How Combinate and Family Day Care Australia automated the National Police Check experience with Machine Learning on AWS
In 2018, Insites and Family Day Care Australia launched an online web application that fully automates the national police check experience. This has sped up the police check application processing from weeks to a few minutes. The online application uses machine learning to automate the identity verification procedures and process images and documents.
This reduces the need for manual intervention. Instead of spending months of work training and building custom machine learning powered solutions, Insites has made the most out of the existing AI services of AWS for image and facial analysis. This decision helped reduce the engineering costs and allowed the company to speed up the launch of an online national police check application being used by hundreds of users per day.
AWS has three main layers for its machine learning services collectively known as Amazon AI. The first layer involves the managed AI services which include Amazon Rekognition for image and video analysis, Amazon Transcribe for speech recognition, and many other services and use cases.
The second layer involves ML Services and AI platforms which allow building and training of custom machine learning models. This includes Amazon SageMaker, Amazon Machine Learning, and Amazon EMR. The third layer involves AI engines, ML Frameworks, and Infrastructure that include open-source, deep learning frameworks on a convenient machine image such as Apache MXNet, Tensorflow, PyTorch, and Deep Learning AMIs.
From the broad range of machine learning services, Combinate has made the most out of Amazon Rekognition as building and deploying a reliable machine learning model from scratch for its facial recognition and OCR (Optical Character Recognition) requirements takes months (or years) to build along with a potentially expensive engineering team to manage and maintain.
One of the things that prevent the police check application process from being fully automated is the identity verification process. Without the automated machine learning powered solutions, the identity verification process is handled by human staff members from security agencies and a scheduled appointment might be required to verify a person's existence and identity.
With the availability of AI services such as Amazon Rekognition, this process is reduced to approximately a few minutes to at most an hour for most users given that scheduled appointments are no longer needed to verify the user's identity.
The information provided by the user is matched against the personal details and information presented in the uploaded documents automatically. This is made possible by Amazon's Rekognition DetectText operation that extracts the lines of text along with the bounding boxes where these lines of text are located in the image.
Performing the DetectText API call is almost instant. After several automated steps of processing the information from the user and the uploaded documents, the initial verification step can be completed in about a second. This greatly simplifies the architecture of the application as a synchronous implementation will involve less special cases to deal with. The DetectText operation accepts an image and returns the lines of text along with its properties and metadata. The DetectText operation is capable of handling scenarios where the uploaded image is tilted so no additional image processing steps may be required before the Amazon Rekognition API call.
Once the user has uploaded the documents and the system has performed the initial set of automated verification steps, the user is then required to take a "selfie" using the laptop's camera (or alternative).
If the CompareFaces API operation returns that the face in the “selfie” and in the primary photo do not match (similar to the example above), the identity verification fails, and an error message is shown to the user. Otherwise, the user can proceed with the next step(s). During the initial phases and assessment of the CompareFaces API operation of Amazon Rekognition for this project, images of twins have been used and Amazon Rekognition was able to detect that the faces do not match. Along with that, performing the CompareFaces API call is almost instant and along with the other automated steps of data processing, this verification step can be completed in about a second or two.
