CVANet: Cascaded visual attention network for single image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (DCNNs) excel at feature extraction for single image super-resolution (SISR).
• Existing DCNN methods often fail to fully leverage complementary information across feature maps, channels, and pixels.
• This limitation hinders the comprehensive feature representation capabilities of DCNNs for SISR.

Purpose of the Study:

• To introduce a novel network, CVANet (Cascaded Visual Attention Network), for SISR.
• To simulate human visual attention mechanisms for enhanced detail reconstruction in SISR.
• To improve the feature representation and image reconstruction quality in SISR tasks.

Main Methods:

• Developed a Cascaded Visual Attention Network (CVANet) incorporating feature, channel, and pixel attention modules.
• Implemented a trainable Feature Attention Module (FAM) for feature-level attention learning.
• Introduced a Channel Attention Module (CAM) for channel-level attention and a Pixel Attention Module (PAM) for adaptive feature selection.

Main Results:

• CVANet effectively enhances image resolution by exploiting diverse feature representations and visual perception principles.
• Experiments on four benchmarks demonstrate CVANet's superiority over state-of-the-art (SOTA) methods.
• Performance gains were observed in subjective visual quality, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM).

Conclusions:

• CVANet significantly improves single image super-resolution performance.
• The proposed attention modules effectively capture and utilize complementary information for better feature representation.
• The network demonstrates strong potential for advancing SISR technology.