MADNet: A Fast and Lightweight Network for Single-Image Super Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) have advanced single-image super-resolution (SISR), improving metrics like peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).
• Existing CNN-based super-resolution models often demand substantial computational resources, hindering practical deployment.
• Many current methods underutilize intermediate features crucial for effective image reconstruction.

Purpose of the Study:

• To propose a dense lightweight network, MADNet, for enhanced multiscale feature expression and correlation learning in SISR.
• To develop a computationally efficient yet high-performing model for real-world super-resolution applications.

Main Methods:

• Introduced a residual multiscale module with an attention mechanism (RMAM) for improved multiscale feature representation.
• Developed a dual residual-path block (DRPB) to leverage hierarchical features from low-resolution images.
• Employed dense connections across blocks to effectively utilize multilevel features.

Main Results:

• MADNet demonstrated superior performance in SISR compared to existing methods.
• The proposed model achieved significant improvements in image quality metrics (PSNR, SSIM).
• MADNet requires considerably fewer computational resources (multi-adds) and parameters than conventional models.

Conclusions:

• MADNet offers an effective and efficient solution for single-image super-resolution.
• The network's design addresses the limitations of high computational cost and underutilization of intermediate features in existing CNN-based SISR models.
• This work facilitates the practical application of advanced super-resolution techniques.