Text2ImageParameters

From Transkribus Wiki
Jump to: navigation, search

thresh

          Threshold for text alignment. If the confidence of a text-to-image
           alignment above this threshold, an alignment is done (default = 0.0). A
            good value is between 0.01 and 0.05. Note that the confidence is stored
            in the pageXML anyway, so deleting text alignments with low confidence
            can also be made later.
       

hyphen

can be null, a non-negative double value or a json-string (default: null). If no value is set or value is "Infinity", no hyphenation is done. If value is a positive double value, the value are the additional costs to recognize a hyphenation. The default hyphenation signs at the end of the line are '¬', '-', ':', '='. The default hyphenation signs at the beginning of the line are empty. There can be hyphenations between all letter-pairs. If one wants to use hyphenation rules for a specific language, this can be configured using the key 'hyphen_lang'.
The hyphenation sign in the groundtruth will be '¬'.
If one wants to configure more, one has to write a j-son-string. Keys:
prefixes: list of hyphenation sign that can be hyphens at the beginning of a line (default: empty)
suffixes: list of hyphenation sign that can be hyphens at the end of a line (default: empty)
skipSuffix: boolean if suffix is optional (true) of forced (false) (default: false)
skipPrefix: boolean if prefix is optional (true) of forced (false) (default: false)
hypCosts: non-negative value that produces additional costs to recognize a hyphenation. (default: 0.0)
pattern: language pattern (e.g. EN_GB, EN_US, DE, ES, FR,...) (default: empty)
example: "{
"skipSuffix":false,
"skipPrefix":true,
"suffixes":["¬","-",":","\u003d"],
"prefixes":[":","\u003d"],
"hypCosts":6.0,
"pattern":"EN_GB"
}"
one of the 4 suffixes have to be recognized and one of the both prefixes can be recognized. Hyphenation costs of 6.0 are added. Hyphenation is only possible as defined for language EN_GB.

hyphen_lang

    if hyphen is given, hyphenation-rules from different languages can be
    applied. If value = null or empty, a linebreak between all letters is
    possible (unicode-characters of Category L). Otherwise, a rule is applied
    ( see https://github.com/mfietz/JHyphenator.git for details). The
    language have e.g. "DE" for German and "EN" for English. Default = null.
       

skip_word

    makes it possible to skip a word, for example if a baseline is too short
    (default: null). The value have to be a positive double value. It
    repesents the default delete-costs for each character. A good value is
    4.0. The higher the value, the less words were skipped. If value = 0, a
    word can be deleted without producing costs (destroys the algorithm), if
    value = Infinity, no characters can be deleted.
       

skip_bl

    makes it possible to skip a baseline (default: null). Sometimes the LA
    finds a baseline in noise (aka false positive). It is possible to delete
    those baselines instead of "pressing" a sequence into the line. The value
    has to be positive double value. The lower the value, the easier a line
    is ignored. A good value is 0.2.
       

jump_bl

    makes it possible to handle wrong reading order in the LA (default: null)
    The value makes it possible to jump instead of the after a line to every
    other line. If value = 0, the reading order has no effect at all. If
    value = Infinity is the same like value = null. If you cannot trust the
    reading order, set value = 0.
       

best_pathes

    if the number of confmats and references gets too large, one can only
    keep a specific number of paths at each reference. As default all paths
    are calculated (like setting value = Infinity). A good value is 200.0