本篇文章是旧版本的训练教程，适应于新版本的训练流程请查看：http://www.jianshu.com/p/7a2c40dd6560
题库特殊字符语言包训练流程
题库中会出现很多Tesseract无法识别的特殊字符，例如：≤ ≥ ×等等。本教程旨在训练新的语言包改善这些特殊字符的识别效果。训练思路为，利用Tesseract认识的字符替代这些特殊字符(与特殊字符外形相近，比如，‘,’替代‘，’，或者题库中没有的字符替换原有字符)。
准备训练依赖项
git clone https://github.com/tesseract-ocr/tesseract.git
git clone https://github.com/tesseract-ocr/langdata.git
git clone https://github.com/tesseract-ocr/tessdata.git
cp ./tessdata/eng.traineddata ./tesseract/tessdata
cp ./tessdata/chi_sim.traineddata ./tesseract/tessdata
下载tesseract包后需要本地安装，可看前面发布的安装教程：Tess4.0 windows编译与使用，tesseract linux 安装与编译。
添加新的字体
打开./training/language-specific.sh文件，在CHI_SIM_FONTS栏目下添加Times New Roman字体内容，添加后内容如下：
CHI_SIM_FONTS=( \
"AR PL UKai CN" \
"AR PL UMing Patched Light"
"AR PL UKai TW" \
"AR PL UMing TW MBE Light" \
"AR PL UKai Patched" \
"Arial Unicode MS" \
"Arial Unicode MS Bold" \
"WenQuanYi Zen Hei Medium" \
"Times New Roman, Bold" \
"Times New Roman, Bold Italic" \
"Times New Roman, Italic" \
"Times New Roman," \
)
外形相近的字符替换
应用Times New Roman字体中现有的字符替换外形相近的字符，见以下是替换代码：
text_line.replace('（','(').replace('）',')').replace('﹣','-').replace('．','.')\
.replace('：', ':')
替换其他特殊字符
训练流程
第一步
将文件training/tesstrain.sh复制备份training/tesstrain0.sh，并在tesstrain0.sh中将54行以下的内容添加注释，注释后的代码为:
source "$(dirname $0)/tesstrain_utils.sh"
ARGV=("$@")
parse_flags
echo $WORKSPACE_DIR
mkdir -p ${TRAINING_DIR}
# tlog "\n=== Starting training for language '${LANG_CODE}'"
#
# source "$(dirname $0)/language-specific.sh"
# set_lang_specific_parameters ${LANG_CODE}
#
# initialize_fontconfig
#
# phase_I_generate_image 8
# phase_UP_generate_unicharset
# phase_D_generate_dawg
# if ((LINEDATA)); then
# phase_E_extract_features "lstm.train" 8 "lstmf"
# make__lstmdata
# else
# phase_E_extract_features "box.train" 8 "tr"
# phase_C_cluster_prototypes "${TRAINING_DIR}/${LANG_CODE}.normproto"
# if [[ "${ENABLE_SHAPE_CLUSTERING}" == "y" ]]; then
# phase_S_cluster_shapes
# fi
# phase_M_cluster_microfeatures
# phase_B_generate_ambiguities
# make__traineddata
# fi
#
# tlog "\nCompleted training for language '${LANG_CODE}'\n"
执行以下命令：
training/tesstrain0.sh --fonts_dir /usr/share/fonts \
--training_text ../training_data/input_data0.txt \
--langdata_dir ../langdata --tessdata_dir ./tessdata \
--lang chi_sim --linedata_only --noextract_font_properties \
--exposures "0" --fontlist "SIMSUN" \
--output_dir ~/tesstutorial/chitest
第二步
按照以下步骤对shell文件进行修改：
- 查看系统/tmp目录下的文件夹，找到格式为'tmp.*'的文件夹，并将此文件夹名进行复制。
- 将文件training/tesstrain_utils.sh复制备份training/tesstrain_utils0.sh，并将tesstrain_utils0.sh中的WORKSPACE_DIR变量名赋值为'tmp.*'(/tmp目录下相应的文件夹名称)
- 将training/tesstrain0.sh文件中source "$(dirname $0)/tesstrain_utils.sh"语句改成source "$(dirname $0)/tesstrain_utils0.sh"
- 将training/tesstrain0.sh文件的注释内容进行更改，更改后的内容如下：
source "$(dirname $0)/tesstrain_utils0.sh"
ARGV=("$@")
parse_flags
echo $WORKSPACE_DIR
mkdir -p ${TRAINING_DIR}
tlog "\n=== Starting training for language '${LANG_CODE}'"
source "$(dirname $0)/language-specific.sh"
set_lang_specific_parameters ${LANG_CODE}
initialize_fontconfig
