Transformer-based Super-Resolution (SR) methods have demonstrated superior performance compared to convolutional neural network (CNN)-based SR approaches due to their capability to capture long-range dependencies. However, their high computational complexity necessitates the development of lightweight approaches for practical use. To address this challenge, we propose the Attention-Sharing Information Distillation (ASID) network, a lightweight SR network that integrates attention-sharing and an information distillation structure specifically designed for Transformer-based SR methods. We modify the information distillation scheme, originally designed for efficient CNN operations, to reduce the computational load of stacked self-attention layers, effectively addressing the efficiency bottleneck. Additionally, we introduce attention-sharing across blocks to further minimize the computational cost of self-attention operations. By combining these strategies, ASID achieves competitive performance with existing SR methods while requiring only around 300K parameters – significantly fewer than existing CNN-based and Transformer-based SR models. Furthermore, ASID outperforms state-of-the-art SR methods when the number of parameters is matched, demonstrating its efficiency and effectiveness.

Training deep neural networks has become increasingly demanding, requiring large datasets and significant computational resources, especially as model complexity advances. Data distillation methods, which aim to improve data efficiency, have emerged as promising solutions to this challenge. In the field of single image super-resolution (SISR), the reliance on large training datasets highlights the importance of these techniques. Recently, a generative adversarial network (GAN) inversion-based data distillation framework for SR was proposed, showing potential for better data utilization. However, the current method depends heavily on pre-trained SR networks and class-specific information, limiting its generalizability and applicability. To address these issues, we introduce a new data distillation approach for image SR that does not need class labels or pre-trained SR models. In particular, we first extract high- gradient patches and categorize images based on CLIP features, then fine-tune a diffusion model on the selected patches to learn their distribution and synthesize distilled training images. Experimental results show that our method achieves state-of-the-art performance while using significantly less training data and requiring less computational time. Specifically, when we train a baseline Transformer model for SR with only 0.68% of the original dataset, the performance drop is just 0.3 dB. In this case, diffusion model fine-tuning takes 4 hours, and SR model training completes within 1 hour, much shorter than the 11-hour training time with the full dataset.

Deep learning methods have shown outstanding performance in many applications, including single-image super-resolution (SISR). With residual connection architecture, deeply stacked convolutional neural networks provide a substantial performance boost for SISR, but their huge parameters and computational loads are impractical for real-world applications. Thus, designing lightweight models with acceptable performance is one of the major tasks in current SISR research. The objective of lightweight network design is to balance a computational load and reconstruction performance. Most of the previous methods have manually designed complex and predefined fixed structures, which generally required a large number of experiments and lacked flexibility in the diversity of input image statistics. In this paper, we propose a dynamic residual self-attention network (DRSAN) for lightweight SISR, while focusing on the automated design of residual connections between building blocks. The proposed DRSAN has dynamic residual connections based on dynamic residual attention (DRA), which adaptively changes its structure according to input statistics. Specifically, we propose a dynamic residual module that explicitly models the DRA by finding the interrelation between residual paths and input image statistics, as well as assigning proper weights to each residual path. We also propose a residual self-attention (RSA) module to further boost the performance, which produces 3-dimensional attention maps without additional parameters by cooperating with residual structures. The proposed dynamic scheme, exploiting the combination of DRA and RSA, shows an efficient trade-off between computational complexity and network performance. Experimental results show that the DRSAN performs better than or comparable to existing state-of-the-art lightweight models for SISR.

This paper presents a new variational inference framework for image restoration and a convolutional neural network (CNN) structure that can solve the restoration problems described by the proposed framework. Earlier CNN-based image restoration methods primarily focused on network architecture design or training strategy with non-blind scenarios where the degradation models are known or assumed. For a step closer to real-world applications, CNNs are also blindly trained with the whole dataset, including diverse degradations. However, the conditional distribution of a high-quality image given a diversely degraded one is too complicated to be learned by a single CNN. Therefore, there have also been some methods that provide additional prior information to train a CNN. Unlike previous approaches, we focus more on the objective of restoration based on the Bayesian perspective and how to reformulate the objective. Specifically, our method relaxes the original posterior inference problem to better manageable sub-problems and thus behaves like a divide-and-conquer scheme. As a result, the proposed framework boosts the performance of several restoration problems compared to the previous ones. Specifically, our method delivers state-of-the-art performance on Gaussian denoising, real-world noise reduction, blind image super-resolution, and JPEG compression artifacts reduction. Our code and more details are available on our project page, https://github.com/JWSoh/VDIR.

