Intel's Latest Research: Facial Recognition from Thermal Images

2020-01-13

Is the thermal image detailed enough for AI models to recognize facial features? This is a question raised by Intel researchers at the recent 12th International Conference on Human-Computer Interaction. They studied the performance of models trained on visible light data and subsequently trained on thermal images.


As researchers imagine, thermal imaging is often used in environments such as medical institutions that prioritize or require privacy protection, instead of RGB camera data. This is because it can obscure details of personal identification, such as eye color and chin contour.


The team utilized two facial thermal image datasets. The first dataset (SC3000-DB) was created using the Flir ThermaCam SC3000 infrared camera. It contains 766 images from 19 male and 21 female volunteers. The second dataset comes from the IRIS dataset of the Visual Computing and Image Processing Laboratory at Oklahoma State University, which includes 4190 images collected with the assistance of 30 individuals, many of whom moved their heads and used different facial expressions.


The researchers first cropped the images to only include the region with the face using a machine learning model. Then, they employed another model to extract facial features from the images and represent them as vectors (sets of values corresponding to specific features). Finally, they used a third model trained on visible light images to verify if this model could be applied to thermal images. During these experimental processes, the researchers compared two types of facial feature vectors: one that identifies the person depicted in a given image, and another based on the similarity between the profile and the input image. They found that the model trained on visible light image data could be well generalized to thermal images, successfully extracting facial features and distinguishing between volunteers, with an accuracy of 99.5% for SC3000-DB and 82.14% for IRIS.


The team hopes to expand the scope of the research to data collected in various other situations, such as when the subject's head is rotated horizontally or vertically. The researchers wrote, "Many promising visual processing applications, such as non-contact vital sign estimation and smart home monitoring, may involve sensitive personal data. Therefore, thermal imaging can not only provide useful data but also conceal personal identity, making it suitable for many applications."

RECENT AI NEWS

RECENT AI TOOLS