A Method for Non-Rigid Face Alignment via Combining Local and Holistic Matching.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Non-rigid face alignment is crucial for various applications but is challenged by complex geometric deformations and outliers.
• Existing methods often require multiple templates or struggle with robustness to significant facial expression changes.

Purpose of the Study:

• To develop a robust and accurate non-rigid face alignment method utilizing only a single facial template.
• To address the limitations of conventional methods in handling complex geometric transformations and outliers.

Main Methods:

• Proposed a novel local feature matching method, K Patch Pairs (K-PP), to identify mutual nearest neighbors for robust outlier handling.
• Introduced a modified Lucas-Kanade algorithm incorporating local matching constraints for joint holistic and local face representation.
• Utilized an inverse compositional algorithm for efficient optimization of the face alignment problem.

Main Results:

• The K Patch Pairs (K-PP) method effectively balances similarity and the number of local matches, improving robustness to outliers.
• The combined approach integrates the flexibility of local matching with the robustness of holistic fitting.
• Comparative analyses demonstrate superior accuracy and robustness compared to conventional non-rigid face alignment techniques.

Conclusions:

• The proposed single-template non-rigid face alignment method offers significant improvements in accuracy and robustness.
• The K Patch Pairs (K-PP) feature matching and modified Lucas-Kanade algorithm provide an effective solution for challenging facial alignment tasks.