revert division by 2 assumption of previous commits to resolve scribeocr issue 40

[rmast]

scribeocr/scribeocr#40

[Balearica]

Any change to the grid size is going to cause a non-zero number of errors, and solve a non-zero number of errors. Indeed, this change was made because it fixed errors that occurred for other documents. On conceptual grounds, and based on previous documents, I currently believe that using the smaller grid size for noise removal is a net positive.

However, I could always be wrong. Ultimately, this type of question is currently unanswerable given that all we have are individual examples. To make an educated decision on questions of this nature, the impact would need to be quantified using an automated benchmark tool that tests accuracy against a (relatively) large and diverse corpus of documents. Most of the code for this already exists, so it would be a matter of building a diverse corpus of images and ground truth data.

[rmast]

@GerHobbelt another vote for a testing corpus/facility! Verzonden vanaf Outlook voor Android<https://aka.ms/AAb9ysg>

…

________________________________ From: Balearica ***@***.***> Sent: Monday, September 2, 2024 7:24:31 PM To: Balearica/tesseract ***@***.***> Cc: rmast ***@***.***>; Author ***@***.***> Subject: Re: [Balearica/tesseract] revert division by 2 assumption of previous commits to resolve scribeocr issue 40 (PR #3) Any change to the grid size is going to cause a non-zero number of errors, and solve a non-zero number of errors. Indeed, this change was made because it fixed errors that occurred for other documents. On conceptual grounds, and based on previous documents, I currently believe that using the smaller grid size for noise removal is a net positive. However, I could always be wrong. Ultimately, this type of question is currently unanswerable given that all we have are individual examples. To make an educated decision on questions of this nature, the impact would need to be quantified using an automated benchmark tool that tests accuracy against a (relatively) large and diverse corpus of documents. Most of the code for this already exists, so it would be a matter of building a diverse corpus of images and ground truth data. — Reply to this email directly, view it on GitHub<#3 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAZPZ5TTBTDTUYWCZ3Q7KX3ZUSNM7AVCNFSM6AAAAABMIJMGLSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGMRVGEZDOOBVGY>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Balearica]

@rmast I created a benchmark repo here that contains an initial set of 23 diverse images and ground truth data, along with code for running comparisons against OCR results. The initial results are in the README. I plan to use this as a baseline when assessing future changes.