Distance-Based Classification of Biometric Images: Leveraging Deep Learning Models

In biometric imaging, data are typically pre-processed manually; this slows down data acquisition and may result in undetected mistakes. Parameters like standoff distance are recorded by hand due to state-of-the-art distance verification algorithms suffering from a lack of datasets at long distances. In this work, we discuss the latest multi-spectral face dataset containing face images from visible, mid-wave IR, and standoff distances ranging from 2 to 400 meters. Our study determines an efficient deep learning solution that accurately determines the capture conditions of visible band face images. Using transfer learning, deep learning models are trained to infer face standoff distance from a subset of biometric data collected “in the wild.” Our efforts yield a face image classification accuracy exceeding 99% and a video classification accuracy of 100% with InceptionV3. Automatically classifying face standoff distance is crucial in data organization and labeling efforts such as the ones undertaken in biometric data collection. We demonstrate that face image standoff distance classification is possible for large-scale, long-distance datasets and achieves accurate results; this can accelerate data preprocessing and mitigate manual labeling errors.