場所：工学部 1号館 143講義室
題目： Identity-preserving face processing for better face recognition
講演者：Prof. Chia-Wen LIN
概要：Face detection/recognition is essential in video surveillance applications. However, the performance of current face detection and recognition schemes can be easily significantly degraded for difficult cases in real-world video surveillance applications such as detecting and recognizing low-resolution faces in the wild, and recognizing faces with illumination variations, viewpoint changes, and facial expression variations. In this lecture, I will show some methods and their results in addressing the difficult problems in face recognition for real-world video surveillance applications, including identity-preserving face hallucination, face augmentation and normalization. Our approaches are based on Siamese Generative Adversarial Networks (SiGANs) to achieve identity-preserving face hallucination and normalization. We incorporate the reconstruction error and identity label information in the loss function of SiGAN. By iteratively optimizing the loss functions of the generator and discriminator of SiGAN, we cannot only maximize the visual fidelity between the reconstructed faces and their ground-truths, but also ensure the reconstructed information is useful for identity recognition. Experimental results on large-scale public face datasets demonstrate the efficacy of the proposed approaches.