I am trying to build a document scanner that is able to read text off of any document/card. However, it sometimes has trouble identifying text correctly off of a credit card. The accuracy is decent, but there is definitely room for improvement. I used the VisionTextRecognition framework and have used all the standard settings which are the right ones for setting up text recognition.
This is what I had to setup the text recognition request
textRecognitionRequest = VNRecognizeTextRequest(completionHandler: { (request, error) in
if let results = request.results, !results.isEmpty {
if let requestResults = request.results as? [VNRecognizedTextObservation] {
var foundText = ""
for observation in recognizedText {
guard let candidate = observation.topCandidates(1).first else { continue }
foundText.append(candidate.string + "\n")
}
}
}
})
textRecognitionRequest.recognitionLevel = .accurate
textRecognitionRequest.usesLanguageCorrection = true
Does anyone have any suggestions for improving the identification programmatically by either pre-processing or post-processing the scan at some point?
UPDATE: I've made a fully open source project that may help you do exactly what you need. Check it out: https://github.com/ethanwa/credit-card-scanner-and-validator
**
You can't do much to improve accuracy beyond adding some preset values to specifically look for, which doesn't make sense with CC numbers so I won't even bother showing that code. You'll need to rely on Apple to improve their text recognition model as iOS iterates for it to truly improve.
What I suggest in the meantime are these two things you can do:
Do validation on your credit card numbers that you think you're recieving. For example, Visa starts with 4, MasterCard starts with 5, Discover with 6, Amex with 3, etc. They have specific lengths and so on. See here: https://www.freeformatter.com/credit-card-number-generator-validator.html
Keep iterating over and over on a camera feed until you get a number that validates. I'm not sure if you are currently just taking a picture of the card, and processing that image (which it sounds like you are doing), but you should be processing many images per second until you get a valid CC. This is most likely how Apple does it when adding a card via Apple Pay on your phone, or when depositing checks digitally using banking apps (finding valid routing and account numbers).
Here's an example of what I mean...
I wrote this code that can pick out and validate ISBN numbers (basically 10 and 13 digit numbers that catalog books, which have a check digit for validation) in any given text and will keep looking until it finds all the numbers and then validates. It works extremely well and is very fast. Check out this Swift 5.3 code: