Image super-resolution, aims to recover a corresponding high-resolution image from a given low-resolution image. While most state-of-the-art methods only consider using fixed small-size convolution kernels (e.g., 1 × 1, 3 × 3) to extract image features, few works have been made to large-size convolution kernels for image super-resolution (SR). In this paper, we propose a novel lightweight baseline model LKASR based on large kernel attention (LKA). LKASR consists of three parts, shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module consists of multiple cascaded visual attention modules (VAM), each of which consists of a 1 × 1 convolution, a large kernel attention (acts as Transformer) and a feature refinement module (FRM, acts as CNN). Specifically, VAM applies lightweight architecture like swin transformer to realize iterative extraction of global and local features of images, which greatly improves the effectiveness of SR method (0.049s in Urban100 dataset). For different scales ( × 2, × 3, × 4), extensive experimental results on benchmark demonstrate that LKASR outperforms most lightweight SR methods by up to 0 . 17 ∼ 0 . 34 dB, while the total of parameters and FLOPs remains lightweight. 

LKASR: Large kernel attention for lightweight image super-resolution

Deep learning-based image super-resolution (SR) methods have attracted growing interest due to their outstanding performance. However, most of these methods assume that the degradations of low-resolution (LR) observations are fixed and known (e.g., bicubic downsampling). This is however not always true in real scenarios. Moreover, the mismatch between the intrinsic degradation of LR images and the assumed degradation generally causes artifacts. Therefore, it is essential to move beyond idealized assumptions and make the SR model adapt to variant degradations. To achieve this goal, we propose a Blind image Super-Resolution approach based on contrastive learning-based Implicit Degradation Modeling (IDMBSR). Since it is challenging to explicitly estimate degradation parameters, a representation is learned from each LR image to model its degradation and differentiate between variant degradations, thereby guiding the subsequent reconstruction to achieve image-specific SR. Since there are no representation labels for LR images, contrastive learning is used to train the attention-enhanced encoder for degradation encoding with the help of degradation parameters for more effective model training. Moreover, to make the reconstruction process adapt to LR images with variant degradations, a degradation-guided SR network is developed, in which the degradation representation adaptively influences the SR process from beginning to end by applying data-driven transformation to middle features. Benefitting from the combination of the degradation representation encoder and the degradation-guided SR network, IDMBSR can adapt to various LR observations without any prior knowledge about their degradations. Experimental results show that IDMBSR outperforms several state-of-the-art blind SR methods with fewer parameters and higher efficiency. 

Weakly-supervised contrastive learning-based implicit degradation modeling for blind image super-resolution

Light Field (LF) imaging enables many attractive applications since scene angular and spatial information can be captured simultaneously. However, the limited angular resolution hinders its development. To alleviate this problem, in this paper, we put forward a geometry-assisted multi-representation view reconstruction network for LF angular super-resolution. There are two stages in our method, namely multi-representation view reconstruction and geometry-assisted refinement. The former fully utilizes the LF lenslet image, sub-aperture image array, and pseudo video sequence representations to synthesize a high-angular-resolution dense LF image stack by adequately exploring the structural characteristics of such three representations. The latter aims to exploit the intra-LF spatial–angular information and inter-LF geometry information contained in the derived dense LF image stack to further promote the reconstruction performance. Specifically, a bidirectional view stack structure is constituted first, which is then fed into the proposed geometry-aware refinement network to allow adequate interactions between the intra-LF spatial–angular information and inter-LF geometry information. Comprehensive angular reconstruction experiments on both real-world and synthetic LF scenes validate the effectiveness of the proposed method. Furthermore, extensible applications on scene depth estimation also demonstrate that the proposed method can recover more texture details. 

Geometry-assisted multi-representation view reconstruction network for Light Field image angular super-resolution

Hierarchical full-attention neural architecture search based on search space compression

Under the global epidemic situation, offline face-to-face teaching has not adapted to the current teaching environment. Using computer technology to carry out online assisted teaching has become an inevitable choice for music teaching in colleges and universities. Offline teaching also has many disadvantages, such as single and backward teaching methods, poor sharing of audio-visual resources, and inconsistent teaching contents. The survey found that music teachers urgently need an online teaching platform to enable teachers to obtain effective, standardized, and comprehensive teaching resources anytime and anywhere so that students can enjoy high-quality music teaching resources as much as possible. Based on these requirements, the development of online music teaching system under B/S architecture has an important practical significance.

/pdf/design-of-online-music-teaching-system-based-on-b-s-1h7zw564fq.pdf

Design of Online Music Teaching System Based on B/S Architecture

We propose a plug-and-play neural architecture search (NAS) method to explore diverse architectures for single image super-resolution (SISR). Unlike current NAS-based methods with the single path setting and pipeline setting, our proposed method achieves the trade-off between diverse network architectures and search cost. Our proposed method formulates the task in a differentiable manner, which inherits the architecture parameter optimization method from Discrete Stochastic Neural Architecture Search (DSNAS). Besides the straightforward searching of operations, we also search each node in a cell for the activation function, from-node, and skip-connection node, which diverse the searched architecture topologies. The individually searching of skip-connection node avoids skip-connection excessive phenomenon. Moreover, to alleviate the influence of inconsistent architecture between training and testing periods, we introduce random variables into the architecture parameter as regularization. Benchmark experiments show our state-of-the-art performance under specific parameters and FLOPs constraints. Compared with other NAS-based SISR methods, our proposed methods achieve better performance with less searching time and resources. The superior results further demonstrate the effectiveness of our proposed NAS methods.

XinZhi Liu

Papers

Exploring more diverse network architectures for single image super-resolution