.png){kind=icon}
.png){kind=icon}
Return
Spatial and channel enhanced self-attention network for efficient single image super-resolution
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2024.129258.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, the widespread adoption of the swin transformer model in computer vision has significantly enhanced the quality of super-resolution reconstructions. However, existing single-image super-resolution models, while effective in capturing spatial details, often overlook interactions across channels, limiting their ability to fully leverage the depth of feature information. This deficiency hampers their performance, leading to suboptimal reconstruction quality, as the models cannot adequately capture global dependencies across both spatial and channel dimensions. Consequently, these models tend to produce less accurate high- frequency details, which are critical for enhancing image clarity and preserving fine structures. To address this issue, we introduce the Spatial and Channel Enhanced Self-Attention Network (SCESN), designed to integrate both spatial and channel attention mechanisms, thereby enhancing feature utilization and delivering superior reconstruction performance. In this work, a global self-attention module (GSM) was designed to perform deep feature extraction by utilizing global contextual information in the images. Specifically, the GSM incorporates spatial self-attention block (SSB), channel self-attention block (CSB), and spatial and channel enhanced blocks (SCEB). SCEB boosts the interaction between spatial information from SSBs and channel information from CSBs, thereby maximizing feature utilization and enhancing network performance. Moreover, efficient convolution blocks are integrated to efficiently enhance the model's deep spatial feature extraction capabilities. Quantitative evaluations across multiple test sets, along with visual comparisons, demonstrate that the proposed method achieves superior reconstruction results with fewer parameters compared to existing methodologies.
Keywords:
Image super-resolution
Lightweight network
Self-attention mechanism
Transformer
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
No organization information available