when applied to real photographs. The reason lies in their train- ing data: low-resolution (LR) images are obtained by bicu- bic interpolation of the corresponding high-resolution (HR) images. The applied convolution kernel significantly differs from real-world camera-blur. Consequently, while current CNNs well super-resolve bicubic-downsampled LR images, they often fail on camera-captured LR images. To improve generalization and robustness of deep super- resolution CNNs on real photographs, we present a ker- nel modeling super-resolution network (KMSR) that incor- porates blur-kernel modeling in the training. Our pro- posed KMSR consists of two stages: we first build a pool of realistic blur-kernels with a generative adversarial net- work (GAN) and then we train a super-resolution network with HR and corresponding LR images constructed with the generated kernels. Our extensive experimental vali- dations demonstrate the effectiveness of our single-image super-resolution approach on photographs with unknown blur-kernels.

Example learning-based image super-resolution (SR) is recognized as an effective way to produce a high-resolution (HR) image with the help of an external training set. The effectiveness of learning-based SR methods, however, depends highly upon the consistency between the supporting training set and low-resolution (LR) images to be handled. To reduce the adverse effect brought by incompatible high-frequency details in the training set, we propose a single image SR approach by learning multiscale self-similarities from an LR image itself. The proposed SR approach is based upon an observation that small patches in natural images tend to redundantly repeat themselves many times both within the same scale and across different scales. To synthesize the missing details, we establish the HR-LR patch pairs using the initial LR input and its down-sampled version to capture the similarities across different scales and utilize the neighbor embedding algorithm to estimate the relationship between the LR and HR image pairs. To fully exploit the similarities across various scales inside the input LR image, we accumulate the previous resultant images as training examples for the subsequent reconstruction processes and adopt a gradual magnification scheme to upscale the LR input to the desired size step by step. In addition, to preserve sharper edges and suppress aliasing artifacts, we further apply the nonlocal means method to learn the similarity within the same scale and formulate a nonlocal prior regularization term to well pose SR estimation under a reconstruction-based SR framework. Experimental results demonstrate that the proposed method can produce compelling SR recovery both quantitatively and perceptually in comparison with other state-of-the-art baselines.

Many deep learning (DL) frameworks have demonstrated state-of-the-art performance in the super-resolution (SR) task of magnetic resonance imaging, but most performances have been achieved with simulated low-resolution (LR) images rather than LR images from real acquisition. Due to the limited generalizability of the SR network, enhancement is not guaranteed for real LR images because of the unreality of the training LR images. In this study, we proposed a DL-based SR framework with an emphasis on data construction to achieve better performance on real LR MR images. The framework comprised two steps: (a) downsampling training using a generative adversarial network (GAN) to construct more realistic and perfectly matched LR/high-resolution (HR) pairs. The downsampling GAN input was real LR and HR images. The generator translated the HR images to LR images and the discriminator distinguished the patch-level difference between the synthetic and real LR images. (b) SR training was performed using an enhance4d deep super-resolution network (EDSR). In the controlled experiments, three EDSRs were trained using our proposed method, Gaussian blur, and k-space zero-filling. As for the data, liver MR images were obtained from 24 patients using breath-hold serial LR and HR scans (only HR images were used in the conventional methods). The k-space zero-filling group delivered almost zero enhancement on the real LR images and the Gaussian group produced a considerable number of artifacts. The proposed method exhibited significantly better resolution enhancement and fewer artifacts compared with the other two networks. Our method outperformed the Gaussian method by an improvement of 0.111 ± 0.016 in the structural similarity index and 2.76 ± 0.98 dB in the peak signal-to-noise ratio. The blind/reference-less image spatial quality evaluator metric of the conventional Gaussian method and proposed method were 46.6 ± 4.2 and 34.1 ± 2.4, respectively.

