00001 // Copyright 2010 Google Inc. All Rights Reserved. 00002 // Author: rays@google.com (Ray Smith) 00004 // File: mastertrainer.h 00005 // Description: Trainer to build the MasterClassifier. 00006 // Author: Ray Smith 00007 // Created: Wed Nov 03 18:07:01 PDT 2010 00008 // 00009 // (C) Copyright 2010, Google Inc. 00010 // Licensed under the Apache License, Version 2.0 (the "License"); 00011 // you may not use this file except in compliance with the License. 00012 // You may obtain a copy of the License at 00013 // http://www.apache.org/licenses/LICENSE-2.0 00014 // Unless required by applicable law or agreed to in writing, software 00015 // distributed under the License is distributed on an "AS IS" BASIS, 00016 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 00017 // See the License for the specific language governing permissions and 00018 // limitations under the License. 00019 // 00021 00022 #ifndef TESSERACT_TRAINING_MASTERTRAINER_H__ 00023 #define TESSERACT_TRAINING_MASTERTRAINER_H__ 00024 00028 #include "classify.h" 00029 #include "cluster.h" 00030 #include "intfx.h" 00031 #include "elst.h" 00032 #include "featdefs.h" 00033 #include "fontinfo.h" 00034 #include "indexmapbidi.h" 00035 #include "intfeaturespace.h" 00036 #include "intfeaturemap.h" 00037 #include "intmatcher.h" 00038 #include "params.h" 00039 #include "shapetable.h" 00040 #include "trainingsample.h" 00041 #include "trainingsampleset.h" 00042 #include "unicharset.h" 00043 00044 namespace tesseract { 00045 00046 class ShapeClassifier; 00047 00048 // Simple struct to hold the distance between two shapes during clustering. 00049 struct ShapeDist { 00050 ShapeDist() : shape1(0), shape2(0), distance(0.0f) {} 00051 ShapeDist(int s1, int s2, float dist) 00052 : shape1(s1), shape2(s2), distance(dist) {} 00053 00054 // Sort operator to sort in ascending order of distance. 00055 bool operator<(const ShapeDist& other) const { 00056 return distance < other.distance; 00057 } 00058 00059 int shape1; 00060 int shape2; 00061 float distance; 00062 }; 00063 00064 // Class to encapsulate training processes that use the TrainingSampleSet. 00065 // Initially supports shape clustering and mftrainining. 00066 // Other important features of the MasterTrainer are conditioning the data 00067 // by outlier elimination, replication with perturbation, and serialization. 00068 class MasterTrainer { 00069 public: 00070 MasterTrainer(NormalizationMode norm_mode, bool shape_analysis, 00071 bool replicate_samples, int debug_level); 00072 ~MasterTrainer(); 00073 00074 // Writes to the given file. Returns false in case of error. 00075 bool Serialize(FILE* fp) const; 00076 // Reads from the given file. Returns false in case of error. 00077 // If swap is true, assumes a big/little-endian swap is needed. 00078 bool DeSerialize(bool swap, FILE* fp); 00079 00080 // Loads an initial unicharset, or sets one up if the file cannot be read. 00081 void LoadUnicharset(const char* filename); 00082 00083 // Sets the feature space definition. 00084 void SetFeatureSpace(const IntFeatureSpace& fs) { 00085 feature_space_ = fs; 00086 feature_map_.Init(fs); 00087 } 00088 00089 // Reads the samples and their features from the given file, 00090 // adding them to the trainer with the font_id from the content of the file. 00091 // If verification, then these are verification samples, not training. 00092 void ReadTrainingSamples(FILE *fp, 00093 const FEATURE_DEFS_STRUCT& feature_defs, 00094 bool verification); 00095 00096 // Adds the given single sample to the trainer, setting the classid 00097 // appropriately from the given unichar_str. 00098 void AddSample(bool verification, const char* unichar_str, 00099 TrainingSample* sample); 00100 00101 // Loads all pages from the given tif filename and append to page_images_. 00102 // Must be called after ReadTrainingSamples, as the current number of images 00103 // is used as an offset for page numbers in the samples. 00104 void LoadPageImages(const char* filename); 00105 00106 // Cleans up the samples after initial load from the tr files, and prior to 00107 // saving the MasterTrainer: 00108 // Remaps fragmented chars if running shape anaylsis. 00109 // Sets up the samples appropriately for class/fontwise access. 00110 // Deletes outlier samples. 00111 void PostLoadCleanup(); 00112 00113 // Gets the samples ready for training. Use after both 00114 // ReadTrainingSamples+PostLoadCleanup or DeSerialize. 