Abstract

Several models have been previously developed for learning correlated representations between source and target modalities. In this paper, we present a novel Joint Fusion model for learning cross spectral image representation for heterogenous face recognition. The coupled receptive face recognition model is built using ResNet architecture as backbone, a fully connected neural network and triple auto encoders for learning perceptible feature points invariant to changes in Spectrum. The performance of this model is tested using CelebA and LFW datasets. Moreover, the empirical results show that the learnt Common Latent Embeddings by the integrated networks produce competitive cross-spectrum face recognition results. These results are obtained by training the model using Adam Optimizer and Mean Squared Error (MSE) loss function. The proposed model has shown a performance improvement of 20% in AUC (Area Under the Curve) measure than the State-of-the-art with Polarization State Information, and 23% improvement in AUC over the State-of-the art models in traditional Thermal-to-Visible synthesis process. As well 12% improvement in EER (Equal Error Rate) measure polar measure and 9% improvement in EER (Conventional) are observed while comparing with Sate-of-the-art Models in Traditional thermal case.

Authors

Anita Sigamani¹, Prema Selvaraj²
B.M.S College for Women, India¹, Arulmigu Arthanareeswarar Arts and Science College, India²

Keywords

Flexible Filter, Bi-HCRV, FCNN, Face Detection, Recognition