Accurate Single Image Super-resolution Research Articles

A Multi-Attention Feature Distillation Neural Network for Lightweight Single Image Super-Resolution

In recent years, remarkable performance improvements have been produced by deep convolutional neural networks (CNN) for single image super-resolution (SISR). Nevertheless, a high proportion of CNN-based SISR models are with quite a few network parameters and high computational complexity for deep or wide architectures. How to more fully utilize deep features to make a balance between model complexity and reconstruction performance is one of the main challenges in this field. To address this problem, on the basis of the well-known information multi-distillation model, a multi-attention feature distillation network termed as MAFDN is developed for lightweight and accurate SISR. Specifically, an effective multi-attention feature distillation block (MAFDB) is designed and used as the basic feature extraction unit in MAFDN. With the help of multi-attention layers including pixel attention, spatial attention, and channel attention, MAFDB uses multiple information distillation branches to learn more discriminative and representative features. Furthermore, MAFDB introduces the depthwise over-parameterized convolutional layer (DO-Conv)-based residual block (OPCRB) to enhance its ability without incurring any parameter and computation increase in the inference stage. The results on commonly used datasets demonstrate that our MAFDN outperforms existing representative lightweight SISR models when taking both reconstruction performance and model complexity into consideration. For example, for ×4 SR on Set5, MAFDN (597K/33.79G) obtains 0.21 dB/0.0037 and 0.10 dB/0.0015 PSNR/SSIM gains over the attention-based SR model AFAN (692K/50.90G) and the feature distillation-based SR model DDistill-SR (675K/32.83G), respectively.

Multi-scale receptive field fusion network for lightweight image super-resolution

In recent years, with the rapid development of deep learning, quantities of convolutional neural network (CNN) based methods are applied to the field of single image super-resolution (SISR), which have excellent performance compared with traditional ones. However, these methods have a large amount of calculation and memory consumption, limiting their application in edge devices. To address this problem, we propose a more lightweight multi-scale receptive field fusion network (MRFN) for fast and accurate SISR. Specifically, we first propose a multi-scale receptive field fusion block to fuse different scales of spatial information and enlarge the overall receptive field simultaneously. Secondly, based on the non-local sparse attention, we introduce a hash-learnable cheap non-local attention to capture long-range dependencies with much fewer parameters and calculations. Moreover, channel attention is also very important for SISR, we further combine directional second-order information and spatial coordinate information into it to enhance its capability, which is called second-order coordinate attention. The cooperation of these components leads to the success of our MRFN. Extensive experimental results show that our method achieves a better trade-off against other state-of-the-art methods in terms of SISR performance and model complexity.

Accurate single image super-resolution using multi-path wide-activated residual network

In many recent image super-resolution (SR) methods based on convolutional neural networks (CNNs), the superior performance was achieved by training very large networks, which may not be suitable for real-world applications with limited computing resources. Therefore, it is necessary to develop more compact networks that achieve a better trade-off between the model size and the performance. In this paper, we propose an efficient and effective network called multi-path wide-activated residual network (MWRN). Firstly, as the basic building block of MWRN, the multi-path wide-activated residual block (MWRB) is presented to extract the multi-scale features. MWRB consists of three parallel wide-activated residual paths, where the dilated convolutions with different dilation factors are used to increase the receptive fields. Secondly, the fusional channel attention (FCA) module, which contains a bottleneck layer and a multi-path wide-activated residual channel attention (MWRCA) block, is designed to well exploit the multi-level features in MWRN. In each FCA, the MWRCA block refines the fused features by taking the interdependencies among feature channels into consideration. The experiments demonstrate that, compared with the state-of-the-art methods, the proposed MWRN model is able to provide very competitive performance with a relatively small number of parameters.

