On Designing MWIR and Visible Band based DeepFace Detection Models

In this work, we propose an optimal solution for face detection when operating in the thermal and visible bands. Our aim is to train, fine tune, optimize and validate preexisting object detection models using thermal and visible data separately. Thus, we perform an empirical study to determine the most efficient band specific DeepFace detection model in terms of detection performance. The original object detection models that were selected for our study are the Faster R-CNN (Region based Convolutional Neural Network), SSD (Single-shot Multi-Box Detector) and R-FCN (Region-based Fully Convolutional Network). Also, the dual-band dataset used for this work is composed of two challenging MWIR and visible band face datasets, where the faces were captured under variable conditions, i.e. indoors, outdoors, different standoff distances (5 and 10 meters) and poses. Experimental results show that the proposed detection model yields the highest accuracy independent of the band and scenario used. Specifically, we show that a modified and tuned Faster R-CNN architecture with ResNet 101 is the most promising model when compared to all the other models tested. The proposed model yields accuracy of 99.2% and 98.4% when tested on thermal and visible face data respectively. Finally, while the proposed model is relatively slower than its competitors, our further experiments show that the speed of this network can be increased by reducing the number of proposals in RPN (Region Proposal Network), and thus, the computational complexity challenge is significantly minimized.