Face Detection in MWIR Spectrum

The capability to perform face recognition in the visible and thermal spectra is of prime interest in many law enforcement and military organizations. Face detection is an important pre-processing step for face recognition. Though many algorithms are available for face detection in the visible spectrum, an assessment of how these algorithms can be retrained for the thermal spectrum is an important study. Current available visible-based face detection algorithms are very effective in daytime conditions, however, when there are extreme changes in illumination conditions like very low-light to no light (night-time), these become challenging. Due to limited amount of data available for researchers from sensors in the thermal band (due to the increased cost of having and operating state of the art thermal sensors), there are only a few proposed algorithms. In this work, we conducted a study to determine the impact of factors such as indoor/outdoor environment, distance from the camera, application of sunscreen, training set size, etc. on training deep-learning models for a face detection system in the thermal spectrum that simultaneously performs face detection and frontal/non-frontal classification. Existing deep learning models such as SSD (Single Shot Multi-box Detector), R-FCN (Region Based Fully Convolutional Network) and R-CNN (Region Based Convolutional Neural Network), are re-trained using thermal images for face detection and pose estimation tasks. Results from each model are compared, and the model with the best performance is further trained and tested on different datasets, including indoor, outdoor at different stand-off distances. The highest accuracy is achieved using a Faster R-CNN model with ResNet-101 and the accuracy is 99.4% after a 10-fold cross-validation. More experiments are performed to further study the efficiency and limitations of this model. The data set we use was collected under constrained indoor and unconstrained outdoor conditions.