In this paper, we demonstrate simple algorithms that project low resolution (LR) images differing in subpixel shifts on a high resolution (HR) also called super resolution (SR) grid. The algorithms are very effective in accuracy as well as time efficiency. A number of spatial interpolation techniques using nearest neighbor, inverse-distance weighted averages, Radial Basis Functions (RBF) etc. are used in projection. For best accuracy of reconstructing SR image by a factor of two requires four LR images differing in four independent subpixel shifts. The algorithm has two steps: i) registration of low resolution images and (ii) shifting the low resolution images to align with reference image and projecting them on high resolution grid based on the shifts of each low resolution image using different interpolation techniques. Experiments are conducted by simulating low resolution images by subpixel shifts and subsampling of original high resolution image and the reconstructing the high resolution images from the simulated low resolution images. The results of accuracy of reconstruction are compared by using mean squared error measure between original high resolution image and reconstructed image. The algorithm was tested on remote sensing images and found to outperform previously proposed techniques such as Iterative Back Projection algorithm (IBP), Maximum Likelihood (ML) algorithms. The algorithms are robust and are not overly sensitive to the registration inaccuracies.

Learning based single image super-resolution (SISR) for real-world images has been an active research topic yet a challenging task, due to the lack of paired low-resolution (LR) and high-resolution (HR) training images. Most of the existing unsupervised real-world SISR methods adopt a two-stage training strategy by synthesizing realistic LR images from their HR counterparts first, then training the super-resolution (SR) models in a supervised manner. However, the training of image degradation and SR models in this strategy are separate, ignoring the inherent mutual dependency between downscaling and its inverse upscaling process. Additionally, the ill-posed nature of image degradation is not fully considered. In this paper, we propose an image downscaling and SR model dubbed as SDFlow, which simultaneously learns a bidirectional many-to-many mapping between real-world LR and HR images unsupervisedly. The main idea of SDFlow is to decouple image content and degradation information in the latent space, where content information distribution of LR and HR images is matched in a common latent space. Degradation information of the LR images and the high-frequency information of the HR images are fitted to an easy-to-sample conditional distribution. Experimental results on real-world image SR datasets indicate that SDFlow can generate diverse realistic LR and SR images both quantitatively and qualitatively.

Image upscaling aims to increase the resolution and size of a low resolution (LR) image in order to generate a high resolution (HR) image of high frequency (HF). There are several polynomial methods for obtaining a sharpened, upscaled HR image. The interpolated pixel is measured using a weighted average of the neighboring pixels within the image grid that blur at HF regions in these methods. Edge degradation is also caused by other edge-directed and learning-based upscaling methods, which produce blurring artifacts. A novel approach is proposed to fill these gaps. Using the concept of unsharp masking (USM), the LR image is blurred adaptively based on the region’s local variance. The sharpened high pass filtered (HPF) image is then obtained by subtracting the adaptively blurred image from the LR image. According to USM, the HPF image is combined with the LR image via a gain factor optimized using the cuckoo search (CS) algorithm. To compensate for the loss caused by upscaling, this pre-processing step is performed prior to interpolation. Aside from that, the edge of the B-spline interpolated image is detected and expanded. Edge expansion of the upscaled image is performed to further restore the HF details and reduce zigzag artifacts introduced by upscaling while also preserving the edge boundary. The proposed method outperforms the Lanczos, Bicubic, and Bilinear schemes in terms of peak signal to noise ratio (PSNR) gain of 3.475, 8.3839, and 8.075 dB, respectively. In terms of performance, this method outperforms state-of-the-art techniques both objectively (PSNR and SSIM) and subjectively (visual quality).

This paper presents a technique that used to enhance the image resolution. The input low resolution (LR) image is decomposed into low frequency (LF) and high frequency (HF) sub-bands by using lifting wavelet transform (LWT) and discrete wavelet transform (DWT). The resolution enhancement is done by interpolating HF sub-bands that are generated using LWT from the LR image. The HF sub-bands of DWT are used to preserve the edges of the images. The operation is done on HF sub-bands of LWT and DWT. The HF sub-bands that are generated from the DWT are interpolated and perform addition to the respective interpolated HF sub-bands of LWT. The HF sub-bands that are modified and the LR image will again interpolate and undergoes the inverse lifting wavelet transform (ILWT) to reconstruct the high resolution (HR) image. This implemented method results efficient over conventional methods.

