Semi-TSGAN: Semi-Supervised Learning for Highlight Removal Based on Teacher-Student Generative Adversarial Network

Area of Science:

• Computer Vision
• Image Processing

Background:

• Highlight removal is crucial for image quality but hindered by limited annotated data and lightweight network deployment challenges.
• Existing methods struggle with data scarcity and efficient implementation for real-world applications.

Purpose of the Study:

• To propose a novel semi-supervised learning paradigm for effective highlight removal.
• To develop a lightweight network architecture balancing performance and computational efficiency.
• To address data scarcity and confirmation bias in highlight removal model training.

Main Methods:

• A semi-supervised learning framework merging a teacher-student model and a generative adversarial network (GAN).
• Utilizing a No-Reference Image Quality Assessment (NR-IQA) method to generate pseudo ground truth data.
• Incorporating contrastive regularization to mitigate overfitting on inaccurate labels.
• Designing a comprehensive feature aggregation module and attention mechanism for the generative network.

Main Results:

• The proposed algorithm demonstrates substantial quantitative and qualitative enhancements on highlight benchmarks.
• Experimental evaluations show superior performance compared to existing state-of-the-art methodologies.
• The lightweight network architecture facilitates efficient deployment.

Conclusions:

• The developed semi-supervised approach effectively tackles data scarcity in highlight removal.
• The integration of NR-IQA and contrastive regularization enhances model robustness and accuracy.
• This work presents a significant advancement in efficient and effective highlight removal techniques.