| Title | : | Deep Models for unconstrained Face Recognition |
| Speaker | : | Samik Banerjee (IITM) |
| Details | : | Tue, 27 Mar, 2018 3:00 PM @ A M Turing Hall |
| Abstract: | : | The problem of Face Recognition (FR) under unconstrained scenarios has been a major focus for researchers in the field of Computer Vision (CV), over the last two decades. Degradation of a face image occurs due to several reasons, such as low-contrast and illumination, low-resolution (LR), disguise due to make-up, occlusion, pose, etc. Any of these artifacts lead to unsatisfactory performance of the state-of-the-art FR algorithms, as training (gallery) and test (probes) samples differ considerably. This talk will be a walk-through through several deep-learning techniques designed by us to overcome some challenges resulting due to degraded face probes used for FR. Firstly, we propose a transfer-CNN (also termed deep-DA) architecture with a novel 3-stage mutually exclusive training (3-MET) algorithm for FR under surveillance conditions. The adaptation to degraded probe samples over a model trained with HR gallery samples (captured under laboratory conditions), was done using a novel Mutual Variation of Information (MVI) loss function on an auto-encoder based Transfer Module (TM) in the deep-DA architecture consisting of 3 components. Using Rank-1 Recognition rates as well as ROC and CMC measures, results have been shown to be superior over several state-of-the-art deep-CNN methods, on three real-world surveillance face datasets, one real-world face dataset with non-uniform motion blur and three synthetically degraded large benchmark face datasets. Secondly, to reconstruct realistic super-resolved (SR) mugshot images from LR probe samples, we designed an LR-GAN (Low-Resolution Generative Adversarial Network) which optimizes a multi-scale reconstruction loss providing rich performance, as evident by the high (> 90%) rank-1 recognition rates, on four benchmark datasets (3 surveillance and 1 blurred). Furthermore, we propose a bimodal mutually exclusive Structural and Denoising GAN (SD-GAN), to reconstruct the missing parts of the face under occlusion, preserving the intensity distribution (spatial) and identity of the face. A novel adversarial training algorithm minimizes a structural loss function, consisting of a holistic and a local loss criteria. Ablation studies on real and synthetically occluded face datasets reveal that our proposed technique outperforms the recent competing methods by a considerable margin. Finally, we designed an end-to-end Siamese convolutional neural network (SCNN) that simultaneously replicates the make-up on any face (near-frontal) w.r.t. a reference image with facial make-up, as well as verifies the identity of a face sample either with or without facial make-up. The proposed architecture of MakeUpMirror reciprocates the make-up at exact locations of the face without any human interventions using a categorical reconstruction loss. SCNN outperforms recent state-of-the-art techniques for face verification under facial make-up, when verified using three benchmark datasets. |
