Image Interpolation Using Multi-scale Attention-aware Inception Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Image interpolation is crucial for enhancing image resolution and quality.
• Existing methods often suffer from aliasing artifacts and limited accuracy.

Purpose of the Study:

• To propose a novel multi-scale deep learning (MDL) framework for accurate image interpolation.
• To develop an attention-aware inception network (AIN) as the core seeding network for the MDL framework.

Main Methods:

• A novel attention-aware inception network (AIN) was designed, featuring feature extraction and feature-to-image mapping stages.
• Multi-scale image patches were generated using a proposed pyramid cut method to avoid aliasing issues.
• Trained AINs at each scale were combined for image processing during the testing phase.

Main Results:

• The proposed MDL framework, named MAIN, demonstrated consistent delivery of highly accurate interpolated images.
• Extensive simulations on seven image datasets (359 images) validated the framework's effectiveness.
• The pyramid cut method outperformed conventional image pyramid techniques by preventing aliasing.

Conclusions:

• The developed multi-scale deep learning framework provides a significant advancement in image interpolation.
• The attention-aware inception network (AIN) and pyramid cut method are key innovations for achieving high-accuracy image upscaling.