trained a new OCRopus model

jze · Oct 17, 2017 · 1656704 · 1656704
1 parent 8d30998
commit 1656704
Show file tree

Hide file tree

Showing 54 changed files with 88 additions and 15 deletions.
diff --git a/README.md b/README.md
@@ -11,19 +11,21 @@ In addition to the pyrnn.gz model to be used with `ocropus-rpred` I have also tr
 
 The ground truth images for this model were selected from these historic books:
 
-- Das astronomische Weltbild im Wandel der Zeit (1912) urn:nbn:at:AT-OOeLB-4113427
 - Topographische Chronik von Breslau  (1805)
-- Egger: Die christliche Mutter (1914)
-- Frapan: Bittersüß (1891)
-- Gartenlaube Heft 1, S. 8 (1897)
+- Menzel: Der praktische Maurer  (1847) urn:nbn:de:kobv:b4-25126-9
 - Kiel city directory (1888)
-- Was sollen wir kochen? (1915)  urn:nbn:at:AT-OOeLB-1184253
 - Köln city directory (1891)
-- Ludendorff: Kriegserinnerungen (1921)
-- Menzel: Der praktische Maurer  (1847)
-- Kreis-Kalender für den Kreis Plön (1909) urn:nbn:de:gbv:8:2-2533517
-- Preuschen: Yoshiwara (1920)
+- Gartenlaube Heft 1, S. 8 (1897)
+- Munzinger: Die Japaner (1898) urn:nbn:de:kobv:b4-200905194219
+- Frapan: Bittersüß (1891) urn:nbn:de:kobv:b4-200905191440
 - Schiller und Oberösterreich (1905) urn:nbn:at:AT-OOeLB-1099695
+- Kreis-Kalender für den Kreis Plön (1909) urn:nbn:de:gbv:8:2-2533517
+- Das astronomische Weltbild im Wandel der Zeit (1912) urn:nbn:at:AT-OOeLB-4113427
+- Egger: Die christliche Mutter (1914) urn:nbn:de:kobv:b4-30011-9
+- Was sollen wir kochen? (1915)  urn:nbn:at:AT-OOeLB-1184253
+- Der Große Krieg in Einzeldarstellungen - Die Winterschlacht in Masuren  (1918) urn:nbn:at:AT-OOeLB-1691328 
+- Preuschen: Yoshiwara (1920) urn:nbn:de:kobv:b4-30090-0
+- Ludendorff: Kriegserinnerungen (1921)
 - Ehrenbuch der Gefallenen Stuttgarts (1925)
 
 For some rare characters (Q, Y, Ä, Ö and Ü) I have generated some synthetic training data with OCRopus-linegen using the Walbaum Fraktur font and words from a German dictionary.
@@ -50,7 +52,7 @@ Using a language model you can easily replace the wrong characters. In German an
 
 ### s, tz, ch, ck and others
 
-Fraktur has a few special characters that usually do not occur in 'normal' text, such as the long-S or the tz ligature. This model (and the ground truth) translates them into 'normal' characters. Therefore a long-S will be interpreted as s and the ligatures will be interpreted as two characters.
+Fraktur has a few special characters that usually do not occur in 'normal' text, such as the long-S or the tz ligature. This model (and the ground truth) recognizsed the correct Unicode character LATIN SMALL LETTER LONG S. The tz-ligature is interpredted as two characters. 
 
 ![Wasser](images/long-s.png)
 
@@ -76,3 +78,9 @@ To train the model from scratch you can use this command:
 (Make sure that the OCRopus directory is included in PATH.) If you want to improve the existing model you can start with it:
 
 	ocropus-rtrain -c codec.txt -F 2000 --load fraktur.pyrnn.gz -o fraktur "training/*.bin.png"
+
+You can see the training progress in this image:
+
+![diagram showing the training progress](images/accuracy.png)
+
+If you are interested in the amount of ground-truth data  and the character accuracy for each book used for the model traning, take a look at the `overview-*` diagrams in the `images` folder.
diff --git a/accuracy.gnuplot b/accuracy.gnuplot
@@ -9,5 +9,5 @@ set xrange [0:]
 set yrange [0.95:1]
 
