Adding khmLimon.traineddata

[phyrumsk]

Hi we finetune the tesseract_best engine which support only Khmer unicode fonts with new 6 Limon fonts such as Limon S1 S2 F1 F2 R1 R2 using the same tesseract netspec.

[Shreeshrii]

Thank you for this. Do you have any statistics on improved accuracy with this for Limon fonts?

I am not sure what is the policy regarding community contributed traineddata files, though we do have ancient greek traineddata contributed by Nick White.

[phyrumsk]

Hi Shreeshrii,
Please find the attached of the accuracy results (Sample of testing images are also included).
We used isri-ocr-evaluation-tools to test for accuracy and there are 2 types of images to be tested, images which are generated using text2image command (without noise) and the scan images (with noise).

https://github.com/eddieantonio/isri-ocr-evaluation-tools

2018_06_04_KhmLimon_result_for_github.zip

Thanks,

[Shreeshrii]

Thank you!

@theraysmith @jbreiden @zdenop What is the policy for adding community contributed tessdata?

In addition to this, I have generated traineddata files for Coptic and Javanese.

[Shreeshrii]

Accuracy_Tesseract_4.0

Testing Text:		Characters :	19618			Clusters :		11272

No.	Font Name	Size	Character Accuracy		Cluster Accuracy
			Errors	Accuracy	Errors	Accuracy
1	Khmer OS	12	1495	92,38%	650	94,23%
2	Khmer OS Battambang		562	97,14%	140	98,76%
3	Khmer OS Bokor		3288	83,24%	1846	83,62%
4	Khmer OS Content		644	96,72%	159	98,59%
5	Khmer OS Fasthand		4605	76,53%	2012	82,15%
6	Khmer OS Freehand		2184	88,87%	505	95,52%
7	Khmer OS Metal Chrieng		3230	83,54%	1694	84,97%
8	Khmer OS Muol		525	97,32%	135	98,80%
9	Khmer OS Muol Light		3430	82,52%	2115	81,24%
10	Khmer OS Muol Pali		6697	65,86%	3688	67,28%
11	Khmer OS Siemreap		1318	93,28%	582	94,84%
12	Khmer OS System		2338	88,08%	1076	90,45%
13	Noto Serif Khmer Bold		2006	89,77%	1038	90,79%
14	Noto Serif Khmer Regular		2043	89,59%	1045	90,73%
Average :			2,454,64	87,49%	1,191,79	89,43%

No.	Font Name	Size	Character Accuracy		Cluster Accuracy
			Errors	Accuracy	Errors	Accuracy
1	Limon F1	22	11459	41,59%	7340	34,88%
2	Limon F2		11152	43,15%	7652	32,11%
3	Limon F3		8454	56,91%	5590	50,41%
4	Limon F4		17082	12,93%	10319	8,45%
5	Limon F5		12259	37,51%	7689	31,79%
6	Limon F6		6146	68,67%	3858	65,77%
7	Limon F7		6750	65,59%	4031	64,24%
8	Limon F8		5512	71,90%	3482	69,11%
9	Limon R1		7053	64,05%	4336	61,53%
10	Limon R2		8044	59,00%	4584	59,33%
11	Limon R3		10467	46,65%	5357	52,48%
12	Limon R4		11815	39,77%	7068	37,30%
13	Limon R5		9479	51,68%	5811	48,45%
14	Limon S1		2704	86,22%	1479	86,88%
15	Limon S2		2277	88,39%	1248	88,93%
16	Limon S3		2863	85,41%	1413	87,46%
17	Limon S4		4631	76,39%	3074	72,73%
18	Limon S5		2552	86,99%	1488	86,80%
19	Limon S6		4577	76,67%	2700	76,05%
20	Limon S7		2773	85,87%	1485	86,83%
Average :			7,402,45	62,27%	4,500,20	60,08%

[Shreeshrii]

Acc_LimonS1S2F1F2R1R2Unicode

Engine: Fine Tune For Limon S1, S2, F1, F2, R1 & R2 Unicode

Net_spec value “[1,48,0,1 Ct3,3,16 Mp3,3 Lfys64 Lfx96 Lrx96 Lfx384 O1c1]”

No.	Font Name	Size	Character Accuracy		Cluster Accuracy
			Errors	Accuracy	Misrecognized	Accuracy
1	Khmer OS	12	1818	90,73%	900	92,02%
2	Khmer OS Battambang		1058	94,61%	476	95,78%
3	Khmer OS Bokor		2746	86,00%	1599	85,81%
4	Khmer OS Content		844	95,70%	353	96,87%
5	Khmer OS Fasthand		5063	74,19%	2296	79,63%
6	Khmer OS Freehand		2370	87,92%	674	94,02%
7	Khmer OS Metal Chrieng		3428	82,53%	1811	83,93%
8	Khmer OS Muol		946	95,18%	464	95,88%
9	Khmer OS Muol Light		1313	93,31%	720	93,61%
10	Khmer OS Muol Pali		6194	68,43%	3413	69,72%
11	Khmer OS Siemreap		1939	90,12%	1041	90,76%
12	Khmer OS System		2562	86,94%	1248	88,93%
13	Noto Serif Khmer Bold		2270	88,43%	1256	88,86%
14	Noto Serif Khmer Regular		2334	88,10%	1278	88,66%
Average :			2,491,79	87,30%	1,252,07	88,89%

