Edge high quality case (skips numbers)

[mmacvicar]

I am dealing with this edge case where an excellent quality image is almost perfectly recognized, but surya will skip the number 4 every single time. I tested different bounding boxes, zoom levels and different types of noise but none made any difference. The detector is working fine but the recognizer reads R$33,00 instead of R$334,00 if I specify only portuguese. However, if I specify portuguese and english, it comes out correctly.

How would you reason about the influence of adding english when recognizing numbers? Is there anything particular in the training data that would make reasonable to always add english?

surya_ocr 1_z0.png --images --langs pt

{"1_z0": [{"text_lines": [{"polygon": [[0.0, 35.0], [1409.0, 35.0], [1409.0, 80.0], [0.0, 80.0]], "confidence": 0.9638850092887878, "text": " Valor: R$33,00 (TREZENTOS E TRINTA E QUATRO REAIS) ", "bbox": [0.0, 35.0, 1409.0, 80.0]}], "languages": ["pt"], "image_bbox": [0.0, 0.0, 1411.0, 116.0], "page": 1}]}

surya_ocr 1_z1.png --images --langs pt

{"1_z1": [{"text_lines": [{"polygon": [[1.0, 22.0], [605.0, 28.0], [597.0, 93.0], [0.0, 87.0]], "confidence": 0.9626762866973877, "text": "Valor: R$33,00", "bbox": [1.0, 22.0, 605.0, 93.0]}], "languages": ["pt"], "image_bbox": [0.0, 0.0, 605.0, 131.0], "page": 1}]}

[hvaz]

I am having the same issue

[israelsaba]

Me too