MAMNet: Multi-path adaptive modulation network for image super-resolution

Kim, Junhyuk; Choi, Jun‐Ho; Cheon, Manri; Lee, Jong-Seok

doi:10.1016/j.neucom.2020.03.069

Cited by 30 publications

(21 citation statements)

References 3 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The spatial attention mechanism transforms the spatial information in the original image into another space by learning the high-level feature information between pixels and retaining the key information. The super-resolution residual attention module (SRRAM) [34] fuses channel attention and spatial attention in one module and uses depth-wise convolution to conduct the feature extraction of spatial attention. At the same time, the SE module is used to adjust the attention of the input feature channel.…”

Section: B Attention Mechanismmentioning

confidence: 99%

“…As shown in Figure 5, we refer to the attention mechanism in SRRAM [34] to integrate the spatial attention and channel attention into one module. Channel attention adopts the squeeze-extraction (SE) method.…”

Section: E Spatial and Channal Attention Modulementioning

confidence: 99%

“…We will use the network with different magnification training to test the corresponding magnification and compare the test results with the advanced super-resolution model developed in recent years. Table 1, 2, 3 present the quantitative comparison results [22], DRCN [43], DRRN [44], LapSRN [45], IDN [26], MemNet [46], SRMDNF [47], IMDN [27], SRRAM [34], and CARN [48]. PSNR and SSIM using different super-resolution models at different up-scale factor of ×2, ×3 and ×4.…”

Section: Performance Comparisonmentioning

confidence: 99%

“…To build a lightweight SR model, it is necessary to have fewer network parameters while also ensuring the image reconstruction effect. From Figure 6, we can observe that compared with other, most advanced methods, such as IDN [26], CARN [48], SRRAM [34] and IMDN [27], the proposed FGRDN has fewer parameters and better image reconstruction performance, achieving a tradeoff between performance and model size.…”

Section: Number Of Parameter Comparisonmentioning

confidence: 99%

See 3 more Smart Citations

Lightweight Feedback Convolution Neural Network for Remote Sensing Images Super-Resolution

Jin

Wang

et al. 2021

IEEE Access

View full text Add to dashboard Cite

There are lots of image data in the field of remote sensing, most of which have low-resolution due to the limited image sensor. The super-resolution method can effectively restore the low-resolution image to the high-resolution image. However, the existing super-resolution method has both heavy computing burden and number of parameters. For saving costs, we propose the feedback ghost residual dense network (FGRDN), which considers the feedback mechanism as the framework to attain lower features through high-level refining. Further, for feature extraction, we replace the convolution of the residual dense blocks (RDBs) with ghost modules (GMs), which can remove the redundant channels and avoid the increase of parameters along with the network depth. Finally, the spatial and channel attention module (SCM) is employed in the end of the RDB to learn more useful information from features. Compared to other SOTA lightweight algorithms, our proposed algorithm can reach convergences more rapidly with fewer parameters, and the performance of the network can be markedly enhanced on the image texture and object contour reconstruction with better peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

show abstract

Section: B Attention Mechanismmentioning

confidence: 99%

Section: E Spatial and Channal Attention Modulementioning

confidence: 99%

Section: Performance Comparisonmentioning

confidence: 99%

Section: Number Of Parameter Comparisonmentioning

confidence: 99%

See 2 more Smart Citations

Lightweight Feedback Convolution Neural Network for Remote Sensing Images Super-Resolution

Jin

Wang

et al. 2021

IEEE Access

View full text Add to dashboard Cite

show abstract

“…Fang et al [27] proposed a multi-scale adaptive sparse representation (MASR) model, the multi-scale spatial information is effectively utilized to limit the pixels from different scales to achieve better image representation. Kim et al [28] proposed a characteristic response modulation network based on adaptive convolution, it solves the inherent inadaptability problem of CNN-based network. Since the input image is limited by the pixel size, the performance of the network in large-size image input is affected.…”

Section: Related Workmentioning

confidence: 99%

Urban Remote Sensing Scene Recognition Based on Lightweight Convolution Neural Network

2021

View full text Add to dashboard Cite

The use rate of urban land is a significant sign to evaluate urban construction, and scene recognition has important application value in improving urban land use rate. In this paper, a new lightweight model based on VGG16 is proposed to extract distinct features of remote sensing images through five convolution modules. This model uses depthwise separable convolution to reduce the network parameters. An adaptive pooling layer is added to solve the inherent non-adaptive problem of the convolution network. It makes the network insensitive to the size of the input image. The global average pooling layer is used to sum the information to make the input spatial transformation more stable. This paper conducts training and testing on two data sets, NWPU-RESISC45 Dataset and SIRI-WHU Dataset, and the recognition scenarios are 13 and 12 categories. Experimental results show that this method is better than other models in recognition accuracy and model size.

show abstract

Multi Attention Neural Network for Digital Rock CT Images Super-Resolution

Xing

Yao

Liu

et al. 2022

Preprint

View full text Add to dashboard Cite

High-quality digital rock images are essential for subsequent high-precision numerical simulations. But limited by the imaging capability of computed tomography (CT), high resolution digital rock images with wide imaging field of view (FOV) cannot be acquired simultaneously. To cope with this constraint, we propose a novel Multi Attention Super-Resolution Neural Network (MASR) that enhances the resolution of images with wide FOV. Considering that textures and edges are more crucial in digital rocks, MASR introduces the component attention mechanism of Component Divide-and-Conquer Super-Resolution (CDCSR) model. By redesigning the hourglass network with spatial and channel attention mechanisms, proposing a spatial attention-based mask module, and optimizing the component attention mask calculation process, MASR delivers higher information utilization with fewer parameters and faster training than CDCSR. And we optimize the depth of MASR to trade off speed and superresolution quality. Furthermore, we retrained several state-of-the-art models. Through quantitative evaluations and qualitative visualizations, it is verified that MASR can recover sharper edges while removing noise, and obtain digital rock images with superior quality and reliability. The pixelwise relative errors of MASR reconstructions are reduced by 15% to 26% over bicubic interpolation method. Our codes are publicly available at https://github.com/MHDXing/MASR-for-Digital-Rock-Images.

show abstract

MAMNet: Multi-path adaptive modulation network for image super-resolution

Cited by 30 publications

References 3 publications

Lightweight Feedback Convolution Neural Network for Remote Sensing Images Super-Resolution

Lightweight Feedback Convolution Neural Network for Remote Sensing Images Super-Resolution

Urban Remote Sensing Scene Recognition Based on Lightweight Convolution Neural Network

Multi Attention Neural Network for Digital Rock CT Images Super-Resolution

Contact Info

Product

Resources

About