Total Complexity | 167 |
Total Lines | 1038 |
Duplicated Lines | 6.17 % |
Changes | 0 |
Duplicate code is one of the most pungent code smells. A rule that is often used is to re-structure code once it is duplicated in three or more places.
Common duplication problems, and corresponding solutions are:
Complex classes like ocrd.workspace often do a lot of different things. To break such a class down, we need to identify a cohesive component within that class. A common approach to find such a component is to look for fields/methods that share the same prefixes, or suffixes.
Once you have determined the fields that belong together, you can apply the Extract Class refactoring. If the component makes sense as a sub-class, Extract Subclass is also a candidate, and is often faster.
1 | import io |
||
2 | from os import makedirs, unlink, listdir, path |
||
3 | from pathlib import Path |
||
4 | from shutil import move, copyfileobj |
||
5 | from re import sub |
||
6 | |||
7 | from cv2 import COLOR_GRAY2BGR, COLOR_RGB2BGR, cvtColor |
||
8 | from PIL import Image |
||
9 | import numpy as np |
||
10 | from deprecated.sphinx import deprecated |
||
11 | |||
12 | from ocrd_models import OcrdMets, OcrdFile |
||
13 | from ocrd_models.ocrd_page import parse, BorderType, to_xml |
||
14 | from ocrd_modelfactory import exif_from_filename, page_from_file |
||
15 | from ocrd_utils import ( |
||
16 | atomic_write, |
||
17 | getLogger, |
||
18 | image_from_polygon, |
||
19 | coordinates_of_segment, |
||
20 | adjust_canvas_to_rotation, |
||
21 | adjust_canvas_to_transposition, |
||
22 | shift_coordinates, |
||
23 | rotate_coordinates, |
||
24 | transform_coordinates, |
||
25 | transpose_coordinates, |
||
26 | crop_image, |
||
27 | rotate_image, |
||
28 | transpose_image, |
||
29 | bbox_from_polygon, |
||
30 | polygon_from_points, |
||
31 | xywh_from_bbox, |
||
32 | pushd_popd, |
||
33 | MIME_TO_EXT, |
||
34 | MIME_TO_PIL, |
||
35 | MIMETYPE_PAGE, |
||
36 | REGEX_PREFIX |
||
37 | ) |
||
38 | |||
39 | from .workspace_backup import WorkspaceBackupManager |
||
40 | |||
41 | class Workspace(): |
||
42 | """ |
||
43 | A workspace is a temporary directory set up for a processor. It's the |
||
44 | interface to the METS/PAGE XML and delegates download and upload to the |
||
45 | :py:class:`ocrd.Resolver`. |
||
46 | |||
47 | Args: |
||
48 | |||
49 | directory (string) : Filesystem folder to work in |
||
50 | mets (:py:class:`ocrd_models.ocrd_mets.OcrdMets`) : `OcrdMets` representing this workspace. |
||
51 | Loaded from `'mets.xml'` if `None`. |
||
52 | mets_basename (string) : Basename of the METS XML file. Default: Last URL segment of the mets_url. |
||
53 | overwrite_mode (boolean) : Whether to force add operations on this workspace globally |
||
54 | baseurl (string) : Base URL to prefix to relative URL. |
||
55 | """ |
||
56 | |||
57 | def __init__(self, resolver, directory, mets=None, mets_basename='mets.xml', automatic_backup=False, baseurl=None): |
||
58 | self.resolver = resolver |
||
59 | self.directory = directory |
||
60 | self.mets_target = str(Path(directory, mets_basename)) |
||
61 | self.overwrite_mode = False |
||
62 | if mets is None: |
||
63 | mets = OcrdMets(filename=self.mets_target) |
||
64 | self.mets = mets |
||
65 | self.automatic_backup = automatic_backup |
||
66 | self.baseurl = baseurl |
||
67 | # print(mets.to_xml(xmllint=True).decode('utf-8')) |
||
68 | |||
69 | def __str__(self): |
||
70 | return 'Workspace[directory=%s, baseurl=%s, file_groups=%s, files=%s]' % ( |
||
71 | self.directory, |
||
72 | self.baseurl, |
||
73 | self.mets.file_groups, |
||
74 | [str(f) for f in self.mets.find_all_files()], |
||
75 | ) |
||
76 | |||
77 | def reload_mets(self): |
||
78 | """ |
||
79 | Reload METS from the filesystem. |
||
80 | """ |
||
81 | self.mets = OcrdMets(filename=self.mets_target) |
||
82 | |||
83 | def merge(self, other_workspace, copy_files=True, **kwargs): |
||
84 | """ |
||
85 | Merge ``other_workspace`` into this one |
||
86 | |||
87 | See :py:func:OcrdMets.merge: for the ``kwargs`` |
||
88 | |||
89 | Keyword Args: |
||
90 | copy_files (boolean): Whether to copy files from ``other_workspace`` to this one |
||
91 | """ |
||
92 | def after_add_cb(f): |
||
93 | if not copy_files: |
||
94 | return |
||
95 | fpath_src = Path(other_workspace.directory, f.url) |
||
96 | fpath_dest = Path(self.directory, f.url) |
||
97 | if fpath_src.exists(): |
||
98 | if fpath_dest.exists(): |
||
99 | raise Exception("Copying %s to %s would overwrite the latter" % (fpath_src, fpath_dest)) |
||
100 | if not fpath_dest.parent.is_dir(): |
||
101 | makedirs(str(fpath_dest.parent)) |
||
102 | with open(str(fpath_src), 'rb') as fstream_in, open(str(fpath_dest), 'wb') as fstream_out: |
||
103 | copyfileobj(fstream_in, fstream_out) |
||
104 | self.mets.merge(other_workspace.mets, after_add_cb=after_add_cb, **kwargs) |
