Abstract:
Name boards of places that are the most frequent visual aids in roadways, help to
identify the areas by guiding the traveller. Their most significant utility is in aiding
drivers who are inexperienced with the region and attempting to verify road names and
locations as they follow a map or route. Although Optical Character Recognition (OCR)
software is commonly available, OCR is still challenging in uncontrolled environments,
such as natural scenery, because of geometrical distortions, complicated backgrounds,
and various fonts. This study investigates the performance of some state-of-the-art
models on OCR introduced by Python on scene text images: Keras-OCR, Pytesseract,
and Easy-OCR. Besides traditional metrics such as Character Recognition Rate (CRR),
minimum edit distance distribution is included to reflect more on the obtained results
from each Python OCR tool. According to the minimum edit distance distribution,
Easy OCR produces somewhat better transcription results than Pytesseract, while
Keras-OCR produces noticeably better scene text transcription overall. CRR assessment
measure is employed to provide transcription outcomes for various category name board
pictures. Overall recognition rates at English character level for Pytesseract, Keras-
OCR, and Easy OCR are 34.87%, 94.19%, and 83.09% respectively. Recognition rates
at multi language character level for Pytesseract, Keras-OCR and Easy OCR are 66.66%,
35.80%, and 65.43%. The best Python tools for identifying name boards with improperly
aligned text, incomplete letters, and background noise are Keras-OCR and Easy OCR.
When detecting multilingual name boards with Tamil and English characters, Easy
OCR and Pytesseract outperform Keras-OCR. The quality of the input data affects
the Pytesseract output. Better outcomes come from precise text segmentation and a
backdrop free of background noise. Thus, to improve Pytesseract results, various preprocessing
approaches are needed.