phase_I_generate_image 8
# phase_UP_generate_unicharset
# phase_D_generate_dawg
# if ((LINEDATA)); then
# phase_E_extract_features "lstm.train" 8 "lstmf"
# make__lstmdata
# else
# phase_E_extract_features "box.train" 8 "tr"
# phase_C_cluster_prototypes "${TRAINING_DIR}/${LANG_CODE}.normproto"
# if [[ "${ENABLE_SHAPE_CLUSTERING}" == "y" ]]; then
# phase_S_cluster_shapes
# fi
# phase_M_cluster_microfeatures
# phase_B_generate_ambiguities
# make__traineddata
# fi
#
# tlog "\nCompleted training for language '${LANG_CODE}'\n"
执行以下命令：
training/tesstrain0.sh --fonts_dir /usr/share/fonts \
--training_text ../training_data/input_data0.txt \
--langdata_dir ../langdata --tessdata_dir ./tessdata \
--lang chi_sim --linedata_only --noextract_font_properties \
--exposures "0" --fontlist "SIMSUN" \
--output_dir ~/tesstutorial/chitest
第三步
在系统桌面进入"/tmp/tmp.J8JYbdYYrv/chi_sim"目录，修改后缀名为.box文件，将此文件中的特殊字符修改为想要替换的其他字符，注意：每个字符的替换需要保持一致哦。
例如：本文例将所有的'≤'符号替换成"小"。
第四步
修改training/tesstrain0.sh文件中的注释内容，修改后如下：
source "$(dirname $0)/tesstrain_utils0.sh"
ARGV=("$@")
parse_flags
# mkdir -p ${TRAINING_DIR}
# tlog "\n=== Starting training for language '${LANG_CODE}'"
#
# source "$(dirname $0)/language-specific.sh"
# set_lang_specific_parameters ${LANG_CODE}
#
# initialize_fontconfig
#
# phase_I_generate_image 8
echo $TRAINING_DIR
phase_UP_generate_unicharset
phase_D_generate_dawg
if ((LINEDATA)); then
phase_E_extract_features "lstm.train" 8 "lstmf"
make__lstmdata
else
phase_E_extract_features "box.train" 8 "tr"
phase_C_cluster_prototypes "${TRAINING_DIR}/${LANG_CODE}.normproto"
if [[ "${ENABLE_SHAPE_CLUSTERING}" == "y" ]]; then
phase_S_cluster_shapes
fi
phase_M_cluster_microfeatures
phase_B_generate_ambiguities
make__traineddata
fi
tlog "\nCompleted training for language '${LANG_CODE}'\n"
执行以下命令：
training/tesstrain0.sh --fonts_dir /usr/share/fonts \
--training_text ../training_data/input_data0.txt \
--langdata_dir ../langdata --tessdata_dir ./tessdata \
--lang chi_sim --linedata_only --noextract_font_properties \
--exposures "0" --fontlist "SIMSUN" \
--output_dir ~/tesstutorial/chitest
第五步
按步骤执行以下命令:
- mkdir -p ~/tesstutorial/chituned_from_chisim
- combine_tessdata -e ../tessdata/chi_sim.traineddata ~/tesstutorial/chituned_from_chisim/chi_sim.lstm
- lstmtraining --model_output ~/tesstutorial/chituned_from_chisim/chituned --continue_from ~/tesstutorial/chituned_from_chisim/chi_sim.lstm --train_listfile ~/tesstutorial/chitest/chi_sim.training_files.txt --eval_listfile ~/tesstutorial/chitest/chi_sim.training_files.txt --target_error_rate 0.01
- lstmtraining --model_output ~/tesstutorial/chituned_from_chisim/chituned.lstm
--continue_from ~/tesstutorial/chituned_from_chisim/chituned_checkpoint
--stop_training - combine_tessdata -o ./tessdata/chi_sim.traineddata
~/tesstutorial/chituned_from_chisim/chituned.lstm
~/tesstutorial/chitest/chi_sim.lstm-number-dawg
~/tesstutorial/chitest/chi_sim.lstm-punc-dawg
~/tesstutorial/chitest/chi_sim.lstm-word-dawg
新生成的traineddata在tesseract/tessdata目录下。
网友评论