tesseract  5.0.0-alpha-619-ge9db
blamer.h File Reference
#include <cstdint>
#include <cstring>
#include "boxword.h"
#include <tesseract/genericvector.h>
#include "params_training_featdef.h"
#include "ratngs.h"
#include "rect.h"
#include <tesseract/strngs.h>
#include "tprintf.h"
#include <tesseract/unichar.h>

Go to the source code of this file.

Classes

struct  BlamerBundle
 

Namespaces

 tesseract
 

Enumerations

enum  IncorrectResultReason {
  IRR_CORRECT, IRR_CLASSIFIER, IRR_CHOPPER, IRR_CLASS_LM_TRADEOFF,
  IRR_PAGE_LAYOUT, IRR_SEGSEARCH_HEUR, IRR_SEGSEARCH_PP, IRR_CLASS_OLD_LM_TRADEOFF,
  IRR_ADAPTION, IRR_NO_TRUTH_SPLIT, IRR_NO_TRUTH, IRR_UNKNOWN,
  IRR_NUM_REASONS
}
 

Enumeration Type Documentation

◆ IncorrectResultReason

Enumerator
IRR_CORRECT 
IRR_CLASSIFIER 
IRR_CHOPPER 
IRR_CLASS_LM_TRADEOFF 
IRR_PAGE_LAYOUT 
IRR_SEGSEARCH_HEUR 
IRR_SEGSEARCH_PP 
IRR_CLASS_OLD_LM_TRADEOFF 
IRR_ADAPTION 
IRR_NO_TRUTH_SPLIT 
IRR_NO_TRUTH 
IRR_UNKNOWN 
IRR_NUM_REASONS 

Definition at line 52 of file blamer.h.

52  {
53  // The text recorded in best choice == truth text
55  // Either: Top choice is incorrect and is a dictionary word (language model
56  // is unlikely to help correct such errors, so blame the classifier).
57  // Or: the correct unichar was not included in shortlist produced by the
58  // classifier at all.
60  // Chopper have not found one or more splits that correspond to the correct
61  // character bounding boxes recorded in BlamerBundle::truth_word.
63  // Classifier did include correct unichars for each blob in the correct
64  // segmentation, however its rating could have been too bad to allow the
65  // language model to pull out the correct choice. On the other hand the
66  // strength of the language model might have been too weak to favor the
67  // correct answer, this we call this case a classifier-language model
68  // tradeoff error.
70  // Page layout failed to produce the correct bounding box. Blame page layout
71  // if the truth was not found for the word, which implies that the bounding
72  // box of the word was incorrect (no truth word had a similar bounding box).
74  // SegSearch heuristic prevented one or more blobs from the correct
75  // segmentation state to be classified (e.g. the blob was too wide).
77  // The correct segmentaiton state was not explored because of poor SegSearch
78  // pain point prioritization. We blame SegSearch pain point prioritization
79  // if the best rating of a choice constructed from correct segmentation is
80  // better than that of the best choice (i.e. if we got to explore the correct
81  // segmentation state, language model would have picked the correct choice).
83  // Same as IRR_CLASS_LM_TRADEOFF, but used when we only run chopper on a word,
84  // and thus use the old language model (permuters).
85  // TODO(antonova): integrate the new language mode with chopper
87  // If there is an incorrect adaptive template match with a better score than
88  // a correct one (either pre-trained or adapted), mark this as adaption error.
90  // split_and_recog_word() failed to find a suitable split in truth.
92  // Truth is not available for this word (e.g. when words in corrected content
93  // file are turned into ~~~~ because an appropriate alignment was not found.
95  // The text recorded in best choice != truth text, but none of the above
96  // reasons are set.
98 
100 };
IRR_SEGSEARCH_PP
Definition: blamer.h:82
IRR_NO_TRUTH_SPLIT
Definition: blamer.h:91
IRR_UNKNOWN
Definition: blamer.h:97
IRR_CHOPPER
Definition: blamer.h:62
IRR_CLASS_LM_TRADEOFF
Definition: blamer.h:69
IRR_ADAPTION
Definition: blamer.h:89
IRR_CLASSIFIER
Definition: blamer.h:59
IRR_CORRECT
Definition: blamer.h:54
IRR_NO_TRUTH
Definition: blamer.h:94
IRR_NUM_REASONS
Definition: blamer.h:99
IRR_PAGE_LAYOUT
Definition: blamer.h:73
IRR_CLASS_OLD_LM_TRADEOFF
Definition: blamer.h:86
IRR_SEGSEARCH_HEUR
Definition: blamer.h:76