PyTorchKR Papers

최신 논문

총 3,696편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2202.03716

Binary Neural Networks as a general-propose compute paradigm for on-device computer vision

For binary neural networks (BNNs) to become the mainstream on-device computer vision algorithm, they must achieve a superior speed-vs-accuracy tradeoff than 8-bit quantization and establish a similar degree of general applicability in vision tasks. To this end, we propose a BNN framework comprising 1) a minimalistic inference scheme for hardware-friendliness, 2) an over-parameterized training scheme for high accuracy, and 3) a simple procedure to adapt to different vision tasks. The resultant framework overtakes 8-bit quantization in the speed-vs-accuracy tradeoff for classification, detection, segmentation, super-resolution and matching: our BNNs not only retain the accuracy levels of their 8-bit baselines but also showcase 1.3-2.4$\times$ faster FPS on mobile CPUs. Similar conclusions can be drawn for prototypical systolic-array-based AI accelerators, where our BNNs promise 2.8-7$\times$ fewer execution cycles than 8-bit and 2.1-2.7$\times$ fewer cycles than alternative BNN designs. These results suggest that the time for large-scale BNN adoption could be upon us.

arXiv:2202.03674공식 코드 · Papers with Code

Trained Model in Supervised Deep Learning is a Conditional Risk Minimizer

We proved that a trained model in supervised deep learning minimizes the conditional risk for each input (Theorem 2.1). This property provided insights into the behavior of trained models and established a connection between supervised and unsupervised learning in some cases. In addition, when the labels are intractable but can be written as a conditional risk minimizer, we proved an equivalent form of the original supervised learning problem with accessible labels (Theorem 2.2). We demonstrated that many existing works, such as Noise2Score, Noise2Noise and score function estimation can be explained by our theorem. Moreover, we derived a property of classification problem with noisy labels using Theorem 2.1 and validated it using MNIST dataset. Furthermore, We proposed a method to estimate uncertainty in image super-resolution based on Theorem 2.2 and validated it using ImageNet dataset. Our code is available on github.

arXiv:2202.03564

Accurate super-resolution low-field brain MRI

The recent introduction of portable, low-field MRI (LF-MRI) into the clinical setting has the potential to transform neuroimaging. However, LF-MRI is limited by lower resolution and signal-to-noise ratio, leading to incomplete characterization of brain regions. To address this challenge, recent advances in machine learning facilitate the synthesis of higher resolution images derived from one or multiple lower resolution scans. Here, we report the extension of a machine learning super-resolution (SR) algorithm to synthesize 1 mm isotropic MPRAGE-like scans from LF-MRI T1-weighted and T2-weighted sequences. Our initial results on a paired dataset of LF and high-field (HF, 1.5T-3T) clinical scans show that: (i) application of available automated segmentation tools directly to LF-MRI images falters; but (ii) segmentation tools succeed when applied to SR images with high correlation to gold standard measurements from HF-MRI (e.g., r = 0.85 for hippocampal volume, r = 0.84 for the thalamus, r = 0.92 for the whole cerebrum). This work demonstrates proof-of-principle post-processing image enhancement from lower resolution LF-MRI sequences. These results lay the foundation for future work to enhance the detection of normal and abnormal image findings at LF and ultimately improve the diagnostic performance of LF-MRI. Our tools are publicly available on FreeSurfer (surfer.nmr.mgh.harvard.edu/).

arXiv:2202.03163공식 코드 · Papers with Code

Patch-Based Stochastic Attention for Image Editing

Attention mechanisms have become of crucial importance in deep learning in recent years. These non-local operations, which are similar to traditional patch-based methods in image processing, complement local convolutions. However, computing the full attention matrix is an expensive step with heavy memory and computational loads. These limitations curb network architectures and performances, in particular for the case of high resolution images. We propose an efficient attention layer based on the stochastic algorithm PatchMatch, which is used for determining approximate nearest neighbors. We refer to our proposed layer as a "Patch-based Stochastic Attention Layer" (PSAL). Furthermore, we propose different approaches, based on patch aggregation, to ensure the differentiability of PSAL, thus allowing end-to-end training of any network containing our layer. PSAL has a small memory footprint and can therefore scale to high resolution images. It maintains this footprint without sacrificing spatial precision and globality of the nearest neighbors, which means that it can be easily inserted in any level of a deep architecture, even in shallower levels. We demonstrate the usefulness of PSAL on several image editing tasks, such as image inpainting, guided image colorization, and single-image super-resolution. Our code is available at: https://github.com/ncherel/psal