00115 // Re-indexes the features and computes canonical and cloud features. 00116 void PreTrainingSetup(); 00117 00118 // Sets up the master_shapes_ table, which tells which fonts should stay 00119 // together until they get to a leaf node classifier. 00120 void SetupMasterShapes(); 00121 00122 // Adds the junk_samples_ to the main samples_ set. Junk samples are initially 00123 // fragments and n-grams (all incorrectly segmented characters). 00124 // Various training functions may result in incorrectly segmented characters 00125 // being added to the unicharset of the main samples, perhaps because they 00126 // form a "radical" decomposition of some (Indic) grapheme, or because they 00127 // just look the same as a real character (like rn/m) 00128 // This function moves all the junk samples, to the main samples_ set, but 00129 // desirable junk, being any sample for which the unichar already exists in 00130 // the samples_ unicharset gets the unichar-ids re-indexed to match, but 00131 // anything else gets re-marked as unichar_id 0 (space character) to identify 00132 // it as junk to the error counter. 00133 void IncludeJunk(); 00134 00135 // Replicates the samples and perturbs them if the enable_replication_ flag 00136 // is set. MUST be used after the last call to OrganizeByFontAndClass on 00137 // the training samples, ie after IncludeJunk if it is going to be used, as 00138 // OrganizeByFontAndClass will eat the replicated samples into the regular 00139 // samples. 00140 void ReplicateAndRandomizeSamplesIfRequired(); 00141 00142 // Loads the basic font properties file into fontinfo_table_. 00143 // Returns false on failure. 00144 bool LoadFontInfo(const char* filename); 00145 00146 // Loads the xheight font properties file into xheights_. 00147 // Returns false on failure. 00148 bool LoadXHeights(const char* filename); 00149 00150 // Reads spacing stats from filename and adds them to fontinfo_table. 00151 // Returns false on failure. 00152 bool AddSpacingInfo(const char *filename); 00153 00154 // Returns the font id corresponding to the given font name. 00155 // Returns -1 if the font cannot be found. 00156 int GetFontInfoId(const char* font_name); 00157 // Returns the font_id of the closest matching font name to the given 00158 // filename. It is assumed that a substring of the filename will match 00159 // one of the fonts. If more than one is matched, the longest is returned. 00160 int GetBestMatchingFontInfoId(const char* filename); 00161 00162 // Sets up a flat shapetable with one shape per class/font combination. 00163 void SetupFlatShapeTable(ShapeTable* shape_table); 00164 00165 // Sets up a Clusterer for mftraining on a single shape_id. 00166 // Call FreeClusterer on the return value after use. 00167 CLUSTERER* SetupForClustering(const ShapeTable& shape_table, 00168 const FEATURE_DEFS_STRUCT& feature_defs, 00169 int shape_id, int* num_samples); 00170 00171 // Writes the given float_classes (produced by SetupForFloat2Int) as inttemp 00172 // to the given inttemp_file, and the corresponding pffmtable. 00173 // The unicharset is the original encoding of graphemes, and shape_set should 00174 // match the size of the shape_table, and may possibly be totally fake. 00175 void WriteInttempAndPFFMTable(const UNICHARSET& unicharset, 00176 const UNICHARSET& shape_set, 00177 const ShapeTable& shape_table, 00178 CLASS_STRUCT* float_classes, 00179 const char* inttemp_file, 00180 const char* pffmtable_file); 00181 00182 const UNICHARSET& unicharset() const { 00183 return samples_.unicharset(); 00184 } 00185 TrainingSampleSet* GetSamples() { 00186 return &samples_; 00187 } 00188 const ShapeTable& master_shapes() const { 00189 return master_shapes_; 00190 } 00191 00192 // Generates debug output relating to the canonical distance between the 00193 // two given UTF8 grapheme strings. 00194 void DebugCanonical(const char* unichar_str1, const char* unichar_str2); 00195 // Debugging for cloud/canonical features. 00196 // Displays a Features window containing: 00197 // If unichar_str2 is in the unicharset, and canonical_font is non-negative, 00198 // displays the canonical features of the char/font combination in red. 00199 // If unichar_str1 is in the unicharset, and cloud_font is non-negative, 00200 // displays the cloud feature of the char/font combination in green. 