Low-resolution Input Research Articles

Personalized Super Resolution with Face Prior

Face Super-Resolution (FSR) is a critical technology in computer vision that aims to reconstruct high-resolution facial images from low-resolution inputs. Despite recent advancements, current FSR methods struggle to accurately reconstruct personalized and detailed features. This paper proposes a novel FSR approach that addresses these challenges through a personalized feature extraction and fusion framework. Our method integrates a U-Net based downsampling mechanism to extract individual- specific features from high-resolution reference images, which are then fused with a pre-trained Generative Adversarial Network (GAN) for enhanced reconstruction. We introduce a comprehensive loss function that combines reconstruction, adversarial, facial component, and identity preservation losses to guide the learning process. Extensive experiments on the augmented FFHQ dataset demonstrate that our approach significantly improves the reconstruction of rich facial features, particularly for older individuals, outperforming existing state-of-the-art methods in both quantitative metrics and qualitative visual assessments.

Target-specified reference-based deep learning network for joint image deblurring and resolution enhancement in surgical zoom lens camera calibration

Background and objectiveFor the augmented reality of surgical navigation, which overlays a 3D model of the surgical target on an image, accurate camera calibration is imperative. However, when the checkerboard images for calibration are captured using a surgical microscope having high magnification, blur owing to the narrow depth of focus and blocking artifacts caused by limited resolution around the fine edges occur. These artifacts strongly affect the localization of corner points of the checkerboard in these images, resulting in inaccurate calibration, which leads to a large displacement in augmented reality. To solve this problem, in this study, we proposed a novel target-specific deep learning network that simultaneously enhances both the blur and spatial resolution of an image for surgical zoom lens camera calibration. MethodsAs a scheme of an end-to-end convolutional deep neural network, the proposed network is specifically intended for the checkerboard image enhancement used in camera calibration. Through the symmetric architecture of the network, which consists of encoding and decoding layers, the distinctive spatial features of the encoding layers are transferred and merged with the output of the decoding layers. Additionally, by integrating a multi-frame framework including subpixel motion estimation and ideal reference image with the symmetric architecture, joint image deblurring and enhanced resolution were efficiently achieved. ResultsFrom experimental comparisons, we verified the capability of the proposed method to improve the subjective and objective performances of surgical microscope calibration. Furthermore, we confirmed that the augmented reality overlap ratio, which quantitatively indicates augmented reality accuracy, from calibration with the enhanced image of the proposed method is higher than that of the previous methods. ConclusionsThese findings suggest that the proposed network provides sharp high-resolution images from blurry low-resolution inputs. Furthermore, we demonstrate superior performance in camera calibration by using surgical microscopic images, thus showing its potential applications in the field of practical surgical navigation.

SRARDA: A lightweight adaptive residual dense attention generative adversarial network for image super-resolution

Image super-resolution (SR) is the task of inferring a high resolution (HR) image from one/multiple single low resolution (LR) input(s). Traditional networks are evaluated by pixel-level metrics such as Peak-Signal-to-Noise Ratio (PSNR) etc., which do not always align with human perception of image quality. They often produce excessively smooth images that lack high-frequency texture and appear unnatural. Therefore, in this paper, we propose a lightweight adaptive residual dense attention generative adversarial network (SRARDA) for image SR. Firstly, our generator adopts the residual in residual (RIR) structure but redesigns the basic module. By using dynamic residual connection (ARC) to dynamically adjust the importance of residual and main paths, we design a novel adaptive residual dense attention block (ARDAB) that enhances the feature extraction capability of the generator. In addition, we build a high-frequency filtering unit (HFU) to extract more high-frequency features from the LR space. Finally, to fully utilize the discriminator, we use WGAN to compute the difference between the HR image and the reconstructed image. Experiments demonstrate that SRARDA effectively addresses the issue of excessive smoothing in reconstructed images, while also enhancing visual quality.

