|Title||:||Face Recognition under Surveillance Scenario using Domain Adaptation and Deep Learning|
|Speaker||:||Samik Banerjee (IITM)|
|Details||:||Tue, 25 Apr, 2017 3:00 PM @ BSB 361|
|Abstract:||:||Face Recognition (FR) in surveillance scenarios has recently attracted the attention of researchers over the last few years. The bottleneck as a large gap in both resolution and contrast between the training (high-resolution gallery) and testing (degraded, low quality probes) sets, must be overcome using efficient statistical learning methods. Performances of the recent state-of-the-art algorithms designed for FR are unsatisfactory when surveillance conditions severely degrade the test probes.
The talk will be a walkthrough from two of our proposed shallow learning techniques to a recently proposed deep learning technique. Firstly, we proposed a Bag-of-Words (BOW) based approach for face recognition combined with Domain Adaptation (DA), to overcome this challenging task of FR in degraded conditions. Transformation from source to target is estimated using eigen-analysis of the BOW-based features. Secondly, we present a novel technique to find the optimal feature-kernel combination by SML_MFKC (Soft-margin Learning for Multi-Feature-Kernel Combination) to solve the problem of FR in surveillance, followed by an Eigen Domain Transformation (EDT) to bridge the gap between the distributions of the gallery and the probe samples.
Finally, in the deep learning paradigm, we designed a novel transfer-CNN architecture of deep learning refurbished for domain adaptation (DA), to overcome the difference in feature distributions between the gallery and probe samples. A novel 2-stage algorithm for Mutually Exclusive Training (2-MET) based on stochastic gradient descent, has also been proposed for training the network. Rigorous experimentation has been performed on three real-world surveillance face datasets: FR_SURV_VID, SCFace and ChokePoint, as well as one real-world face dataset with non-uniform motion blur and three synthetically degraded large benchmark face datasets. Results have been shown using Rank-1 Recognition rates, ROC and CMC measures. Experiments also include performance analysis under unbiased training with two large-scale chimeric face datasets.