Suggestion for preprocessing of Image.

[erictan23]

Hello! Thank you for sharing PaddlePaddleOCR, I am trying to read a text from a CCTV surveillance video footage. However, I encounter some characters that are black and background is black, resulting in a poor performance of the OCR model. The Image format is in PNG.

The results by the model are: 02+60023-We1-1249:05

I have tried some pre-processing methods such as changing of grayscale and median blurring and the results slightly improve.
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
median = cv2.medianBlur(gray, 7)

The results by the model are: 09-20023Wed-1249:05

are there any suggestions to improve on the results?

[GreatV]

In my experience you need to fine-tune the model by collecting some data for the corresponding scenarios. A generic model may not be sufficient.

[zhangyubo0722]

Yes, you need to collect some data to retrain the model.