Lightweight image super-resolution with sliding Proxy Attention Network

Recently, image super-resolution (SR) models using window-based Transformers have demonstrated superior performance compared to SR models based on convolutional neural networks. Nevertheless, Transformer-based SR models often entail high computational demands. This is due to the adoption of shifted window self-attention following the window self-attention layer to model long-range relationships, resulting in additional computational overhead. Moreover, extracting local image features only with the self-attention mechanism is insufficient to reconstruct rich high-frequency image content. To overcome these challenges, we propose the Sliding Proxy Attention Network (SPAN), capable of recovering high-quality High-Resolution (HR) images from Low-Resolution (LR) inputs with substantially fewer model parameters and computational operations. The primary innovation of SPAN lies in the Sliding Proxy Transformer Block (SPTB), integrating the local detail sensitivity of convolution with the long-range dependency modeling of self-attention mechanism. Key components within SPTB include the Enhanced Local Feature Extraction Block (ELFEB) and the Sliding Proxy Attention Block (SPAB). ELFEB is designed to enhance the local receptive field with lightweight parameters for high-frequency details compensation. SPAB optimizes computational efficiency by implementing intra-window and cross-window attention in a single operation through leveraging window overlap. Experimental results demonstrate that SPAN can produce high-quality SR images while effectively managing computational complexity. The code is publicly available at: https://github.com/zononhzy/SPAN.

Revealing the Dark Side of Non-Local Attention in Single Image Super-Resolution.

Single Image Super-Resolution (SISR) aims to reconstruct a high-resolution image from its corresponding low-resolution input. A common technique to enhance the reconstruction quality is Non-Local Attention (NLA), which leverages self-similar texture patterns in images. However, we have made a novel finding that challenges the prevailing wisdom. Our research reveals that NLA can be detrimental to SISR and even produce severely distorted textures. For example, when dealing with severely degrade textures, NLA may generate unrealistic results due to the inconsistency of non-local texture patterns. This problem is overlooked by existing works, which only measure the average reconstruction quality of the whole image, without considering the potential risks of using NLA. To address this issue, we propose a new perspective for evaluating the reconstruction quality of NLA, by focusing on the sub-pixel level that matches the pixel-wise fusion manner of NLA. From this perspective, we provide the approximate reconstruction performance upper bound of NLA, which guides us to design a concise yet effective Texture-Fidelity Strategy (TFS) to mitigate the degradation caused by NLA. Moreover, the proposed TFS can be conveniently integrated into existing NLA-based SISR models as a general building block. Based on the TFS, we develop a Deep Texture-Fidelity Network (DTFN), which achieves state-of-the-art performance for SISR. Our code and a pre-trained DTFN are available on GitHub for verification.

A novel theoretical analysis on optimal pipeline of multi-frame image super-resolution using sparse coding

Super-resolution is the process of obtaining a high-resolution (HR) image from one or more low-resolution (LR) images. Single image super-resolution (SISR) deals with one LR image while multi-frame super-resolution (MFSR) employs several LR ones to reach the HR output. MFSR pipeline consists of alignment, fusion, and reconstruction. We conduct a theoretical analysis using sparse coding (SC) and iterative shrinkage-thresholding algorithm to fill the gap of mathematical justification in the execution order of the optimal MFSR pipeline. Our analysis recommends executing alignment and fusion before the reconstruction stage (whether through deconvolution or SISR techniques). The suggested order ensures enhanced performance in terms of peak signal-to-noise ratio and structural similarity index. The optimal pipeline also reduces computational complexity compared to intuitive approaches that apply SISR to each input LR image. Also, we demonstrate the usefulness of SC in analysis of computer vision tasks such as MFSR, leveraging the sparsity assumption in natural images. Simulation results support the findings of our theoretical analysis, both quantitatively and qualitatively.

Multi-scale transformer network for super-resolution of visible and thermal air images

Reference image-based Super-Resolution (RefSR) is introduced to improve the quality of a Low-resolution (LR) input image by leveraging the additional information provided by a High-Resolution (HR) reference image (Ref). While existing RefSR methods focus on thermal or visible flows separately, they often struggle to enhance the resolution of small objects such as Mini/Micro UAVs (Unmanned Aerial Vehicle) due to the resolution disparities between the input and reference images. To cope with these challenges when dealing with UAV early detection in context of video surveillance, we propose ThermoVisSR, a multiscale texture transformer for enhancing the Super-Resolution (SR) of visible and thermal images of Mini/Micro UAVs. Our approach tries to reconstruct the fine details of these objects while preserving their approximation (the body form and color of the different scene objects) already contained in the LR image. Hence, our model is divided up into two streams dealing separately with approximation and detail reconstruction. In the first one, we introduce a Convolution Neural Network (CNN) fusion backbone to extract the Low-Frequency (LF) approximation from the original LR image pairs. In the second one and to extract the details from the Ref image, our approach involves blending features from both visible and thermal sources to make the most of what each offer. Subsequently, we introduce the High-Frequency Texture Transformer (HFTT) across various resolutions of the merged features to ensure an accurate correspondence matching and significant transfer of High-Frequency (HF) patches from Ref to LR images. Moreover, to adapt the injection to the different bands well, we incorporate the separable software decoder (SSD) into the HFTT allowing to capture channel-specific details during the reconstruction phase. We validated our approach using a newly created dataset of Air images of Mini/Micro UAVs. Experimental results demonstrate that the proposed model consistently outperforms the state-of-the-art approaches on both qualitative and quantitative assessments.