arXiv:2202.03046공식 코드 · Papers with Code

A new face swap method for image and video domains: a technical report

Deep fake technology became a hot field of research in the last few years. Researchers investigate sophisticated Generative Adversarial Networks (GAN), autoencoders, and other approaches to establish precise and robust algorithms for face swapping. Achieved results show that the deep fake unsupervised synthesis task has problems in terms of the visual quality of generated data. These problems usually lead to high fake detection accuracy when an expert analyzes them. The first problem is that existing image-to-image approaches do not consider video domain specificity and frame-by-frame processing leads to face jittering and other clearly visible distortions. Another problem is the generated data resolution, which is low for many existing methods due to high computational complexity. The third problem appears when the source face has larger proportions (like bigger cheeks), and after replacement it becomes visible on the face border. Our main goal was to develop such an approach that could solve these problems and outperform existing solutions on a number of clue metrics. We introduce a new face swap pipeline that is based on FaceShifter architecture and fixes the problems stated above. With a new eye loss function, super-resolution block, and Gaussian-based face mask generation leads to improvements in quality which is confirmed during evaluation.

arXiv:2202.01731

Fast Online Video Super-Resolution with Deformable Attention Pyramid

Video super-resolution (VSR) has many applications that pose strict causal, real-time, and latency constraints, including video streaming and TV. We address the VSR problem under these settings, which poses additional important challenges since information from future frames is unavailable. Importantly, designing efficient, yet effective frame alignment and fusion modules remain central problems. In this work, we propose a recurrent VSR architecture based on a deformable attention pyramid (DAP). Our DAP aligns and integrates information from the recurrent state into the current frame prediction. To circumvent the computational cost of traditional attention-based methods, we only attend to a limited number of spatial locations, which are dynamically predicted by the DAP. Comprehensive experiments and analysis of the proposed key innovations show the effectiveness of our approach. We significantly reduce processing time and computational complexity in comparison to state-of-the-art methods, while maintaining a high performance. We surpass state-of-the-art method EDVR-M on two standard benchmarks with a speed-up of over $3\times$.

arXiv:2202.01116

An Optimal Transport Perspective on Unpaired Image Super-Resolution

Real-world image super-resolution (SR) tasks often do not have paired datasets, which limits the application of supervised techniques. As a result, the tasks are usually approached by unpaired techniques based on Generative Adversarial Networks (GANs), which yield complex training losses with several regularization terms, e.g., content or identity losses. We theoretically investigate optimization problems which arise in such models and find two surprizing observations. First, the learned SR map is always an optimal transport (OT) map. Second, we theoretically prove and empirically show that the learned map is biased, i.e., it does not actually transform the distribution of low-resolution images to high-resolution ones. Inspired by these findings, we propose an algorithm for unpaired SR which learns an unbiased OT map for the perceptual transport cost. Unlike the existing GAN-based alternatives, our algorithm has a simple optimization objective reducing the need for complex hyperparameter selection and an application of additional regularizations. At the same time, it provides a nearly state-of-the-art performance on the large-scale unpaired AIM19 dataset.

arXiv:2202.00997공식 코드 · Papers with Code

Gradient Variance Loss for Structure-Enhanced Image Super-Resolution

Recent success in the field of single image super-resolution (SISR) is achieved by optimizing deep convolutional neural networks (CNNs) in the image space with the L1 or L2 loss. However, when trained with these loss functions, models usually fail to recover sharp edges present in the high-resolution (HR) images for the reason that the model tends to give a statistical average of potential HR solutions. During our research, we observe that gradient maps of images generated by the models trained with the L1 or L2 loss have significantly lower variance than the gradient maps of the original high-resolution images. In this work, we propose to alleviate the above issue by introducing a structure-enhancing loss function, coined Gradient Variance (GV) loss, and generate textures with perceptual-pleasant details. Specifically, during the training of the model, we extract patches from the gradient maps of the target and generated output, calculate the variance of each patch and form variance maps for these two images. Further, we minimize the distance between the computed variance maps to enforce the model to produce high variance gradient maps that will lead to the generation of high-resolution images with sharper edges. Experimental results show that the GV loss can significantly improve both Structure Similarity (SSIM) and peak signal-to-noise ratio (PSNR) performance of existing image super-resolution (SR) deep learning models.

