[Paper Review] Lightweight Image Super-Resolution with Adaptive Weighted Learning Network

Chaofeng Wang, Zheng Li+1 more|arXiv (Cornell University)|Apr 4, 2019

Advanced Image Processing Techniques38 references82 citations

TL;DR

Proposes a lightweight SISR network (AWSRN) with a local fusion block and adaptive weighted multi-scale module, achieving superior performance with similar parameter count and computation.

ABSTRACT

Deep learning has been successfully applied to the single-image super-resolution (SISR) task with great performance in recent years. However, most convolutional neural network based SR models require heavy computation, which limit their real-world applications. In this work, a lightweight SR network, named Adaptive Weighted Super-Resolution Network (AWSRN), is proposed for SISR to address this issue. A novel local fusion block (LFB) is designed in AWSRN for efficient residual learning, which consists of stacked adaptive weighted residual units (AWRU) and a local residual fusion unit (LRFU). Moreover, an adaptive weighted multi-scale (AWMS) module is proposed to make full use of features in reconstruction layer. AWMS consists of several different scale convolutions, and the redundancy scale branch can be removed according to the contribution of adaptive weights in AWMS for lightweight network. The experimental results on the commonly used datasets show that the proposed lightweight AWSRN achieves superior performance on x2, x3, x4, and x8 scale factors to state-of-the-art methods with similar parameters and computational overhead. Code is avaliable at: https://github.com/ChaofWang/AWSRN

Motivation & Objective

• Reduce computational cost for single-image super-resolution.
• Design a lightweight architecture with effective residual learning.
• Utilize adaptive weighting to select useful scales and features in reconstruction.

Proposed method

• Introduce a lightweight Adaptive Weighted Super-Resolution Network (AWSRN).
• Develop a local fusion block (LFB) consisting of stacked adaptive weighted residual units (AWRU) and a local residual fusion unit (LRFU).
• Propose an adaptive weighted multi-scale (AWMS) module with multiple scale convolutions and a removable redundancy branch guided by adaptive weights.
• Leverage adaptive weighted residual learning for efficient feature fusion in reconstruction.
• Evaluate performance on standard SR benchmarks across x2, x3, x4, and x8 scales.

Experimental results

Research questions

• RQ1 Can a lightweight network with adaptive weighting achieve competitive SR quality compared to heavier models?
• RQ2 How does the local fusion block (LFB) improve residual learning efficiency in a compact architecture?
• RQ3 What is the impact of the adaptive weighted multi-scale (AWMS) module on reconstruction quality and model efficiency?
• RQ4 Does AWRSN maintain superior performance with similar parameter counts and computational overhead across multiple upscaling factors (x2, x3, x4, x8)?

Key findings

• AWSRN achieves superior performance to state-of-the-art methods with similar parameters and computational overhead.
• The local fusion block enables efficient residual learning through stacked AWRUs and LRFU.
• AWMS exploits multi-scale convolutions and adaptive weights to utilize informative scales while removing redundant branches.
• The approach is effective across multiple upscaling factors (x2, x3, x4, x8).
• Results are demonstrated on commonly used SR datasets as per the paper.