Hybrid domain enhancement network for lightweight image super-resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning, particularly convolutional neural networks (CNNs), has advanced image super-resolution (SR).
• Existing SR methods often exhibit high computational complexity and memory usage, limiting their deployment on devices with limited resources.

Purpose of the Study:

• To propose a lightweight hybrid domain enhancement network (HDEN) for efficient image super-resolution.
• To address the challenges of computational complexity and memory demands in current SR approaches.

Main Methods:

• The proposed HDEN utilizes a hybrid domain enhancement module with parallel spatial and frequency domain branches.
• A spatial domain enhancement block (SDEB) extracts multi-scale features using wide-activated residual units with varying dilation factors.
• A frequency domain enhancement block (FDEB) employs wavelet transform to process frequency domain features, enhancing details like edges.

Main Results:

• The HDEN demonstrates superior performance compared to other lightweight SR methods.
• Quantitative metrics and visual quality assessments confirm the effectiveness of the proposed network.

Conclusions:

• The lightweight HDEN effectively enhances image super-resolution by leveraging both spatial and frequency domain features.
• HDEN offers a promising solution for deploying high-quality image super-resolution on resource-constrained devices.