Wavelet-fusion image super-resolution model with deep learning for downscaling remotely-sensed, multi-band spectral albedo imagery

Generating granular-scale surface albedo data is extremely important for solar photovoltaic site planning and to optimise renewable energy yield of bifacial panel installations. The albedo effect brings about a significant increase in power in bifacial photovoltaic systems, compared to their mono-facial counterparts, since the spectral response of bifacial solar panels correlates with the incident solar radiation wavelength on the back of the panel, to provide additional power generation capacity. Thus, harnessing the albedo data at relatively local scales is critical towards boosting solar power generation and providing greater power density in local electricity grids. This paper develops novel modelling approaches to produce high-resolution spectral albedo imagery across the Visible and Near Infrared (VNIR) bands, using the Wavelet-Fusion super-resolution model (i.e., Wavelet-FusionSR) trained with the Learned Gamma Correction approach by applying satellite image enhancement methodology. The proposed Wavelet-FusionSR model utilises the low-resolution moderate-resolution Imaging Spectroradiometer (MODIS) as well as high-resolution multi-spectral Advanced Space-borne Thermal Emission Reflection Radiometer (ASTER) satellite images, as critical inputs and ground-truth imagery, respectively, in order to perform sensor-to-sensor deep downscaling, without employing any synthetic or low-resolution satellite imagery data pairs. To augment the proposed deep learning algorithm across the decomposed sub-images of low-resolution inputs, we integrate local and global feature representation learning to train the proposed Wavelet-FusionSR model with Cauchy loss functions. In comparison with five competing benchmark models, the proposed Wavelet-FusionSR model demonstrates performance superiority using quantitative image downscaling metrics and visual assessments of the downscaled images for the visible band of solar radiation. The proposed Wavelet-FusionSR model yielded a Mean Square Error (MSE) of 0.00017, Signal-to-noise-ratio (PSNR) of 37.80, Structural Similarity Index (SSIM) of 0.999 and combined loss, MS-SSIMLoss, based on Multi Structural Similarity and Mean Absolute Error of 2.354 for the Visible Band images, and an MSE of 0.0014, PSNR of 28.43, SSIM of 0.999 and MS-SSIMLoss of 7.426 for the NIR spectral bands, demonstrating high efficacy of the proposed Wavelet-FusionSR method. The Wavelet-FusionSR method therefore attains high-resolution spectral albedo imagery outputs.

Comparison of super-resolution deep learning models for flow imaging

The primary goal of this study is to introduce deep learning (DL) methods as a cost-effective alternative to the computationally intensive Direct Numerical Simulation (DNS) simulations. We show that one can obtain a parametric field from a low-resolution input and map it to a fine grid output, significantly reducing the computational burden. We assess five super-resolution models for up-scaling low-resolution flow data into fine-grid numerical simulations’ output for accuracy and efficiency. The proposed architectures employ convolutional neural networks interconnected in encoder/decoder branches. We investigate these models using turbulent velocity fields inside a suddenly expanded channel characterized by complex features, including turbulence, instabilities, asymmetries, separation, and reattachment. Our results reveal that an encoder/decoder model with residual connections delivers the fastest results, a U-Net-based model with skip connections excels at producing sharper edges in regions prone to blurring, while deeper models incorporating maximum and average pooling layers show superior performance in reconstructing velocity profiles. These findings significantly contribute to our understanding of the potential of deep learning in fluid mechanics. The models presented in this study are trained and validated on standard computer hardware and can be easily adapted to other problems. The findings are promising for discovering and analyzing flow physics, highlighting the potential for DL techniques to improve the accuracy of the available fluid mechanics computational tools.