arXiv:2202.00416

CAESR: Conditional Autoencoder and Super-Resolution for Learned Spatial Scalability

In this paper, we present CAESR, an hybrid learning-based coding approach for spatial scalability based on the versatile video coding (VVC) standard. Our framework considers a low-resolution signal encoded with VVC intra-mode as a base-layer (BL), and a deep conditional autoencoder with hyperprior (AE-HP) as an enhancement-layer (EL) model. The EL encoder takes as inputs both the upscaled BL reconstruction and the original image. Our approach relies on conditional coding that learns the optimal mixture of the source and the upscaled BL image, enabling better performance than residual coding. On the decoder side, a super-resolution (SR) module is used to recover high-resolution details and invert the conditional coding process. Experimental results have shown that our solution is competitive with the VVC full-resolution intra coding while being scalable.

arXiv:2201.12348공식 코드 · Papers with Code

End-to-End Optimization of Metasurfaces for Imaging with Compressed Sensing

We present a framework for the end-to-end optimization of metasurface imaging systems that reconstruct targets using compressed sensing, a technique for solving underdetermined imaging problems when the target object exhibits sparsity (i.e. the object can be described by a small number of non-zero values, but the positions of these values are unknown). We nest an iterative, unapproximated compressed sensing reconstruction algorithm into our end-to-end optimization pipeline, resulting in an interpretable, data-efficient method for maximally leveraging metaoptics to exploit object sparsity. We apply our framework to super-resolution imaging and high-resolution depth imaging with a phase-change material. In both situations, our end-to-end framework computationally discovers optimal metasurface structures for compressed sensing recovery, automatically balancing a number of complicated design considerations to select an imaging measurement matrix from a complex, physically constrained manifold with millions ofdimensions. The optimized metasurface imaging systems are robust to noise, significantly improving over random scattering surfaces and approaching the ideal compressed sensing performance of a Gaussian matrix, showing how a physical metasurface system can demonstrably approach the mathematical limits of compressed sensing.

arXiv:2201.12576

Scale-arbitrary Invertible Image Downscaling

Conventional social media platforms usually downscale the HR images to restrict their resolution to a specific size for saving transmission/storage cost, which leads to the super-resolution (SR) being highly ill-posed. Recent invertible image downscaling methods jointly model the downscaling/upscaling problems and achieve significant improvements. However, they only consider fixed integer scale factors that cannot downscale HR images with various resolutions to meet the resolution restriction of social media platforms. In this paper, we propose a scale-Arbitrary Invertible image Downscaling Network (AIDN), to natively downscale HR images with arbitrary scale factors. Meanwhile, the HR information is embedded in the downscaled low-resolution (LR) counterparts in a nearly imperceptible form such that our AIDN can also restore the original HR images solely from the LR images. The key to supporting arbitrary scale factors is our proposed Conditional Resampling Module (CRM) that conditions the downscaling/upscaling kernels and sampling locations on both scale factors and image content. Extensive experimental results demonstrate that our AIDN achieves top performance for invertible downscaling with both arbitrary integer and non-integer scale factors. Code will be released upon publication.

arXiv:2201.13256공식 코드 · Papers with Code

Proximal Denoiser for Convergent Plug-and-Play Optimization with Nonconvex Regularization

Plug-and-Play (PnP) methods solve ill-posed inverse problems through iterative proximal algorithms by replacing a proximal operator by a denoising operation. When applied with deep neural network denoisers, these methods have shown state-of-the-art visual performance for image restoration problems. However, their theoretical convergence analysis is still incomplete. Most of the existing convergence results consider nonexpansive denoisers, which is non-realistic, or limit their analysis to strongly convex data-fidelity terms in the inverse problem to solve. Recently, it was proposed to train the denoiser as a gradient descent step on a functional parameterized by a deep neural network. Using such a denoiser guarantees the convergence of the PnP version of the Half-Quadratic-Splitting (PnP-HQS) iterative algorithm. In this paper, we show that this gradient denoiser can actually correspond to the proximal operator of another scalar function. Given this new result, we exploit the convergence theory of proximal algorithms in the nonconvex setting to obtain convergence results for PnP-PGD (Proximal Gradient Descent) and PnP-ADMM (Alternating Direction Method of Multipliers). When built on top of a smooth gradient denoiser, we show that PnP-PGD and PnP-ADMM are convergent and target stationary points of an explicit functional. These convergence results are confirmed with numerical experiments on deblurring, super-resolution and inpainting.