 plot 'errors.csv' u 1:(1-$2) with lines title 'testing',\
-'' u 1:(1-$3) title 'training',\
-'' u 1:(1-$3) title 'exclusive testing'
+'' u 1:(1-$3) with lines title 'training',\
+'' u 1:(1-$4) with lines title 'exclusive testing'
diff --git a/errors.csv b/errors.csv
@@ -0,0 +1,24 @@
+# iteration	testing	training	exclusive testing
+2000	0.142825631411	0.14939694083	0.124767801858
+4000	0.0636373662733	0.0596750420105	0.0597523219814
+6000	0.0379949266571	0.0323524182587	0.0266253869969
+8000	0.0291717216279	0.0224648860697	0.0238390092879
+10000	0.0292268666593	0.0223788576574	0.0213622291022
+12000	0.0259181647734	0.0186050779704	0.0167182662539
+14000	0.0226094628874	0.0150033551081	0.0164086687307
+16000	0.0268004852763	0.0188516927524	0.015479876161
+18000	0.0207896768501	0.0138677800655	0.0123839009288
+20000	0.0219477225102	0.0146936528237	0.0130030959752
+22000	0.0205690967244	0.0127952925253	0.012693498452
+24000	0.0207345318187	0.013615430056	0.0133126934985
+26000	0.0231609132017	0.0144183619043	0.0145510835913
+28000	0.0201830815044	0.012462649331	0.0139318885449
+30000	0.0195213411272	0.0111148708714	0.0123839009288
+32000	0.0184735855299	0.0108969322268	0.0120743034056
+34000	0.0180324252785	0.010346350388	0.0136222910217
+36000	0.0177015550899	0.0107764924496	0.0108359133127
+38000	0.0175361199956	0.00975562195674	0.00959752321981
+40000	0.0227748979817	0.0153417335299	0.0139318885449
+42000	0.0200727914415	0.0137989573356	0.0148606811146
+44000	0.0199073563472	0.0137588107432	0.012693498452
+46000	0.0149443035183	0.0100481185586	0.0130030959752
diff --git a/fraktur.pyrnn.gz b/fraktur.pyrnn.gz
diff --git a/images/accuracy.png b/images/accuracy.png
diff --git a/images/overview-characters.png b/images/overview-characters.png
diff --git a/images/overview-lines.png b/images/overview-lines.png
diff --git a/images/overview-testing.png b/images/overview-testing.png
diff --git a/images/overview-training.png b/images/overview-training.png
diff --git a/overview.csv b/overview.csv
@@ -0,0 +1,20 @@
+# book	training lines	training characters	training error	test lines	test characters	test error
+AC09345259	116	5119	0.00655391120507	6	190	0.00564971751412
+astro1912	75	4993	0.0117195823567	11	707	0.0345345345345
+breslau1805	21	920	0.0689252336449	2	62	0.107142857143
+egger	419	16510	0.00754962577286	48	1910	0.00964265456608
+frapan	504	25074	0.00335570469799	57	2852	0.00265957446809
+gartenlaube	69	4409	0.00896535013327	6	388	0.0301369863014
+grosserkrieg20	43	2316	0.0221811460259	18	1116	0.0191938579655
+kiel1888	373	13065	0.0188726305288	33	960	0.0368239355581
+kochen1915	119	6502	0.00713337757133	4	237	0.00448430493274
+koeln1891	44	3306	0.0175382653061	2	148	0.013986013986
+linegen	190	1702	0.0500747384155	3	20	0.214285714286
+ludendorf	351	23000	0.00475158001568	33	1891	0.00789622109419
+menzel	213	13479	0.0167192429022	20	1233	0.0128755364807
+munzinger	202	10576	0.010013744355	62	3286	0.0148593107809
+nn	19	861	0.0718711276332	3	196	0.0641711229947
+ploen1909	87	3478	0.0436667698978	10	438	0.0565110565111
+preuschen	19	953	0.010067114094	4	194	0.00549450549451
+schiller	364	20494	0.00508339952343	37	2076	0.00831600831601
+stuttgart	537	29414	0.00777550010603	41	2338	0.0120751341682
diff --git a/overview.sh b/overview.sh
@@ -2,6 +2,11 @@
 
 # Generate a table with an overview (number of lines, number of characters, recognition error) 
 # over the books used in the training process.
+#
+# After running this script you can use
+#    gnuplot overview.gnuplot
+# to genernate a few nice diagrams showing the number of ground-truth data and the character accuracy
+# for each of the books used in the model training process.
 