Downscaling sea surface height and currents in coastal regions using convolutional neural network

Downscaling is the process to obtain high-resolution data from low-resolution data. Recently statistical models using convolutional neural networks have gained popularity for fast downscaling of environmental fields, while their application to the coastal sea surface height and currents is lacking. This research aims to downscale sea surface height and depth-averaged current to a resolution of hundreds of meters in coastal regions with dynamic shorelines using convolutional neural networks. Hourly outputs over one year from a physical numerical model for a coastal region in the German Bight are used as the low-resolution input and high-resolution ground truth for the network. The results show that the network effectively reconstructs sea surface height and current in the region to a resolution of hundreds of meters with a scale factor of 16 or even 64, and accurately traces the moving sea surface and shorelines. The global mean absolute error and root mean square error for the sea surface height are found to be less than 0.03 m and 0.07 m, respectively, and for the current less than 0.03 m/s and 0.05 m/s, respectively. These values are around ten times smaller than those obtained from interpolation methods including nearest neighbor, bilinear and bicubic. The network also effectively replicates the distribution of high-resolution data. The errors in the reconstructed time average, 1st percentile and 99th percentile are significantly smaller than those from interpolation methods, especially for the current. These results highlight the ability of the network to downscale sea surface height and currents in regions with complex shorelines, and have implications for downscaling other coastal fields and shoreline tracking.

Comparison of different algorithms to improve the quality of Image

Super-resolution (SR) become the most important task in image processing because of requirement of high quality of images. It becomes popular in technology like computer vision and pattern recognition. In Single Image Super-resolution (SR) uses the sequences of single low resolution (LR) images from high resolution (HR) to improve the quality of image. The SR is an image processing problem which aims to generate a high resolution (HR) image from low resolution input image. The SR is difficult because of the missing information in the given LR image. Thus in SR it is very important to boost the reconstruction performance so that, there is need of effective prior knowledge. The prior knowledge of image will solve the problem to improve the quality of generated image. The main aim of SR is to generate the HR image with good visual perception as similar as original image. The paper compare two methods of SR that is bi-linear method, and direction group sparsity of image method that is proposed one. These methods will evaluate the accuracy, subjective visual effect, quality of image, time complexity and PSNR, etc. These calculations will enhance nature of picture like better visual impact quality, lower reconstruction error and adequate calculation proficiency, and so on

STDAN: Deformable Attention Network for Space-Time Video Super-Resolution.

The target of space-time video super-resolution (STVSR) is to increase the spatial-temporal resolution of low-resolution (LR) and low-frame-rate (LFR) videos. Recent approaches based on deep learning have made significant improvements, but most of them only use two adjacent frames, that is, short-term features, to synthesize the missing frame embedding, which cannot fully explore the information flow of consecutive input LR frames. In addition, existing STVSR models hardly exploit the temporal contexts explicitly to assist high-resolution (HR) frame reconstruction. To address these issues, in this article, we propose a deformable attention network called STDAN for STVSR. First, we devise a long short-term feature interpolation (LSTFI) module that is capable of excavating abundant content from more neighboring input frames for the interpolation process through a bidirectional recurrent neural network (RNN) structure. Second, we put forward a spatial-temporal deformable feature aggregation (STDFA) module, in which spatial and temporal contexts in dynamic video frames are adaptively captured and aggregated to enhance SR reconstruction. Experimental results on several datasets demonstrate that our approach outperforms state-of-the-art STVSR methods. The code is available at https://github.com/littlewhitesea/STDAN.

Terrain Amplification using Multi Scale Erosion

Modeling high-resolution terrains is a perennial challenge in the creation of virtual worlds. In this paper, we focus on the amplification of a low-resolution input terrain into a high-resolution, hydrologically consistent terrain featuring complex patterns by a multi-scale approach. Our framework combines the best of both worlds, relying on physics-inspired erosion models producing consistent erosion landmarks and introducing control at different scales, thus bridging the gap between physics-based erosion simulations and multi-scale procedural modeling. The method uses a fast and accurate approximation of different simulations, including thermal, stream power erosion and deposition performed at different scales to obtain a range of effects. Our approach provides landscape designers with tools for amplifying mountain ranges and valleys with consistent details.

Multi-Resolution Learning and Semantic Edge Enhancement for Super-Resolution Semantic Segmentation of Urban Scene Images.

