This paper introduces a novel adapter framework to improve face recognition performance in real-world conditions with low-quality images. The key aspects of the approach are:
Dual-Input Processing: The framework processes both the original low-quality (LQ) images and the high-quality (HQ) images restored by a face restoration model. This dual-input design minimizes the domain gap and provides complementary perspectives for the face recognition model.
Adapter Design: The adapter consists of a trainable HQ branch that processes the restored HQ images, while the pre-trained face recognition model handles the original LQ images. This allows the adapter to leverage the capabilities of the pre-trained model without losing its knowledge.
Fusion Structure: The framework employs a fusion structure with nested Cross-Attention and Self-Attention mechanisms to effectively integrate the features from the LQ and HQ branches. This fusion process enhances the model's ability to recognize faces accurately in diverse image quality conditions.
Extensive Experiments: The authors conduct experiments on both synthetic and real-world datasets, demonstrating the effectiveness of their approach. The results show significant improvements in face recognition accuracy compared to baseline methods, especially in zero-shot settings and under real-world degradation conditions like atmospheric turbulence.
The proposed adapter framework sets a new standard in face recognition, offering a robust and versatile solution for various applications, including surveillance, mobile authentication, and other real-world scenarios with varying image quality.
他の言語に翻訳
原文コンテンツから
arxiv.org
深掘り質問