 ocropus-rpred -h > /dev/null 2>/dev/null
 if [ $? != 0 ]; then
@@ -10,9 +15,11 @@ if [ $? != 0 ]; then
 fi
 
 echo "Recognizing training data..."
-#ocropus-rpred -q -n -m `pwd`/fraktur.pyrnn.gz -Q6 "training/*.bin.png"
+ocropus-rpred -q -n -m `pwd`/fraktur.pyrnn.gz -Q6 "training/*.bin.png"
+ocropus-rpred -q -n -m `pwd`/fraktur.pyrnn.gz -Q6 "training/*.nrm.png"
 echo "Recognizing testing data..."
-#ocropus-rpred -q -n -m `pwd`/fraktur.pyrnn.gz -Q6 "testing/*.bin.png"
+ocropus-rpred -q -n -m `pwd`/fraktur.pyrnn.gz -Q6 "testing/*.bin.png"
+ocropus-rpred -q -n -m `pwd`/fraktur.pyrnn.gz -Q6 "testing/*.nrm.png"
 
 ls training/*.gt.txt | cut -c 10- | sed 's/_.*//' | uniq  | while read book; do echo -ne "$book\t"; LANG=  wc -m -l training/${book}_*.gt.txt |grep total | sed 's/[^ 0-9].*//' ; done > r1.txt
 ls training/*.gt.txt | cut -c 10- | sed 's/_.*//' | uniq  | while read book; do echo -ne "$book\t"; LANG=  wc -m -l testing/${book}_*.gt.txt |grep total | sed 's/[^ 0-9].*//' ; done > r2.txt