This article presents a new method for understanding and visualizing convolutional neural networks (CNNs). Most existing approaches to this problem focus on a global score and evaluate the pixelwise contribution of inputs to the score. The analysis of CNNs for multilabeled outputs or regression has not yet been considered in the literature, despite their success on image classification tasks with well-defined global scores. To address this problem, we propose a new inverse-based approach that computes the inverse of a feedforward pass to identify activations of interest in lower layers. We developed a layerwise inverse procedure based on two observations: 1) inverse results should have consistent internal activations to the original forward pass and 2) a small amount of activation in inverse results is desirable for human interpretability. Experimental results show that the proposed method allows us to analyze CNNs for classification and regression in the same framework. We demonstrated that our method successfully finds attributions in the inputs for image classification with comparable performance to state-of-the-art methods. To visualize the tradeoff between various methods, we developed a novel plot that shows the tradeoff between the amount of activations and the rate of class reidentification. In the case of regression, our method showed that conventional CNNs for single image super-resolution overlook a portion of frequency bands that may result in performance degradation.

Real-noise denoising is a challenging task because the statistics of real-noise do not follow the normal distribution, and they are also spatially and temporally changing. In order to cope with various and complex real-noise, we propose a well-generalized denoising architecture and a transfer learning scheme. Specifically, we adopt an adaptive instance normalization to build a denoiser, which can regularize the feature map and prevent the network from overfitting to the training set. We also introduce a transfer learning scheme that transfers knowledge learned from synthetic-noise data to the real-noise denoiser. From the proposed transfer learning, the synthetic-noise denoiser can learn general features from various synthetic-noise data, and the real-noise denoiser can learn the real-noise characteristics from real data. From the experiments, we find that the proposed denoising method has great generalization ability, such that our network trained with synthetic-noise achieves the best performance for Darmstadt Noise Dataset (DND) among the methods from published papers. We can also see that the proposed transfer learning scheme robustly works for real-noise images through the learning with a very small number of labeled data.

Convolutional neural networks (CNNs) have shown dramatic improvements in single image super-resolution (SISR) by using large-scale external samples. Despite their remarkable performance based on the external dataset, they cannot exploit internal information within a specific image. Another problem is that they are applicable only to the specific condition of data that they are supervised. For instance, the low-resolution (LR) image should be a "bicubic" downsampled noise-free image from a high-resolution (HR) one. To address both issues, zero-shot super-resolution (ZSSR) has been proposed for flexible internal learning. However, they require thousands of gradient updates, i.e., long inference time. In this paper, we present Meta-Transfer Learning for Zero-Shot Super-Resolution (MZSR), which leverages ZSSR. Precisely, it is based on finding a generic initial parameter that is suitable for internal learning. Thus, we can exploit both external and internal information, where one single gradient update can yield quite considerable results. With our method, the network can quickly adapt to a given image condition. In this respect, our method can be applied to a large spectrum of image conditions within a fast adaptation process.

Deep convolutional neural networks have shown significant improvement in the single image super-resolution (SISR) field. Recently, there have been attempts to solve the SISR problem using lightweight networks, considering limited computational resources for real-world applications. Especially for lightweight networks, balancing between parameter demand and performance is very difficult to adjust, and most lightweight SISR networks are manually designed based on a huge number of brute-force experiments. Besides, a critical key to the network performance relies on the skip connection of building blocks that are repeatedly in the architecture. Notably, in previous works, these connections are pre-defined and manually determined by human researchers. Hence, they are less flexible to the input image statistics, and there can be a better solution for the given number of parameters. Therefore, we focus on the automated design of networks regarding the connection of basic building blocks (residual networks), and as a result, propose a dynamic residual attention network (DRAN). The proposed method allows the network to dynamically select residual paths depending on the input image, based on the idea of attention mechanism. For this, we design a dynamic residual module that determines the residual paths between the basic building blocks for the given input image. By finding optimal residual paths between the blocks, the network can selectively bypass informative features needed to reconstruct the target high-resolution (HR) image. Experimental results show that our proposed DRAN outperforms most of the existing state-of-the-arts lightweight models in SISR.

Image denoising is an essential part of many image processing and computer vision tasks due to inevitable noise corruption during image acquisition. Traditionally, many researchers have investigated image priors for the denoising, within the Bayesian perspective based on image properties and statistics. Recently, deep convolutional neural networks (CNNs) have shown great success in image denoising by incorporating large-scale synthetic datasets. However, they both have pros and cons. While the deep CNNs are powerful for removing the noise with known statistics, they tend to lack flexibility and practicality for the blind and real-world noise. Moreover, they cannot easily employ explicit priors. On the other hand, traditional non-learning methods can involve explicit image priors, but they require considerable computation time and cannot exploit large-scale external datasets. In this paper, we present a CNN-based method that leverages the advantages of both methods based on the Bayesian perspective. Concretely, we divide the blind image denoising problem into sub-problems and conquer each inference problem separately. As the CNN is a powerful tool for inference, our method is rooted in CNNs and propose a novel design of network for efficient inference. With our proposed method, we can successfully remove blind and real-world noise, with a moderate number of parameters of universal CNN.