Image super-resolution is of great interest in medical imaging field. However, different from natural images studied in computer vision field, the low-resolution (LR) medical imaging data is often a stack of high-resolution (HR) 2D slices with large slice thickness. Consequently, the goal of super-resolution for medical imaging data is to reconstruct the missing slice(s) between any two consecutive slices. Since some modalities (e.g., T1-weighted MR image) are often acquired with high-resolution (HR) image, it is intuitive to harness the prior self-similarity information in the HR image for guiding the super-resolution of LR image (e.g., T2-weighted MR image). The conventional way is to find the profile of patchwise self-similarity in the HR image and then use it to reconstruct the missing information at the same location of LR image. However, the local morphological patterns could vary significantly across the LR and HR images, due to the use of different imaging protocols. Therefore, such direct (un-supervised) adaption of self-similarity profile from HR image is often not effective in revealing the actual information in the LR image. To this end, we propose to employ the existing image information in the LR image to supervise the estimation of self-similarity profile by requiring it not only being optimal in representing patches in the HR image, but also producing less reconstruction errors for the existing image information in the LR image. Moreover, to make the anatomical structures spatially consistent in the reconstructed image, we simultaneously estimate the self-similarity profiles for a stack of patches across consecutive slices by solving a group sparse patch representation problem. We have evaluated our proposed super-resolution method on both simulated brain MR images and real patient images with multiple sclerosis lesion, achieving promising results with more anatomical details and sharpness.

We propose a new high-quality up-scaling technique that extends the existing example based super-resolution (SR) framework. Our approach is based on the fundamental idea that a low-resolution (LR) image could be generated from any of the multiple possible high-resolution (HR) images. Therefore it would be more natural to use multiple predictors of HR patch from LR patch instead of single one. In this work we build a generic framework to estimate an HR image from LR one using an adaptive prior (select the predictor locally) based on the local statistics of LR images. We use natural image patch prior as the HR image statistics. We partition the natural images into documents and group them to discover the inherent topics using probabilistic Latent Semantic Analysis (pLSA) and also learn the dual dictionaries of HR and LR image patch pairs for each of the topics using sparse dictionary learning technique. Then for test image we infer locally which topic it corresponds to and then we use the corresponding learned dual dictionary to generate HR image. Experimental results show the effectiveness of our method over existing state-of-art methods.KeywordsSparse RepresentationSparse CodeSuper ResolutionProbabilistic Latent Semantic AnalysisBicubic InterpolationThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Deep Feature Blend Attention: A New Frontier in Super Resolution Image Generation

Modern face recognition systems extract face representations using deep neural networks (DNNs) and give excellent identification and verification results, when tested on high resolution (HR) images. However, the performance of such an algorithm degrades significantly for low resolution (LR) images. A straight forward solution could be to train a DNN, using simultaneously, high and low resolution face images. This approach yields a definite improvement at lower resolutions but suffers a performance degradation for high resolution images. To overcome this shortcoming, we propose to train a network using both HR and LR images under the guidance of a fixed network, pretrained on HR face images. The guidance is provided by minimising the KL-divergence between the output Softmax probabilities of the pretrained (i.e., Teacher) and trainable (i.e., Student) network as well as by sharing the Softmax weights between the two networks. The resulting solution is tested on down-sampled images from FaceScrub and MegaFace datasets and shows a consistent performance improvement across various resolutions. We also tested our proposed solution on standard LR benchmarks such as TinyFace and SCFace. Our algorithm consistently outperforms the state-of-the-art methods on these datasets, confirming the effectiveness and merits of the proposed method.

In this paper we propose a fast approach for edge preserving super-resolution (SR) based on learning of contourlet coefficients. Given a low resolution test image, we first obtain an initial HR estimate i.e., a close approximation to SR image by learning the contourlet coefficients from a training database consisting of low resolution (LR) and high resolution (HR) images. The final SR image is obtained by using a regularization framework in which both the SR and the LR images are modeled as separate homogeneous Markov Random Fields (MRFs). The LR image formation process is modeled as a decimated and noisy version of the SR image and the final cost function is minimized by using a gradient descent method. Novelty of our approach lies in preserving the edges in the final SR image while using a non edge preserving MRF prior. This is definitely advantageous since it avoids the use of discontinuity preserving prior and hence the computationally taxing optimization methods. The edges in the final SR correspond to those learned from the initial HR estimate. The use of MRF on the low resolution image imposes an additional constraint on the final solution and hence we expect a better solution. In addition, we use the initial HR image for estimating the decimation matrix entries as well as for learning the corresponding MRF parameter. We show the effectiveness of the proposed approach by conducting the experiments on images captured using a real camera.