diff --git a/testing/grosserkrieg20_1535a2b29b0b81f6676a41ece9e70162.gt.txt b/testing/grosserkrieg20_1535a2b29b0b81f6676a41ece9e70162.gt.txt
@@ -1 +1 @@
-Durch die getroffenen. Vorſichtsmaßregeln war es gelungen,
+Durch die getroffenen Vorſichtsmaßregeln war es gelungen,
diff --git a/testing/grosserkrieg20_24116d1ac52738440f068cedb9df96f2.bin.png b/testing/grosserkrieg20_24116d1ac52738440f068cedb9df96f2.bin.png
diff --git a/testing/grosserkrieg20_24116d1ac52738440f068cedb9df96f2.gt.txt b/testing/grosserkrieg20_24116d1ac52738440f068cedb9df96f2.gt.txt
@@ -0,0 +1 @@
+heranzuwerfen. Tatſächlich gingen die Ruſſen bei Sztabin und
diff --git a/testing/grosserkrieg20_24116d1ac52738440f068cedb9df96f2.nrm.png b/testing/grosserkrieg20_24116d1ac52738440f068cedb9df96f2.nrm.png
diff --git a/testing/grosserkrieg20_4b433d29271382cbde1c753617d03468.bin.png b/testing/grosserkrieg20_4b433d29271382cbde1c753617d03468.bin.png
diff --git a/testing/grosserkrieg20_4b433d29271382cbde1c753617d03468.gt.txt b/testing/grosserkrieg20_4b433d29271382cbde1c753617d03468.gt.txt
@@ -0,0 +1 @@
+hinter dem Njemen lebhafter, doch hatte auch hier die deutſche
diff --git a/testing/grosserkrieg20_4b433d29271382cbde1c753617d03468.nrm.png b/testing/grosserkrieg20_4b433d29271382cbde1c753617d03468.nrm.png
diff --git a/testing/grosserkrieg20_4ecb4a95fa59eb20b937239686b0cbfa.bin.png b/testing/grosserkrieg20_4ecb4a95fa59eb20b937239686b0cbfa.bin.png
diff --git a/testing/grosserkrieg20_4ecb4a95fa59eb20b937239686b0cbfa.gt.txt b/testing/grosserkrieg20_4ecb4a95fa59eb20b937239686b0cbfa.gt.txt
@@ -0,0 +1 @@
+ſal überlaſſen hatte, als er keinen Ausweg mehr ſah, mag von
diff --git a/testing/grosserkrieg20_4ecb4a95fa59eb20b937239686b0cbfa.nrm.png b/testing/grosserkrieg20_4ecb4a95fa59eb20b937239686b0cbfa.nrm.png
diff --git a/testing/grosserkrieg20_503c040a919be30042b9682f746a0683.bin.png b/testing/grosserkrieg20_503c040a919be30042b9682f746a0683.bin.png
diff --git a/testing/grosserkrieg20_503c040a919be30042b9682f746a0683.gt.txt b/testing/grosserkrieg20_503c040a919be30042b9682f746a0683.gt.txt
@@ -0,0 +1 @@
+Grodno aus die ruſſiſche Heeresleitung beſchworen haben, alles
diff --git a/testing/grosserkrieg20_503c040a919be30042b9682f746a0683.nrm.png b/testing/grosserkrieg20_503c040a919be30042b9682f746a0683.nrm.png
diff --git a/testing/grosserkrieg20_57aefadbfd0855609a269e513833e27d.bin.png b/testing/grosserkrieg20_57aefadbfd0855609a269e513833e27d.bin.png
diff --git a/testing/grosserkrieg20_57aefadbfd0855609a269e513833e27d.gt.txt b/testing/grosserkrieg20_57aefadbfd0855609a269e513833e27d.gt.txt
@@ -0,0 +1 @@
+General Siewers, der vor Tagen bereits ſeine Armee ihrem Schick-
diff --git a/testing/grosserkrieg20_57aefadbfd0855609a269e513833e27d.nrm.png b/testing/grosserkrieg20_57aefadbfd0855609a269e513833e27d.nrm.png
diff --git a/testing/grosserkrieg20_6b2e38d9cbec46fefd2f3dba9f66b267.bin.png b/testing/grosserkrieg20_6b2e38d9cbec46fefd2f3dba9f66b267.bin.png
diff --git a/testing/grosserkrieg20_6b2e38d9cbec46fefd2f3dba9f66b267.gt.txt b/testing/grosserkrieg20_6b2e38d9cbec46fefd2f3dba9f66b267.gt.txt
@@ -0,0 +1 @@
+Führung bereits genügend Kräfte zur Abwehr eines etwaigen
diff --git a/testing/grosserkrieg20_6b2e38d9cbec46fefd2f3dba9f66b267.nrm.png b/testing/grosserkrieg20_6b2e38d9cbec46fefd2f3dba9f66b267.nrm.png
diff --git a/testing/grosserkrieg20_8aa4713109e0bac645323c46ae41bee1.bin.png b/testing/grosserkrieg20_8aa4713109e0bac645323c46ae41bee1.bin.png
diff --git a/testing/grosserkrieg20_8aa4713109e0bac645323c46ae41bee1.gt.txt b/testing/grosserkrieg20_8aa4713109e0bac645323c46ae41bee1.gt.txt
@@ -0,0 +1 @@
+(ſiehe Skizze 2). Auch in Gegend der Feſtung Olita wurde es
diff --git a/testing/grosserkrieg20_8aa4713109e0bac645323c46ae41bee1.nrm.png b/testing/grosserkrieg20_8aa4713109e0bac645323c46ae41bee1.nrm.png
diff --git a/testing/grosserkrieg20_9e218636b314d1e43a5cf46efe0afbb9.bin.png b/testing/grosserkrieg20_9e218636b314d1e43a5cf46efe0afbb9.bin.png
diff --git a/testing/grosserkrieg20_9e218636b314d1e43a5cf46efe0afbb9.gt.txt b/testing/grosserkrieg20_9e218636b314d1e43a5cf46efe0afbb9.gt.txt
@@ -0,0 +1 @@
+Geſpenſt der Schlacht von Tannenberg haben heraufſteigen ſehen.
diff --git a/testing/grosserkrieg20_9e218636b314d1e43a5cf46efe0afbb9.nrm.png b/testing/grosserkrieg20_9e218636b314d1e43a5cf46efe0afbb9.nrm.png
