Hallucinating Face Image by Regularization Models in High-Resolution Feature Space.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Conventional methods for reconstructing high-resolution (HR) face images from low-resolution (LR) inputs often rely on local geometry consistency assumptions.
• These assumptions can be challenging to maintain across different resolutions, limiting reconstruction quality.
• Existing techniques struggle with preserving intricate details and high-frequency information in complex facial structures.

Purpose of the Study:

• To propose two novel regularization models for efficient HR face image reconstruction from LR inputs.
• To develop a method that directly regularizes relationships in the HR space, bypassing local geometry consistency issues.
• To enhance the reconstruction of fine facial details and preserve high-frequency information, particularly on complex edges.

Main Methods:

• Introduced a patch-wise regularization model that directly relates target patches to the training set in HR space.
• Utilized kernel functions to capture nonlinear characteristics in a high-dimensional kernel space for patch-based reconstruction.
• Developed a pixel-wise regularization model to enhance fuzzy details by prioritizing reconstruction along dominant structural orientations.

Main Results:

• The proposed patch-wise and pixel-wise models were combined into a unified framework for iterative optimization.
• Experimental results demonstrated superior performance compared to existing state-of-the-art face hallucination methods.
• The method effectively reconstructs HR face images, preserving intricate details and high-frequency information.

Conclusions:

• The novel regularization models offer an efficient and effective approach to face hallucination.
• Directly regularizing HR space relationships and incorporating pixel-wise detail enhancement significantly improves reconstruction quality.
• The unified framework provides a robust solution for generating high-resolution face images from low-resolution inputs.