Learning lightweight super-resolution networks with weight pruning.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) have significantly advanced single image super-resolution (SISR).
• Current deep learning-based SISR methods are computationally intensive and memory-demanding, hindering deployment on resource-constrained devices like mobile phones.

Purpose of the Study:

• To develop a lightweight and accurate SISR network through network compression.
• To enable the practical deployment of state-of-the-art SISR models on mobile devices.

Main Methods:

• Implemented a novel weight pruning approach to compress SR networks.
• Utilized a progressive optimization method for gradual parameter zeroing.
• Developed a sparse-aware attention module using a pruning-based strategy.
• Proposed an information multi-slicing network for multi-scale feature extraction and integration.

Main Results:

• The proposed weight pruning method effectively reduces model size.
• Performance degradation is minimal, enabling practical application of SISR models.
• The pruned versions achieve superior accuracy and visual quality compared to existing methods.

Conclusions:

• Weight pruning is a viable strategy for creating efficient and effective SISR networks.
• The developed techniques facilitate the real-world application of advanced image super-resolution.
• The proposed method offers a balance between model size reduction and performance enhancement.