Features Guided Face Super-Resolution via Hybrid Model of Deep Learning and Random Forests.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Face hallucination (super-resolution) is challenging due to pose, illumination, and expression variations.
• Existing generative neural networks often prioritize perceptual realism over image fidelity, impacting downstream tasks.
• Many Convolutional Neural Network (CNN) methods use cascaded networks, obscuring investigative details.

Purpose of the Study:

• To develop a novel approach for face super-resolution that addresses the fidelity limitations of current methods.
• To create a generalizable method capable of handling diverse facial image conditions without pre-processing.
• To combine the strengths of deep learning and random forests for improved face super-resolution.

Main Methods:

• A Hierarchical CNN based Random Forests (HCRF) approach is proposed, operating in a coarse-to-fine manner.
• Two novel CNN models are developed for initial coarse facial image super-resolution and segmentation.
• Random forests are employed for targeted refinement of local facial features, utilizing segmentation outputs.

Main Results:

• The HCRF approach demonstrates comparable speed and competitive performance against state-of-the-art methods for very low-resolution images.
• Extensive benchmark experiments confirm the effectiveness of HCRF in both subjective and objective evaluations.
• The method successfully handles facial images under various conditions without requiring pre-processing.

Conclusions:

• HCRF offers a robust solution for face super-resolution, balancing perceptual quality with essential image fidelity.
• This work represents the first integration of deep learning and random forests for face super-resolution.
• The proposed method provides a significant advancement in generating high-quality super-resolved facial images from low-resolution inputs.