Facial recognition in schools: How the new SchoolBench™ web app identifies student images more accurately than humans
Facial recognition in schools is the next inevitable step for this technology, which has taken substantial steps forward during the past decade, and major leaps since Woodrow Wilson Bledsoe pioneered the field in the 1960s.
Back then, Bledsoe used a RAND tablet to manually trace facial features, such as eyes, nose, hairline, and mouth, and plot them into a database.
Later, when a new photo was manually scanned, the database could retrieve images deemed to be a close match.
Fast forward to 2017, and developers at Adelaide’s Parashift have rolled out SchoolBench™, a cross platform, java-based, web application that can be deployed locally on existing school infrastructure, using facial recognition technology to capture, index, and share digital photos and videos of students and staff.
Machines now better than humans at facial recognition
In the paper, Face Recognition Algorithms Surpass Humans, researchers concluded that while most of us, including computer vision researchers and psychologists, assume humans are better than machines at matching faces, that assumption is now challenged.
This is especially so when “matching face identity between photographs that are taken under different illumination conditions”.
Experiments conducted as part of the research show that machines leave humans behind except where the subjects are known to the humans doing the identification.
In school situations, especially schools with many hundreds of students, communication and administration staff often face situations in which they are sorting images of students unknown to them personally.
This is where processing images through an application such as SchoolBench™ not only delivers correct identification but also the accurate parsing of usage rights per student.
SchoolBench™’s Neural Network for Facial Recognition
Developers at Parashift based the neural network on Google’s Facenet because it has continued to prove itself as a leader in accuracy.
In fact, Google researchers published a paper, FaceNet: A Unified Embedding for Face Recognition and Clustering, in which they show that Facenet achieved nearly 100 per cent accuracy on a dataset of human images called Labeled Faces
Training happens by having the neural network guess the pictures based on known labels and then comparing it against how close it was to being correct. It will then continue to make guesses, tweaking values slightly to get closer and closer to an accurate way of guessing.
By using the LFW dataset to train Facenet and SchoolBench™ in facial recognition, the systems “learn” how to contrast and compare faces in different poses and lighting environments.
Furthermore, approximately one in ten of the people in the LFW database have two or more distinct photos in the collection, which adds extra finesse and rigour to the artificial intelligence.
It is common for researchers in the facial recognition field to benchmark their systems against Labeled Faces in the Wild because it provides a sound “verification” test.
This means high accuracy rates correctly analysing LFW images lead to algorithms performing at or above industry standard.
Humans teaching machines: Faster, better facial recognition in schools
Out of the box, SchoolBench™’s digital assets are stored in a single repository.
It then automatically classifies them using metadata from the image and video files, and applies indexation by referencing school-oriented taxonomy, such as class, year, term, and pupil names for later use.
The index of these images and media files is built up using a reverse term index, which stores the terms of each media file and a pointer to them rather than the media files individually. For example if you search for images in “Term 3” it will look up “Term 3” in the index and find all media files associated with that term. In this way, you don’t need to look through every image to find out which ones are in “Term 3”.
The resulting output means authorised users can quickly sort, view, and share digital media filtered by one or multiple fields, from class to location to usage rights.
While other systems can achieve this sorting output, it is the introduction of facial recognition using Facenet that speeds up the processing and accuracy of finding and sorting student record files.
From hours to seconds: We know who that is
By using reference images for each student, taken from annual school photo shoots, SchoolBench™ can create a unique, 128 byte number, or “signature”, for each student and then compare that to similar signatures extracted from faces found in images and videos.
Similar faces will have similar signatures, and by doing a distance comparison between each number, we can find all images that exist within a set threshold.
This means staff and other users can upload images and video in bulk and have them automatically scanned by SchoolBench™’s facial recognition system to:
- identify faces
- create a unique signature for each face
- look for matches it to the database
- add supporting data such as time, place, activity
- link the file to a student’s usage rights settings
In rare circumstances where an angle of a photo leads to a variation in the signature file for a student, a second signature is generated and added to that child’s record so they can be identified in various poses into the future.
Once a list of known faces to labels is generated for a given school’s data set, retraining can be applied just like using Labelled Faces in the Wild, except with the data from the school included, allowing SchoolBench™ to learn uniquely to your school.
What once took hours will now take seconds, with the ability to apply manual oversight to correct matches and help SchooBench become even more adaptable and robust in the future.
Our team enjoys sharing insights into the technical infrastructure of SchoolBench™ and would be happy to meet with you and your ICT colleagues to conduct a demonstration and Q and A session.