Deep fractal residual network for fast and accurate single image super resolution

Recently, deep convolutional neural networks (CNNs) have achieved excellent performance improvement in Single Image Super-Resolution (SISR) tasks. However, it is still difficult for most of the existing CNN-based models to learn and make full use of adequate hierarchical features, which is crucial to ideally recover both high-frequency and low-frequency information. Generally, many networks tend to incorporate numerous parameters aiming to improve their learning capability, which demands more storage and is less applicable to mobile systems. To address this issue, we propose a novel method for SISR by using a deep fractal residual network (DFRN) composed of a series of fractal blocks containing multiple paths for feature extraction. The fractal blocks learn and combine different hierarchical features to generate finer features favoring the reconstruction of high resolution (HR) images. Moreover, we integrate both local and global residual learning to preserve the low-level features and decrease the difficulty of training. Finally, we propose a weight-sharing version with fewer parameters to reduce the space complexity while keeping comparable performance. Experimental results demonstrate that DFRN has the capability of learning different hierarchical features and outperforms other state-of-the-art methods for SISR by reconstructing low resolution (LR) images into the HR images quickly and accurately.

SCRSR: An efficient recursive convolutional neural network for fast and accurate image super-resolution

Convolutional neural networks have recently demonstrated high-quality reconstruction for single image super-resolution (SR). These CNN networks effectively recover a high-resolution (HR) image from a low-resolution (LR) image, at the cost of enormous parameters and heavy computational burden. In this work, we propose a recursive efficient deep convolutional network for fast and accurate single-image SR with only 0.28M parameters. A Split-Concatenate-Residual (SCR) block is proposed to reduce computation and parameters. With downsampling block and upsampling block, we significantly reduce computational complexity and enlarge the size of the receptive field. Specifically, two-level recursive learning is proposed which can improve accuracy by increasing depth without adding any weight parameters. We also employ local, semi-global and global residual techniques to train our very deep network steadily and improve its performance. Extensive experiments indicate that our proposed method Split-Concatenate-Residual Super Resolution (SCRSR) yields promising SR performance while maintaining shorter running time and fewer parameters.

MRFN: Multi-Receptive-Field Network for Fast and Accurate Single Image Super-Resolution

Recently, convolutional neural network (CNN) based models have shown great potential in the task of single image super-resolution (SISR). However, many state-of-the-art SISR solutions are reproducing some tricks proven effective in other vision tasks, such as pursuing a deeper model. In this paper, we propose a new solution (named as Multi-Receptive-Field Network - MRFN), which outperforms existing SISR solutions in three different aspects. First, from receptive field: a novel multi-receptive-field (MRF) module is proposed to extract and fuse features in different receptive fields from local to global. Integrating these hierarchical features can generate better mappings on recovering high-fidelity details at different scales. Second, from network architectures: both dense skip connections and deep supervision are utilized to combine features from the current MRF module and preceding ones for training more representative features. Moreover, a deconvolution layer is embedded at the end of the network to avoid artificial priors induced by numerical data pre-processing (e.g., bicubic stretching), and speed up the restoration process. Finally, from error modeling: different from $L1$ and $L2$ loss functions, we proposed a novel two-parameter training loss called Weighted Huber loss function which can adaptively adjust the value of back-propagated derivative according to the residual value, thus fit the reconstruction error more effectively. Extensive qualitative and quantitative evaluation results on benchmark datasets demonstrate that our proposed MRFN can achieve more accurate recovering results than most state-of-the-art methods with significantly less complexity.

Fast and accurate single image super-resolution via an energy-aware improved deep residual network

Recently, convolutional neural network (CNN) based single image super-resolution (SISR) solutions have demonstrated significant progress on restoring accurate high-resolution image based on its corresponding low-resolution version. However, most state-of-the-art SISR approaches attempt to achieve higher accuracy by pursuing deeper or more complicated models, which adversely increases computational cost. To achieve a good balance between restoration accuracy and computational speed, we make simple but effective modifications to the structure of residual blocks and skip-connections between stacked layers, and then propose a novel energy-aware training loss to adaptively adjust the restoration of high-frequency and low-frequency image regions. Extensive qualitative and quantitative evaluation results on benchmark datasets verify the effectiveness of the proposed techniques that they significantly improve SISR accuracy while causing no/ignorable extra computational loads.