arXiv:2201.11998

Image Superresolution using Scale-Recurrent Dense Network

Recent advances in the design of convolutional neural network (CNN) have yielded significant improvements in the performance of image super-resolution (SR). The boost in performance can be attributed to the presence of residual or dense connections within the intermediate layers of these networks. The efficient combination of such connections can reduce the number of parameters drastically while maintaining the restoration quality. In this paper, we propose a scale recurrent SR architecture built upon units containing series of dense connections within a residual block (Residual Dense Blocks (RDBs)) that allow extraction of abundant local features from the image. Our scale recurrent design delivers competitive performance for higher scale factors while being parametrically more efficient as compared to current state-of-the-art approaches. To further improve the performance of our network, we employ multiple residual connections in intermediate layers (referred to as Multi-Residual Dense Blocks), which improves gradient propagation in existing layers. Recent works have discovered that conventional loss functions can guide a network to produce results which have high PSNRs but are perceptually inferior. We mitigate this issue by utilizing a Generative Adversarial Network (GAN) based framework and deep feature (VGG) losses to train our network. We experimentally demonstrate that different weighted combinations of the VGG loss and the adversarial loss enable our network outputs to traverse along the perception-distortion curve. The proposed networks perform favorably against existing methods, both perceptually and objectively (PSNR-based) with fewer parameters.

arXiv:2201.11996

Deep Networks for Image and Video Super-Resolution

Efficiency of gradient propagation in intermediate layers of convolutional neural networks is of key importance for super-resolution task. To this end, we propose a deep architecture for single image super-resolution (SISR), which is built using efficient convolutional units we refer to as mixed-dense connection blocks (MDCB). The design of MDCB combines the strengths of both residual and dense connection strategies, while overcoming their limitations. To enable super-resolution for multiple factors, we propose a scale-recurrent framework which reutilizes the filters learnt for lower scale factors recursively for higher factors. This leads to improved performance and promotes parametric efficiency for higher factors. We train two versions of our network to enhance complementary image qualities using different loss configurations. We further employ our network for video super-resolution task, where our network learns to aggregate information from multiple frames and maintain spatio-temporal consistency. The proposed networks lead to qualitative and quantitative improvements over state-of-the-art techniques on image and video super-resolution benchmarks.

arXiv:2201.11793공식 코드 · Papers with Code

Denoising Diffusion Restoration Models

Many interesting tasks in image restoration can be cast as linear inverse problems. A recent family of approaches for solving these problems uses stochastic algorithms that sample from the posterior distribution of natural images given the measurements. However, efficient solutions often require problem-specific supervised training to model the posterior, whereas unsupervised methods that are not problem-specific typically rely on inefficient iterative methods. This work addresses these issues by introducing Denoising Diffusion Restoration Models (DDRM), an efficient, unsupervised posterior sampling method. Motivated by variational inference, DDRM takes advantage of a pre-trained denoising diffusion generative model for solving any linear inverse problem. We demonstrate DDRM's versatility on several image datasets for super-resolution, deblurring, inpainting, and colorization under various amounts of measurement noise. DDRM outperforms the current leading unsupervised methods on the diverse ImageNet dataset in reconstruction quality, perceptual quality, and runtime, being 5x faster than the nearest competitor. DDRM also generalizes well for natural images out of the distribution of the observed ImageNet training set.

arXiv:2201.12288공식 코드 · Papers with Code

VRT: A Video Restoration Transformer

Video restoration (e.g., video super-resolution) aims to restore high-quality frames from low-quality frames. Different from single image restoration, video restoration generally requires to utilize temporal information from multiple adjacent but usually misaligned video frames. Existing deep methods generally tackle with this by exploiting a sliding window strategy or a recurrent architecture, which either is restricted by frame-by-frame restoration or lacks long-range modelling ability. In this paper, we propose a Video Restoration Transformer (VRT) with parallel frame prediction and long-range temporal dependency modelling abilities. More specifically, VRT is composed of multiple scales, each of which consists of two kinds of modules: temporal mutual self attention (TMSA) and parallel warping. TMSA divides the video into small clips, on which mutual attention is applied for joint motion estimation, feature alignment and feature fusion, while self attention is used for feature extraction. To enable cross-clip interactions, the video sequence is shifted for every other layer. Besides, parallel warping is used to further fuse information from neighboring frames by parallel feature warping. Experimental results on five tasks, including video super-resolution, video deblurring, video denoising, video frame interpolation and space-time video super-resolution, demonstrate that VRT outperforms the state-of-the-art methods by large margins ($\textbf{up to 2.16dB}$) on fourteen benchmark datasets.

