Heterogeneous face recognition aims to acknowledge faces across totally different sensor modalities. Typically, gallery images are normal visible spectrum pictures, and probe images are infrared images / sketches or thermal images. Now a day's vital improvements in face recognition are obtained by CNNs, gained knowledge from massive training datasets. In this paper, we are trying to find a match between a sketch with a digital photograph, a thermal image with a digital photograph, and an

more »

... d image with a digital photograph. We explored different machine learning methods to reduce the discrepancies between the various modalities. In this paper, we make use of some high-level features of deep convolutional neural networks that are trained on the digital photographs and these images do not belong to any type or any domain. CNN can also be used for encoding the images that are taken from various media. A generic framework for Heterogeneous Face Recognition is planned by making use of Deep Convolutional Neural Networks low-level features for each domain and it is called Domain-Specific Units. Domain-Specific Units extracts the shallow features for every new image domain. It also handles the transformation to a generic face space shared between all image domains. Experiments carried out on four face databases i.e CUHK Face Sketch Database (CUFS), CASIA NIR-VIS face database (CASIA), Near-Infrared and Visible Light (NIVL) dataset, Polarimetric And Thermal Database (Pola Thermal) covering 3 different image domains i.e. sketch, thermal image, infrared image, and show improvements, in terms of matching ratio.