00201 // The canonical features are drawn first to show which ones have no 00202 // matches in the cloud features. 00203 // Until the features window is destroyed, each click in the features window 00204 // will display the samples that have that feature in a separate window. 00205 void DisplaySamples(const char* unichar_str1, int cloud_font, 00206 const char* unichar_str2, int canonical_font); 00207 00208 // Tests the given test_classifier on the internal samples. 00209 // See TestClassifier for details. 00210 void TestClassifierOnSamples(int report_level, 00211 bool replicate_samples, 00212 ShapeClassifier* test_classifier, 00213 STRING* report_string); 00214 // Tests the given test_classifier on the given samples 00215 // report_levels: 00216 // 0 = no output. 00217 // 1 = bottom-line error rate. 00218 // 2 = bottom-line error rate + time. 00219 // 3 = font-level error rate + time. 00220 // 4 = list of all errors + short classifier debug output on 16 errors. 00221 // 5 = list of all errors + short classifier debug output on 25 errors. 00222 // If replicate_samples is true, then the test is run on an extended test 00223 // sample including replicated and systematically perturbed samples. 00224 // If report_string is non-NULL, a summary of the results for each font 00225 // is appended to the report_string. 00226 double TestClassifier(int report_level, 00227 bool replicate_samples, 00228 TrainingSampleSet* samples, 00229 ShapeClassifier* test_classifier, 00230 STRING* report_string); 00231 00232 // Returns the average (in some sense) distance between the two given 00233 // shapes, which may contain multiple fonts and/or unichars. 00234 // This function is public to facilitate testing. 00235 float ShapeDistance(const ShapeTable& shapes, int s1, int s2); 00236 00237 private: 00238 // Replaces samples that are always fragmented with the corresponding 00239 // fragment samples. 00240 void ReplaceFragmentedSamples(); 00241 00242 // Runs a hierarchical agglomerative clustering to merge shapes in the given 00243 // shape_table, while satisfying the given constraints: 00244 // * End with at least min_shapes left in shape_table, 00245 // * No shape shall have more than max_shape_unichars in it, 00246 // * Don't merge shapes where the distance between them exceeds max_dist. 00247 void ClusterShapes(int min_shapes, int max_shape_unichars, 00248 float max_dist, ShapeTable* shape_table); 00249 00250 private: 00251 NormalizationMode norm_mode_; 00252 // Character set we are training for. 00253 UNICHARSET unicharset_; 00254 // Original feature space. Subspace mapping is contained in feature_map_. 00255 IntFeatureSpace feature_space_; 00256 TrainingSampleSet samples_; 00257 TrainingSampleSet junk_samples_; 00258 TrainingSampleSet verify_samples_; 00259 // Master shape table defines what fonts stay together until the leaves. 00260 ShapeTable master_shapes_; 00261 // Flat shape table has each unichar/font id pair in a separate shape. 00262 ShapeTable flat_shapes_; 00263 // Font metrics gathered from multiple files. 00264 UnicityTable<FontInfo> fontinfo_table_; 00265 // Array of xheights indexed by font ids in fontinfo_table_; 00266 GenericVector<int> xheights_; 00267 00268 // Non-serialized data initialized by other means or used temporarily 00269 // during loading of training samples. 00270 // Number of different class labels in unicharset_. 00271 int charsetsize_; 00272 // Flag to indicate that we are running shape analysis and need fragments 00273 // fixing. 00274 bool enable_shape_anaylsis_; 00275 // Flag to indicate that sample replication is required. 00276 bool enable_replication_; 00277 // Flag to indicate that junk should be included in samples_. 00278 bool include_junk_; 00279 // Array of classids of fragments that replace the correctly segmented chars. 00280 int* fragments_; 00281 // Classid of previous correctly segmented sample that was added. 00282 int prev_unichar_id_; 00283 // Debug output control. 00284 int debug_level_; 00285 // Feature map used to construct reduced feature spaces for compact 00286 // classifiers. 00287 IntFeatureMap feature_map_; 00288 // Vector of Pix pointers used for classifiers that need the image. 00289 // Indexed by page_num_ in the samples. 00290 // These images are owned by the trainer and need to be pixDestroyed. 00291 GenericVector<Pix*> page_images_; 00292 }; 00293 00294 } // namespace tesseract. 00295 00296 #endif