Pipeline

Every image provided to the SDK will pass through the same pipeline:

• 1. STN (Spatial Transformer Network) processing to retrieve the document position and attention map.

• 2. The document will be rectified (de-skewed) and anything outside the attention map will be ignored.

• 3. Perform full MRZ search on the rectified image.

• 4. Perform MRZ recognition on the image (if it has some).

• 5. Perform document classification on the detected cards (in batches if multiple). Filter the possibilities using the MRZ result if the card has one.

• 6. Perform graph-match on the detected cards (against the matched graphs from the dataset) to localize each field (text, portrait, signature…)

• 7. Perform image registration on each field and run the OCR to extract the text.

• 8. Return the result as JSON string.

The next sections explain how each module (MRZ recognizer, Card Detector, Card Classifier, Graph computation…) work.