Super-resolution semantic segmentation (SRSS) is a technique that aims to obtain high-resolution semantic segmentation results based on resolution-reduced input images. SRSS can significantly reduce computational cost and enable efficient, high-resolution semantic segmentation on mobile devices with limited resources. Some of the existing methods require modifications of the original semantic segmentation network structure or add additional and complicated processing modules, which limits the flexibility of actual deployment. Furthermore, the lack of detailed information in the low-resolution input image renders existing methods susceptible to misdetection at the semantic edges. To address the above problems, we propose a simple but effective framework called multi-resolution learning and semantic edge enhancement-based super-resolution semantic segmentation (MS-SRSS) which can be applied to any existing encoder-decoder based semantic segmentation network. Specifically, a multi-resolution learning mechanism (MRL) is proposed that enables the feature encoder of the semantic segmentation network to improve its feature extraction ability. Furthermore, we introduce a semantic edge enhancement loss (SEE) to alleviate the false detection at the semantic edges. We conduct extensive experiments on the three challenging benchmarks, Cityscapes, Pascal Context, and Pascal VOC 2012, to verify the effectiveness of our proposed MS-SRSS method. The experimental results show that, compared with the existing methods, our method can obtain the new state-of-the-art semantic segmentation performance.

High-flexibility reconstruction of small-scale motions in wall turbulence using a generalized zero-shot learning

This study proposes a novel super-resolution (or SR) framework for generating high-resolution turbulent boundary layer (TBL) flow from low-resolution inputs. The framework combines a super-resolution generative adversarial neural network (SRGAN) with down-sampling modules (DMs), integrating the residual of the continuity equation into the loss function. The DMs selectively filter out components with excessive energy dissipation in low-resolution fields prior to the super-resolution process. The framework iteratively applies the SRGAN and DM procedure to fully capture the energy cascade of multi-scale flow structures, collectively termed the SRGAN-based energy cascade reconstruction framework (EC-SRGAN). Despite being trained solely on turbulent channel flow data (via ‘zero-shot transfer’), EC-SRGAN exhibits remarkable generalization in predicting TBL small-scale velocity fields, accurately reproducing wavenumber spectra compared to direct numerical simulation (DNS) results. Furthermore, a super-resolution core is trained at a specific super-resolution ratio. By leveraging this pretrained super-resolution core, EC-SRGAN efficiently reconstructs TBL fields at multiple super-resolution ratios from various levels of low-resolution inputs, showcasing strong flexibility. By learning turbulent scale invariance, EC-SRGAN demonstrates robustness across different TBL datasets. These results underscore the potential of EC-SRGAN for generating and predicting wall turbulence with high flexibility, offering promising applications in addressing diverse TBL-related challenges.

Improved neurological diagnoses and treatment strategies via automated human brain tissue segmentation from clinical magnetic resonance imaging

ObjectiveSegmentation of medical images is a crucial process in various image analysis applications. Automated segmentation methods excel in accuracy when compared to manual segmentation in the context of medical image analysis. One of the essential phases in the quantitative analysis of the brain is automated brain tissue segmentation using clinically obtained magnetic resonance imaging (MRI) data. It allows for precise quantitative examination of the brain, which aids in diagnosis, identification, and classification of disorders. Consequently, the efficacy of the segmentation approach is crucial to disease diagnosis and treatment planning. MethodsThis study presented a hybrid optimization method for segmenting brain tissue in clinical MRI scans using a fractional Henry horse herd gas optimization-based Shepard convolutional neural network (FrHHGO-based ShCNN). To segment the clinical brain MRI images into white matter (WM), grey matter (GM), and cerebrospinal fluid (CSF) tissues, the proposed framework was evaluated on the Lifespan Human Connectome Projects (HCP) database. The hybrid optimization algorithm, FrHHGO, integrates the fractional Henry gas optimization (FHGO) and horse herd optimization (HHO) algorithms. Training required 30 min, whereas testing and segmentation of brain tissues from an unseen image required an average of 12 s. ResultsCompared to the results obtained with no refinements, the Skull stripping refinement showed significant improvement. As the method included a preprocessing stage, it was flexible enough to enhance image quality, allowing for better results even with low-resolution input. Maximum precision of 93.2%, recall of 91.5%, Dice score of 91.1%, and F1-score of 90.5% were achieved using the proposed FrHHGO-based ShCNN, which was superior to all other approaches. ConclusionThe proposed method may outperform existing state-of-the-art methodologies in qualitative and quantitative measurements across a wide range of medical modalities. It might demonstrate its potential for real-life clinical application.

