| Title | : | Face Recognition and Generation using Deep Learning techniques |
| Speaker | : | Avishek Bhattacharjee (IITM) |
| Details | : | Wed, 20 Nov, 2019 4:00 PM @ AM Turing Hall |
| Abstract: | : | Face Recognition (FR) has been a well-studied topic in the field of Computer Vision and Machine Learning. The performance of FR algorithms has vastly improved with the advent of Deep Learning. However, most of the recent literature on FR has dealt with the situation where the training and test data belong to an identical distribution, which does not relate to the real-world scenarios where the training and test data belong to different domains. Presentation of the work done, deals with a variety of problems, such as FR under constrained scenarios having poor illumination, pose variation, occlusion, blur as well as makeup. Additionally, we have also proposed a novel classroom dataset called Indian Classroom Face Dataset (ICFD), which ex plores the challenges faced while performing FR in classroom scenarios, and is a first of its kind dataset, to the best of our knowledge. Firstly, we have proposed a generative adversarial network (GAN) based approach (PosIX-GAN) to perform pose-invariant face recognition along with generating faces at nine different poses when given an input face image at any arbitrary pose, with the help of a novel “Patchwise MSE lossâ€. Secondly, we have proposed a fully convolutional deep model called SpoofNet, which performs the task of makeup removal from a makeup-induced face along with performing the task of FR on the face. The proposed network also generates the face of the person from a closed set of subjects who could have been spoofed by the makeup face, solely based on makeup, with the help of a novel reconstruction loss function. Thirdly, we have proposed a dual-pathway GAN based domain adaptation approach (DP-GAN) to bridge the gap between the source and target domains by generating crisp detailed gallery images, when provided with a probe image as input to the network. The generator of the GAN has two pathway s, which uses a novel Jensen-Shannon divergence-based loss, for capturing different components in the two pathways for superior reconstruction results. The GAN also incorporates a novel Multiscale Reconstruction loss to provide enhanced performance. Finally, we have proposed a dual-channeled GAN based domain adaptation approach (D2SC-GAN) to perform the task of generating gallery images from input probe images (captured in classroom scenarios) while also categorizing the generated images into corresponding classes. The approach makes use of a novel multi-resolution reconstruction loss consisting of a multi-resolution Patchwise MSE and a normalized chi-squared distance-based loss while also making use of a novel Kullback-Leibler divergence-based loss function to ensure that the two channels capture the low and high-frequency components of the generated image for superior reconstruction results. An online demo will also be shown at the end. Extensive experimentations, when performed on various datasets, in each work, have consistently shown the superiority of the proposed methods when compared to recent state-of-the-art methods. An online demo will also be shown at the end. |