This paper presents methods based on convolutional neural networks (CNNs) for removing compression artifacts. We modify the Inception module for the image restoration problem and use it as a building block for constructing blind and non-blind artifact removal networks. It is known that a CNN trained in a non-blind scenario (known compression quality factor) performs better than the one trained in a blind scenario (unknown factor), and our network is not an exception. However, the blind system is more practical because the compression quality factor is not always available or does not reflect the actual quality when the image is a transcoded or requantized image. Hence, in this paper, we also propose a pseudo-blind system that estimates the quality factor for a given compressed image and then applies a network that is trained with a similar quality factor. For this purpose, we propose a CNN that estimates the compression quality factor and prepare several non-blind artifact removal networks that are trained for some specific compression quality factors. We train the networks and conduct experiments on widely used compression standards, such as JPEG, MPEG-2, H.264, and HEVC. In addition, we conduct experiments for dynamically changing and transcoded videos to demonstrate the effectiveness of the quality estimation method. The experimental results show that the proposed pseudo-blind network performs better than the blind one for the various cases stated above and requires fewer computations.

Recently, many convolutional neural networks for single image super-resolution (SISR) have been proposed, which focus on reconstructing the high-resolution images in terms of objective distortion measures. However, the networks trained with objective loss functions generally fail to reconstruct the realistic fine textures and details that are essential for better perceptual quality. Recovering the realistic details remains a challenging problem, and only a few works have been proposed which aim at increasing the perceptual quality by generating enhanced textures. However, the generated fake details often make undesirable artifacts and the overall image looks somewhat unnatural. Therefore, in this paper, we present a new approach to reconstructing realistic super-resolved images with high perceptual quality, while maintaining the naturalness of the result. In particular, we focus on the domain prior properties of SISR problem. Specifically, we define the naturalness prior in the low-level domain and constrain the output image in the natural manifold, which eventually generates more natural and realistic images. Our results show better naturalness compared to the recent super-resolution algorithms including perception-oriented ones.

This paper presents a new framework that controls feature maps of a convolutional neural network (CNN) according to the noise level such that the network can have different properties to different levels. Unlike the conventional non-blind approach which reloads all the parameters of CNN or switches to other CNNs for different noise levels, we adjust the CNN activation feature maps without changing the parameters at the test phase. For this, we additionally construct a noise level indicator network, which gives appropriate weighting values to the feature maps for the given situation. The noise level indicator network is so simple that it can be implemented as a low-dimensional look-up table at the test phase and thus does not increase the overall complexity. From the experiments on noise reduction, we can observe that the proposed method achieves better performance compared to the baseline network.

In this paper, we propose a trainable Gabor wavelet (TGW) layer and cascade it with a convolutional neural network (CNN) for the age estimation. Unlike an existing method that uses fixed (hand-tuned) Gabor filters at the head of a CNN, we use Gabor wavelets that can be adapted for the given input as well as for the targeting task. This is enabled by (a) estimating hyperparameters of Gabor wavelets from the input and (b) using a 1 × 1 convolution layer for the selection of orientation parameter. The proposed TGW layers are trained with the standard gradient-descent method and can be easily incorporated with conventional CNNs in an end-to-end training manner. We conduct experiments on the Adience dataset and show that the proposed network outperforms the baseline CNN without TGW layers and efficiently used trainable parameters than ordinary CNN based methods.

This paper presents a hierarchical convolutional neural network (CNN) for single image super-resolution (SISR), which exploits the controlled multi-context features. We focus on the method to extract more enriched features than the case of using fixed size kernels. For this, we attempt to bring out the best of the given parameter capacity through the design of some sophisticated networks in a hierarchical manner. First, we exploit the multi-kernel dilated convolution for extracting multi-size contexts from the image and combine them with the proposed trainable parameters. The multi-kernel network with some new pre-and post-processing blocks forms our basic building block. Then the basic building blocks are densely connected with a new feature fusion schemes, which makes the upper level building block. Then, by connecting the upper level blocks, we can use various features which can enrich the representation of the images. In the experiments, it is shown that the proposed method achieves significant PSNR gain compared to recent lightweight models with comparable numbers of parameters.