Improved dual-scale residual network for image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Convolutional neural networks (CNNs) have advanced single image super-resolution (SISR).
• Existing CNNs offer improvements in accuracy and speed for SISR tasks.
• Further optimization is needed for efficient and high-performance image reconstruction.

Purpose of the Study:

• To propose an improved dual-scale residual network (IDSRN) for SISR.
• To enhance reconstruction performance while maintaining computational efficiency.
• To effectively capture both dual-scale features and high-frequency image details.

Main Methods:

• Developed an IDSRN with two parallel branches: dual-scale feature extraction and texture attention.
• Utilized improved dual-scale residual blocks (IDSRB) with active weighted mapping for dual-scale feature capture.
• Employed an encoder-decoder network with a symmetric full convolutional-deconvolution structure for high-frequency detail enhancement.

Main Results:

• The IDSRN effectively integrates dual-scale features with high-frequency information.
• Experimental results demonstrate competitive accuracy and efficiency compared to state-of-the-art methods.
• The proposed network shows promising reconstruction performance in SISR tasks.

Conclusions:

• The IDSRN offers a balanced approach to accuracy and computational cost in SISR.
• The network's architecture successfully captures essential image features for high-quality super-resolution.
• IDSRN represents a significant advancement in efficient and effective single image super-resolution.