No.	Font Name	Size	Character Accuracy		Cluster Accuracy
			Errors	Accuracy	Misrecognized	Accuracy
1	Limon F1	22	2599	86,75%	1696	84,95%
2	Limon F2		3896	80,14%	2661	76,39%
3	Limon F3		4178	78,70%	2786	75,28%
4	Limon F4		14807	24,52%	9374	16,84%
5	Limon F5		4859	75,23%	3282	70,88%
6	Limon F6		2576	86,87%	1622	85,61%
7	Limon F7		4115	79,02%	2145	80,97%
8	Limon F8		3580	81,75%	2166	80,78%
9	Limon R1		2144	89,07%	1358	87,95%
10	Limon R2		3681	81,24%	1626	85,57%
11	Limon R3		8224	58,08%	3482	69,11%
12	Limon R4		9265	52,77%	4504	60,04%
13	Limon R5		3120	84,10%	1995	82,30%
14	Limon S1		1262	93,57%	612	94,57%
15	Limon S2		1557	92,06%	756	93,29%
16	Limon S3		2100	89,30%	1070	90,51%
17	Limon S4		2515	87,18%	1670	85,18%
18	Limon S5		2069	89,45%	1122	90,05%
19	Limon S6		2885	85,29%	1473	86,93%
20	Limon S7		1852	90,56%	889	92,11%
Average :			4,064,20	79,28%	2,314,45	79,47%

[Shreeshrii]

@zdenop I suggest that you merge this PR. The accuracy reports show that there is over 10% improvement in the recognition on newly finetuned fonts without any appreciable loss in accuracy for the earlier list.

[zdenop]

It is s question if we want to mix google produce data with custom data...

[Shreeshrii]

@zdenop In that case may I suggest an additional repo, tessdata_contrib

[stweil]

Adding a new repository would be a possible solution. But wouldn't it be easier if people managed their own repositories with their tessdata contribution, and Tesseract could maintain a list with such repositories in the Wiki?

In any case such tessdata contributions should ideally document everything needed to reproduce the training process (fonts, images, ground truth, texts, scripts, documentation, ...).

[zdenop]

I agree with @stweil: this is the best solution. Google will not update files of contributor...

[Shreeshrii]

ok. I have added a page at https://github.com/tesseract-ocr/tesseract/wiki/Data-Files-Contributions ShreeDevi

…

____________________________________________________________ भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com

On Tue, Jun 5, 2018 at 3:25 PM, zdenop ***@***.***> wrote: I agree with @stweil <https://github.com/stweil>: this is the best solution. Google will not update files of contributor... — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#27 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AE2_o7BEjrv1YCIgbTtb10HXwtMsfbmmks5t5lV5gaJpZM4UPlyj> .

[stweil]

Thanks.

[zdenop]

Created https://github.com/tesseract-ocr/tessdata_contrib

[Shreeshrii]

@zdenop Will you be able to apply this PR to the new repo or does @phyrumsk need to submit a new PR for https://github.com/tesseract-ocr/tessdata_contrib?

[Shreeshrii]

tesseract-ocr/tessdata_contrib#1

adds this to tessdata_contrib repo.

[stweil]

It's also possible to apply the original pull request in tessdata_contrib. That would preserve the original author and time, so that is the preferred solution. @zdenop, I currently cannot do this because of missing rights for tessdata_contrib.

@phyrumsk, @Shreeshrii, it would be good to also have some documentation for each contributed model (who generated it, what is the name of the language or script, how can the training be reproduced, accuracy reports, ...). I suggest one documentation file per model (khmLimon.md) in markdown syntax.

[Shreeshrii]

t's also possible to apply the original pull request in tessdata_contrib. That would preserve the original author and time, so that is the preferred solution.

I would prefer it too :-)

[Shreeshrii]

https://github.com/tesseract-ocr/tesseract/wiki/Data-Files-Contributions

I have added some more links there. It is difficult to do accuracy estimates because there we do not have testdata for all languages. Also, my training is mainly proof of concept, I do not have system resources to do millions of iterations to get high accuracy (error rates below 1).

@stweil If you are able to run the training with more fonts/lines etc. I can provide the training text/scripts used.

[zdenop]

@stweil: try now.

[stweil]

Thank you. This pull request was now merged in tesseract-ocr/tessdata_contrib@05fa41a. It can be closed here (I don't have the rights to do that myself).