Adaptive Transform Domain Image Super-resolution Via Orthogonally Regularized Deep Networks

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) excel at single image Super-Resolution (SR) by mapping low-resolution (LR) to high-resolution (HR) images in the spatial domain.
• Existing deep learning SR methods often require extensive training data and can introduce artifacts, such as those from bicubic interpolation.

Purpose of the Study:

• To propose a novel network structure for image super-resolution operating in an image transform domain.
• To develop an adaptable and efficient deep learning model for SR that mitigates common challenges like data requirements and artifact generation.

Main Methods:

• Introduced a Convolutional DCT (CDCT) layer to integrate the Discrete Cosine Transform (DCT) into a deep learning network.
• Developed the DCT Deep SR (DCT-DSR) network and further enhanced it into the Orthogonally Regularized Deep SR (ORDSR) network with trainable CDCT layers.
• Applied pairwise orthogonality and complexity order constraints to the CDCT layer's basis functions for optimized transform adaptation.

Main Results:

• The ORDSR network achieved state-of-the-art SR image quality using fewer parameters compared to many deep CNN methods.
• ORDSR demonstrated significant success in reducing artifacts commonly introduced by bicubic interpolation.
• The network showed more graceful degradation in performance with reduced training data, offering benefits for limited training scenarios.

Conclusions:

• The proposed ORDSR network effectively leverages the image transform domain for efficient and high-quality super-resolution.
• ORDSR offers a promising alternative to spatial-domain deep learning SR methods, particularly concerning parameter efficiency, artifact reduction, and reduced training data dependency.
• The model's adaptability and efficiency suggest potential for broader applications in image restoration and enhancement.