Pixel flip experiment
The objective of the experiment is to see OCR app can recognize text better if we eliminate the background image.
The experiment is done on images from O3-APR-002 video. The video used annotation style 1 (where white font text is bounded by black box). In this experiment, all pixels that are not black are converted to black and pixels that are pure black are converted to white.
EasyOCR
The pixel flipping method did not improve the EasyOCR's performance. The EasyOCR could not extract text correctly from both images (regular and the pixel flipped images).
Azure
The pixel flipping method worsened the Azure's performance. The Azure OCR is able to extract the text correctly from the regular images.
Error when doing pixel flip on images from 03-APR-004 videos
- There was no problem doing pixel flip on images from 03-APR-002 videos.
- The output images did not come out correctly when doing pixel flip on images from 03-APR-004 videos.