Global Aligned Structured Sparsity Learning for Efficient Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Efficient image super-resolution (SR) models are crucial but often resource-intensive.
• Existing lightweight architectures and compression techniques have limitations in fine-grained redundancy removal.
• Network pruning offers a solution but faces challenges with structured pruning in SR networks due to residual blocks and sparsity determination.

Purpose of the Study:

• To introduce a novel method, Global Aligned Structured Sparsity Learning (GASSL), for efficient network pruning in image super-resolution.
• To address the difficulties of structured pruning in SR networks, specifically aligning pruned indices across layers and determining layerwise sparsities.
• To develop new, highly efficient single image SR networks using the proposed GASSL framework.

Main Methods:

• GASSL comprises two key components: Hessian-Aided Regularization (HAIR) for automatic sparsity selection and Aligned Structured Sparsity Learning (ASSL) for pruning.
• HAIR utilizes a regularization-based approach, implicitly considering Hessian information for sparsity auto-selection, supported by a theoretical proposition.
• ASSL introduces a novel penalty term, Sparsity Structure Alignment (SSA), to ensure consistent pruned indices across different network layers.

Main Results:

• GASSL enables the design of two novel, efficient single image SR networks across different architectural styles.
• The proposed method successfully pushes the efficiency boundaries of SR models.
• Experimental results validate the superiority of GASSL compared to other state-of-the-art methods.

Conclusions:

• GASSL effectively resolves critical challenges in structured pruning for image super-resolution networks.
• The method facilitates the creation of more efficient SR models by optimizing network redundancy at the filter level.
• GASSL represents a significant advancement in developing efficient and high-performing image super-resolution solutions.