Face sketch synthesis via sparse representation-based greedy search.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing face sketch synthesis algorithms struggle with non-facial attributes (e.g., glasses, hairstyles) not present in training data.
• Previous methods are sensitive to variations in background, image size, and alignment, limiting their real-world applicability.

Purpose of the Study:

• To develop a novel face sketch synthesis method robust to non-facial factors and diverse image conditions.
• To improve the accuracy and applicability of face sketch synthesis for law enforcement and digital entertainment.

Main Methods:

• A novel method combining image patch similarity and prior knowledge for face sketch synthesis.
• Learning a photo patch feature dictionary and using sparse coefficients for patch replacement during synthesis.
• Employing Bayesian inference for final sketch generation after purifying nearest neighbors with prior knowledge.

Main Results:

• The proposed method successfully synthesizes face sketches incorporating non-facial factors absent in the training set.
• Demonstrated robustness against varied image backgrounds, sizes, and alignment issues.
• Outperformed several state-of-the-art methods in both perceptual and objective evaluation metrics.

Conclusions:

• The novel face sketch synthesis approach offers significant improvements in handling complex real-world image variations.
• This method enhances the utility of face sketch synthesis in practical applications by overcoming limitations of prior algorithms.