Issues with Multi-Column Text Recognition and the Need for hOCR Support
This is a great project. I tested it with a variety of Arabic texts, and it recognizes the content quite well. However, one issue I noticed is that it struggles with multi-column layouts, especially when a page contains three or more columns. In a test page from an old Arabic magazine with three columns, the model processed the two rightmost columns but ignored the left one. I tried adjusting different settings, but it never managed to process the entire page.
Another issue I encountered with columned articles is the messy output. The model sometimes treats adjacent columns as a single text block, combining lines from different columns into one line in the resulting text. This brings me to the feature request mentioned in the title: support for hOCR output. Having hOCR data would greatly simplify the creation of searchable PDFs, which could then be used in training datasets.
Thanks for sharing. We will look into the hOCR data format. Feel free to share if you find any other edge cases. We would love to fix as many as possible in our next release. If possible, do share one example here; it will be easier for us to collect similar data.
Sure, here are two example images for the same page—one in color and the other in grayscale. I deskewed the images by 1 degree in one run to improve the results, but that didn’t help, and the model still couldn’t recognize the whole pages.