diff --git a/testing/grosserkrieg20_a3d1d3d41374480513a8f09ae73582a9.bin.png b/testing/grosserkrieg20_a3d1d3d41374480513a8f09ae73582a9.bin.png
diff --git a/testing/grosserkrieg20_a3d1d3d41374480513a8f09ae73582a9.gt.txt b/testing/grosserkrieg20_a3d1d3d41374480513a8f09ae73582a9.gt.txt
@@ -0,0 +1 @@
+und in blutigem Ringen unter ſtarken Verluſten zurükgeworfen
diff --git a/testing/grosserkrieg20_a3d1d3d41374480513a8f09ae73582a9.nrm.png b/testing/grosserkrieg20_a3d1d3d41374480513a8f09ae73582a9.nrm.png
diff --git a/testing/grosserkrieg20_b14f5ad16f9d3222402efac09ff418cc.bin.png b/testing/grosserkrieg20_b14f5ad16f9d3222402efac09ff418cc.bin.png
diff --git a/testing/grosserkrieg20_b14f5ad16f9d3222402efac09ff418cc.gt.txt b/testing/grosserkrieg20_b14f5ad16f9d3222402efac09ff418cc.gt.txt
@@ -0,0 +1 @@
+das XL. Reſervekorps und die 4. Kavalleriediviſion aufgehalten
diff --git a/testing/grosserkrieg20_b14f5ad16f9d3222402efac09ff418cc.nrm.png b/testing/grosserkrieg20_b14f5ad16f9d3222402efac09ff418cc.nrm.png
diff --git a/testing/grosserkrieg20_c6a9cda10bb7c4b76b1fa23a22ffc5ac.bin.png b/testing/grosserkrieg20_c6a9cda10bb7c4b76b1fa23a22ffc5ac.bin.png
diff --git a/testing/grosserkrieg20_c6a9cda10bb7c4b76b1fa23a22ffc5ac.gt.txt b/testing/grosserkrieg20_c6a9cda10bb7c4b76b1fa23a22ffc5ac.gt.txt
@@ -0,0 +1 @@
+feindlichen Entlaſtungsvorſtoßes gegen den Rücken der Ein-
diff --git a/testing/grosserkrieg20_c6a9cda10bb7c4b76b1fa23a22ffc5ac.nrm.png b/testing/grosserkrieg20_c6a9cda10bb7c4b76b1fa23a22ffc5ac.nrm.png
diff --git a/testing/grosserkrieg20_c7b286f42a5cb98e42fbe2f340467fce.bin.png b/testing/grosserkrieg20_c7b286f42a5cb98e42fbe2f340467fce.bin.png
diff --git a/testing/grosserkrieg20_c7b286f42a5cb98e42fbe2f340467fce.gt.txt b/testing/grosserkrieg20_c7b286f42a5cb98e42fbe2f340467fce.gt.txt
@@ -0,0 +1 @@
+öſtlich zum Angriff über den Bobr vor, wurden aber hier durch
diff --git a/testing/grosserkrieg20_c7b286f42a5cb98e42fbe2f340467fce.nrm.png b/testing/grosserkrieg20_c7b286f42a5cb98e42fbe2f340467fce.nrm.png
diff --git a/testing/grosserkrieg20_d72cdca2698e7afc285b87adf5fba0da.bin.png b/testing/grosserkrieg20_d72cdca2698e7afc285b87adf5fba0da.bin.png
diff --git a/testing/grosserkrieg20_d72cdca2698e7afc285b87adf5fba0da.gt.txt b/testing/grosserkrieg20_d72cdca2698e7afc285b87adf5fba0da.gt.txt
@@ -0,0 +1 @@
+irgend Erreichbare zum Einſaz ſeiner eingeſchloſſenen Diviſionen
diff --git a/testing/grosserkrieg20_d72cdca2698e7afc285b87adf5fba0da.nrm.png b/testing/grosserkrieg20_d72cdca2698e7afc285b87adf5fba0da.nrm.png
diff --git a/testing/grosserkrieg20_f04c7a10172fb5604ab8c0d92a3e0f72.bin.png b/testing/grosserkrieg20_f04c7a10172fb5604ab8c0d92a3e0f72.bin.png
diff --git a/testing/grosserkrieg20_f04c7a10172fb5604ab8c0d92a3e0f72.gt.txt b/testing/grosserkrieg20_f04c7a10172fb5604ab8c0d92a3e0f72.gt.txt
@@ -0,0 +1 @@
+ſchließungsarmee bereitgeſtellt. In weiſer Vorausſicht war die
diff --git a/testing/grosserkrieg20_f04c7a10172fb5604ab8c0d92a3e0f72.nrm.png b/testing/grosserkrieg20_f04c7a10172fb5604ab8c0d92a3e0f72.nrm.png
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		Durch die getroffenen. Vorſichtsmaßregeln war es gelungen,
		Durch die getroffenen Vorſichtsmaßregeln war es gelungen,
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		heranzuwerfen. Tatſächlich gingen die Ruſſen bei Sztabin und
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		hinter dem Njemen lebhafter, doch hatte auch hier die deutſche
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		ſal überlaſſen hatte, als er keinen Ausweg mehr ſah, mag von
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Grodno aus die ruſſiſche Heeresleitung beſchworen haben, alles
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		General Siewers, der vor Tagen bereits ſeine Armee ihrem Schick-
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Führung bereits genügend Kräfte zur Abwehr eines etwaigen
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		(ſiehe Skizze 2). Auch in Gegend der Feſtung Olita wurde es
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Geſpenſt der Schlacht von Tannenberg haben heraufſteigen ſehen.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		und in blutigem Ringen unter ſtarken Verluſten zurükgeworfen