Search Icon
Back to Blog

Reconstructing Barcodes with Generative AI

While this year has already seen the industry-leading performance from our new DocV 3.0 platform and Capture SDKs, our team is pushing the boundaries for superior performance across every dimension. Using generative AI, we’ve just released improvements to model performance, while enhancing the experience for the end user and increasing funnel conversion rates for our customers as more users successfully go through the process.

These improvements include:

  • Increased the barcade readability rate from 68% to 89% with our detection and reconstruction model
  • Reduced image validation resubmit rates from 11% to just 3%.
  • Conclusive responses now delivered in just 2 seconds.

Why is the Barcode so Important?

With most document verification providers, a waterfall typically performs different checks. The extracted information from the document, like the name and date of birth, often comes from the front of the ID where that information is written in “human-readable” characters. However, extracting written information from an image of a photo ID is an imperfect science. Optical character recognition (OCR) has improved dramatically over the last several years with the recent advances in machine learning, yet even the most accurate systems may produce errors. For example, the perfect glare on an ID might make an “E” look like an “F”. Oftentimes this can be correctable (e.g. JAKF instead of JAKE) but sometimes it is more ambiguous (e.g. CHAFF instead of CHAFE). Often, extracting the right data is very important, and OCR will never be completely perfect given real world conditions.

Barcodes, on the other hand, are intended to be read by machines. In particular, the PDF417 barcode on the back of all US and Canadian driver licenses and ID cards includes rich data duplicated from the front of the ID, including the name and date of birth of the holder. However, it has the benefit of including robust error correction codes. What this means in practical terms is that if a machine is able to read a barcode successfully, you can be confident that what it read was the correct information. This is why when you are checking out at a supermarket, it’s exceedingly rare for the scanner to misread a box of cereal as laundry detergent. The typical way that scanners fail is that it can’t read the barcode, like when you go through self checkout with an item that doesn’t seem to want to scan. The item then needs to be keyed in manually, which is more prone to errors, similar to OCR.

So with this rich data on an ID, why doesn’t all document verification software simply leverage the barcode if it is a source of rich, high-accuracy data? When the setting is not ideal, the software may have trouble scanning the PDF417 barcodes. Perhaps the lighting is low, the camera quality is poor, or the barcode itself has been damaged. In each of these cases, the barcode is often not readable by default. In order to read those barcodes, some system or process would need to make otherwise unreadable barcodes readable to leverage this rich source of information.

Barcode Enhancement Model

As part of our new 3.0 platform, our DocV team has trained a new barcode enhancement model that takes an imperfect barcode as input and using generative AI for the purpose of reconstruction, outputs a clean, “super resolution” barcode.

To protect real PIIs, we used synthetically generated personal data, from which we generated perfectly readable barcodes. We then further artificially degraded the perfect readable barcodes through a degradation process which includes downsampling, blurring, shadowing, adding wear-and-tear effects, and changing background colors. The degraded barcodes were then fed to the generative super resolution model as inputs. The model would learn to output reconstructed super resolution barcodes to match with the original perfectly readable barcodes as much as possible.

Document Verification with Generative AI

When it comes to innovation and market-leading performance in ID document verification, Socure’s fully automated solution is setting the standard.

Want to see Predictive DocV in action? Sign up for a live demo here.

Feng Xiao

Feng Xiao serves as Senior Director and Head of Socure's Computer Vision team, supporting the DocV solution.