Dynamic Merging of Word Bounding Boxes

[satvik-27199]

Currently, the algorithm for merging word bounding boxes combines words into a single bounding box if they are close enough. While this works in many cases, it can cause issues in others where words should remain separate despite their proximity. This leads to inaccurate text extraction and poor representation of the original content.

Example:

Consider the following text snippet:

Word1 Word2 Word3 Word4

In the current implementation, if Word1 and Word2 are close enough, they might be combined into a single bounding box as Word1Word2. Similarly, Word3 and Word4 might also be merged if they are close. This results in incorrect extraction as shown below:

Word1Word2 Word3Word4

I wonder if there any sophisticated merging strategy that takes into account:

1. The average character width and spacing.
2. The context of the text, ensuring words that are naturally separate remain so.
3. A threshold that adjusts dynamically based on font size and text density.

[GreatV]

Could you please tell me which algorithm you are describing and if a corresponding example is provided?

[JoshvirN]

I would like some insight into this as well. I trained paddle detection model on word level but the inference still shows line boxes. Is there a way to get unmerged boxes?