arXiv:2201.11279공식 코드 · Papers with Code

Revisiting RCAN: Improved Training for Image Super-Resolution

Image super-resolution (SR) is a fast-moving field with novel architectures attracting the spotlight. However, most SR models were optimized with dated training strategies. In this work, we revisit the popular RCAN model and examine the effect of different training options in SR. Surprisingly (or perhaps as expected), we show that RCAN can outperform or match nearly all the CNN-based SR architectures published after RCAN on standard benchmarks with a proper training strategy and minimal architecture change. Besides, although RCAN is a very large SR architecture with more than four hundred convolutional layers, we draw a notable conclusion that underfitting is still the main problem restricting the model capability instead of overfitting. We observe supportive evidence that increasing training iterations clearly improves the model performance while applying regularization techniques generally degrades the predictions. We denote our simply revised RCAN as RCAN-it and recommend practitioners to use it as baselines for future research. Code is publicly available at https://github.com/zudi-lin/rcan-it.

arXiv:2201.10747공식 코드 · Papers with Code

Learning Multiple Probabilistic Degradation Generators for Unsupervised Real World Image Super Resolution

Unsupervised real world super resolution (USR) aims to restore high-resolution (HR) images given low-resolution (LR) inputs, and its difficulty stems from the absence of paired dataset. One of the most common approaches is synthesizing noisy LR images using GANs (i.e., degradation generators) and utilizing a synthetic dataset to train the model in a supervised manner. Although the goal of training the degradation generator is to approximate the distribution of LR images given a HR image, previous works have heavily relied on the unrealistic assumption that the conditional distribution is a delta function and learned the deterministic mapping from the HR image to a LR image. In this paper, we show that we can improve the performance of USR models by relaxing the assumption and propose to train the probabilistic degradation generator. Our probabilistic degradation generator can be viewed as a deep hierarchical latent variable model and is more suitable for modeling the complex conditional distribution. We also reveal the notable connection with the noise injection of StyleGAN. Furthermore, we train multiple degradation generators to improve the mode coverage and apply collaborative learning for ease of training. We outperform several baselines on benchmark datasets in terms of PSNR and SSIM and demonstrate the robustness of our method on unseen data distribution. Code is available at https://github.com/sangyun884/MSSR.

arXiv:2201.10521

A Review of Deep Learning Based Image Super-resolution Techniques

Image super-resolution technology is the process of obtaining high-resolution images from one or more low-resolution images. With the development of deep learning, image super-resolution technology based on deep learning method is emerging. This paper reviews the research progress of the application of depth learning method in the field of image super-resolution, introduces this kind of super-resolution work from several aspects, and looks forward to the further application of depth learning method in the field of image super-resolution. By collecting and counting the relevant literature on the application of depth learning in the field of image super-resolution, we preliminarily summarizes the application results of depth learning method in the field of image super-resolution, and reports the latest progress of image super-resolution technology based on depth learning method.

arXiv:2201.10084

Revisiting L1 Loss in Super-Resolution: A Probabilistic View and Beyond

Super-resolution as an ill-posed problem has many high-resolution candidates for a low-resolution input. However, the popular $\ell_1$ loss used to best fit the given HR image fails to consider this fundamental property of non-uniqueness in image restoration. In this work, we fix the missing piece in $\ell_1$ loss by formulating super-resolution with neural networks as a probabilistic model. It shows that $\ell_1$ loss is equivalent to a degraded likelihood function that removes the randomness from the learning process. By introducing a data-adaptive random variable, we present a new objective function that aims at minimizing the expectation of the reconstruction error over all plausible solutions. The experimental results show consistent improvements on mainstream architectures, with no extra parameter or computing cost at inference time.