At Microblink, we’re proud to consider ourselves AI and identity document experts. And yet, we’re always striving to get better — both for our client partners and the end-users who interact with our products. On the heels of an exciting BlinkID release, we chatted with Luka Slibar, our BlinkID Product Manager, and Daniel Balchev, a Staff ML Engineer, about how we think about solving unique digital identity challenges and the image capture that powers our BlinkID suite. The below conversations have been edited and condensed for clarity.
Solving digital identity challenges with Luka Šlibar
I’ve been with Microblink in various capacities since 2014. While my initial focus was on business development, I eventually moved to a product management role for our AI initiatives until taking over as product lead for BlinkID’s capture and extract modules.
Over the years, Microblink has processed over 12 billion diverse identity documents, and in that time, we’ve developed a unique perspective on the nuances and challenges these documents contain. IDs vary by geography, person, and regularly change formats and layouts. Whether it’s opening an account or transferring money, most of the modern, digital user experiences that we take for granted today hinge upon the successful scanning and authentication of these diverse documents. That’s where BlinkID comes in.
When it comes to solving our clients’ challenges and raising the bar for digital identity, we break the process into three key steps: capture, extraction, and verification.
Exceptional document capture is foundational for downstream verification success. Whatever the unique business needs, there is no way for accurate, reliable identity document verification without phenomenal image capture and extraction. Put another way, the ability to verify a document as valid or flag it as potentially fraudulent is only as good as the data that’s extracted, which is only as good as the initial capture. If the capture experience is clunky or confusing or error-prone, users are likely to abandon it altogether.
Traditionally, Microblink has been known for developing and delivering SDKs and APIs to automate the first two steps in the process with industry-leading speed and accuracy. BlinkID’s on-device processing delivers a superior user experience with real-time feedback (e.g. “keep document still” or “move back”), data security, and the ability to support use cases without a strong internet connection.
Despite all the strides of mobile phone processing power, however, it remains secondary to the processing power of a full server. In addition, in many highly-regulated industries like banking & financial services, which our technology serves, captured images are sent to a server anyway for later processing.
While we’ve catered to our diverse client needs with flexible integration options and broad platform coverage, we also saw an opportunity to improve the existing approach, starting with image capture.
Microblink’s experience building highly optimized on-device products uniquely positioned us to solve this challenge. As a team, we know how to develop lightweight ML models and a user-friendly image capture experience without adding unnecessary size to the SDK. This also made it a compelling project for our cross-functional teams to work on.
Our just-released Capture SDK integrates into clients’ mobile applications to enable the fantastic user capture experience our flagship BlinkID product is known for, without adding much to the overall application size. It also allows for more powerful server-side image processing and extraction, which is essential for our long-term vision to deliver the best global identity document verification solution in the market.
Enabling even smarter, ML-powered image capture with Daniel Balchev
I joined Microblink at the end of 2022 as one of the first 3 employees in Sofia, Bulgaria, and I’m excited to release the first product I’ve worked on with the team! With our wealth of existing ML models and vast internal data to train new models, we were able to release this SDK in just 5 months.
It takes a village to build a product like this, and getting here was a close collaboration with multiple ML engineers, cross-platform software engineers, product designers, and our PMs. We also spoke with our customers and client-facing colleagues to incorporate their feedback and ensure we’re building the right product. The image capture SDK release is feature-full, including:
- Automatic document detection: Our SDK is able to detect when the document is visible and where the document is exactly, which allows us to determine when the user is holding a document too close or too far from the camera or at too much of an angle, providing instructions to guide the user to a better scan.
- Document type classification: Our new SDK classifies documents into several categories that help us determine their expected aspect ratio and whether the documents are single- or two-sided. This is particularly exciting because it allows us to tailor the user experience and scanning flow based on the specific document. In other words, if the user shows an ID document, we will know that we need to ask for the front and the backside, as opposed to just the front for a passport.
- Blur detection: detects if the document is blurred and waits until the user gets a better, steadier shot.
- Glare detection: If there’s glare covering the document, we will catch it and instruct the user to move the document so it doesn’t show.
- Tilt detection: knows where the document is in the image or camera frame and allows us to estimate if the document is tilted or not. Although we can support tilted documents in further processing, transforming such images inevitably results in some distortion of the transformed image, which are often undesirable when checking if an image is doctored. There is no way to correct for this post-capture, so the only solution is providing real-time end-user instructions on how to correctly position the document to remove the tilt and ensure the best possible image capture.
- Hand occlusion: This SDK reuses the hand detection models from our verification model to detect if a user’s hand is covering a significant part of the document. (What is considered “significant” can be configured in the setting by the client.) This ensures we don’t capture images where there is a finger covering an important part of the document, as an example.
- Sharpness estimation: We capture multiple frames of the identity document from the video stream, filtering them in order to select the one with the sharpest quality for the most accurate downstream processing.
- Lighting condition estimation: We can detect if the user is scanning in a dark room or if there’s too much light and instruct them to make the change to take a better image.
The result is more proactive, detailed user feedback that uses ML to check the image quality and ensures the best possible first-time scan, reducing the likelihood that users will have to repeat the process and resulting in better extraction.
Along with the superior capture experience and image quality that it provides, our standalone Capture SDK is not limited to a list of supported documents, so it is designed to work with all standard-sized ID cards, driver’s licenses and passports – today and in the future. If a new document version is released, clients don’t need to burden their users with an update but can simply update the backend for extraction.
Thank you to Luka and Daniel for chatting with us!