Low-resolution few-shot learning via multi-space knowledge distillation

Existing few-shot classification models usually rely on limited known support images to form class centers, and classify query images based on the distance between their embedding and the class centers. However, these models assume that the query image is high-resolution (HR), and thus suffer from significant performance degradation when applied to low-resolution (LR) images. Due to the lack of discriminative information in LR images, there is a noticeable discrepancy between the embeddings of LR query images and the class centers formed by HR support images. To address this issue, we first formulate the problem of Low-Resolution Few-Shot Learning (LRFSL), where the support images are HR while the query images are only available in LR. Then, we propose an end-to-end pipeline that leverages mutual learning between a super-resolution (SR) network and a few-shot classification network. To further reduce the domain discrepancy between the embeddings of the SR images and HR class centers, we introduce a multi-space knowledge distillation strategy that aims to transfer pixel-level, feature-level, and logit-level knowledge of the HR domain to the SR domain. We conduct extensive experiments on classic few-shot datasets: miniImageNet, tieredImageNet, and the fine-grained few-shot dataset CUB. Experimental results show that our method can handle few-shot classification with LR input, and achieve performance that is almost comparable to using HR images as input. Specifically, our method achieves an average accuracy improvement of 26.47% with the Meta-Baseline Model and 7.44% with the Meta DeepBDC Model across all datasets compared to LR Query.

Meta-learning based blind image super-resolution approach to different degradations

Although recent studies on blind single image super-resolution (SISR) have achieved significant success, most of them typically require supervised training on synthetic low resolution (LR)-high resolution (HR) paired images. This leads to re-training necessity for different degradations and restricted applications in real-world scenarios with unfavorable inputs. In this paper, we propose an unsupervised blind SISR method with input underlying different degradations, named different degradations blind super-resolution (DDSR). It formulates a Gaussian modeling on blur degradation and employs a meta-learning framework for solving different image degradations. Specifically, a neural network-based kernel generator is optimized by learning from random kernel samples, referred to as random kernel learning. This operation provides effective initialization for blur degradation optimization. At the same time, a meta-learning framework is proposed to resolve multiple degradation modelings on the basis of alternative optimization between blur degradation and image restoration, respectively. Differing from the pre-trained deep-learning methods, the proposed DDSR is implemented in a plug-and-play manner, and is capable of restoring HR image from unfavorable LR input with degradations such as partial coverage, noise addition, and darkening. Extensive simulations illustrate the superior performance of the proposed DDSR approach compared to the state-of-the-arts on public datasets with comparable memory load and time consumption, yet exhibiting better application flexibility and convenience, and significantly better generalization ability towards multiple degradations. Our code is available at https://github.com/XYLGroup/DDSR.

Low-resolution facial emotion recognition on low-cost devices

The low-resolution input image is a crucial challenge for applying facial emotion recognition in real-world scenarios. The critical problem is that valuable object features are relatively lost in the extraction process due to their small size. On the other hand, this vision system is required by a machine to run smoothly on low-cost devices. Facial emotion recognition using a lightweight feature extractor is proposed in this study to effectively capture crucial facial components in a low-resolution image. To compromise the running speed, this work offers an efficient feature convolution to discriminate specific facial features. In addition, the system is embedded with an attentive module to capture important features and correlate them. Our model performance is evaluated on low-resolution public datasets achieving the accuracy of 97.34\%, 81.10\%, and 80.12\% on KDEF, RFDB, and FER-plus, respectively. The practical application demands that the deep learning model can operate fast on inexpensive devices. Consequently, the model achieved a speed of 290 FPS on a CPU device.

Sub2Full: split spectrum to boost optical coherence tomography despeckling without clean data.

Optical coherence tomography (OCT) suffers from speckle noise, causing the deterioration of image quality, especially in high-resolution modalities such as visible light OCT (vis-OCT). Here, we proposed an innovative self-supervised strategy called Sub2Full (S2F) for OCT despeckling without clean data. This approach works by acquiring two repeated B-scans, splitting the spectrum of the first repeat as a low-resolution input, and utilizing the full spectrum of the second repeat as the high-resolution target. The proposed method was validated on vis-OCT retinal images visualizing sublaminar structures in the outer retina and demonstrated superior performance over state-of-the-art Noise2Noise (N2N) and Noise2Void (N2V) schemes.