||
105 | |||
106 | |||
107 | @deprecated(version='1.0.0', reason="Use workspace.download_file") |
||
108 | def download_url(self, url, **kwargs): |
||
109 | """ |
||
110 | Download a URL to the workspace. |
||
111 | |||
112 | Args: |
||
113 | url (string): URL to download to directory |
||
114 | **kwargs : See :py:class:`ocrd_models.ocrd_file.OcrdFile` |
||
115 | |||
116 | Returns: |
||
117 | The local filename of the downloaded file |
||
118 | """ |
||
119 | f = OcrdFile(None, url=url, **kwargs) |
||
120 | f = self.download_file(f) |
||
121 | return f.local_filename |
||
122 | |||
123 | |||
124 | def download_file(self, f, _recursion_count=0): |
||
125 | """ |
||
126 | Download a :py:class:`ocrd_models.ocrd_file.OcrdFile` to the workspace. |
||
127 | """ |
||
128 | log = getLogger('ocrd.workspace.download_file') |
||
129 | log.debug('download_file %s [_recursion_count=%s]' % (f, _recursion_count)) |
||
130 | with pushd_popd(self.directory): |
||
131 | try: |
||
132 | # If the f.url is already a file path, and is within self.directory, do nothing |
||
133 | url_path = Path(f.url).resolve() |
||
134 | if not (url_path.exists() and url_path.relative_to(str(Path(self.directory).resolve()))): |
||
135 | raise Exception("Not already downloaded, moving on") |
||
136 | except Exception as e: |
||
137 | basename = '%s%s' % (f.ID, MIME_TO_EXT.get(f.mimetype, '')) if f.ID else f.basename |
||
138 | try: |
||
139 | f.url = self.resolver.download_to_directory(self.directory, f.url, subdir=f.fileGrp, basename=basename) |
||
140 | except FileNotFoundError as e: |
||
141 | if not self.baseurl: |
||
142 | raise Exception("No baseurl defined by workspace. Cannot retrieve '%s'" % f.url) |
||
143 | if _recursion_count >= 1: |
||
144 | raise Exception("Already tried prepending baseurl '%s'. Cannot retrieve '%s'" % (self.baseurl, f.url)) |
||
145 | log.debug("First run of resolver.download_to_directory(%s) failed, try prepending baseurl '%s': %s", f.url, self.baseurl, e) |
||
146 | f.url = '%s/%s' % (self.baseurl, f.url) |
||
147 | f.url = self.download_file(f, _recursion_count + 1).local_filename |
||
148 | f.local_filename = f.url |
||
149 | return f |
||
150 | |||
151 | def remove_file(self, ID, force=False, keep_file=False, page_recursive=False, page_same_group=False): |
||
152 | """ |
||
153 | Remove a METS `file` from the workspace. |
||
154 | |||
155 | Arguments: |
||
156 | ID (string|:py:class:`ocrd_models.ocrd_file.OcrdFile`): `@ID` of the METS `file` |
||
157 | to delete or the file itself |
||
158 | Keyword Args: |
||
159 | force (boolean): Continue removing even if file not found in METS |
||
160 | keep_file (boolean): Whether to keep files on disk |
||
161 | page_recursive (boolean): Whether to remove all images referenced in the file |
||
162 | if the file is a PAGE-XML document. |
||
163 | page_same_group (boolean): Remove only images in the same file group as the PAGE-XML. |
||
164 | Has no effect unless ``page_recursive`` is `True`. |
||
165 | """ |
||
166 | log = getLogger('ocrd.workspace.remove_file') |
||
167 | log.debug('Deleting mets:file %s', ID) |
||
168 | if not force and self.overwrite_mode: |
||
169 | force = True |
||
170 | if isinstance(ID, OcrdFile): |
||
171 | ID = ID.ID |
||
172 | try: |
||
173 | ocrd_file_ = self.mets.remove_file(ID) |
||
174 | ocrd_files = [ocrd_file_] if isinstance(ocrd_file_, OcrdFile) else ocrd_file_ |
||
175 | if page_recursive: |
||
176 | with pushd_popd(self.directory): |
||
177 | for ocrd_file in ocrd_files: |
||
178 | if ocrd_file.mimetype != MIMETYPE_PAGE: |
||
179 | continue |
||
180 | ocrd_page = parse(self.download_file(ocrd_file).local_filename, silence=True) |
||
181 | for img_url in ocrd_page.get_AllAlternativeImagePaths(): |
||
182 | img_kwargs = {'url': img_url} |
||
183 | if page_same_group: |
||
184 | img_kwargs['fileGrp'] = ocrd_file.fileGrp |
||
185 | for img_file in self.mets.find_files(**img_kwargs): |
||
186 | self.remove_file(img_file, keep_file=keep_file, force=force) |
||
187 | if not keep_file: |
||
188 | with pushd_popd(self.directory): |
||
189 | for ocrd_file in ocrd_files: |
||
190 | if not ocrd_file.local_filename: |
||
191 | log.warning("File not locally available %s", ocrd_file) |
||
192 | if not force: |
||
193 | raise Exception("File not locally available %s" % ocrd_file) |
||
194 | else: |
||
195 | log.info("rm %s [cwd=%s]", ocrd_file.local_filename, self.directory) |
||
196 | unlink(ocrd_file.local_filename) |
||
197 | return ocrd_file_ |
||
198 | except FileNotFoundError as e: |
||
199 | if not force: |
||
200 | raise e |
||
201 | |||
202 | def remove_file_group(self, USE, recursive=False, force=False, keep_files=False, page_recursive=False, page_same_group=False): |
||
203 | """ |
||
204 | Remove a METS `fileGrp`. |
||
205 | |||
206 | Arguments: |
||
207 | USE (string): `@USE` of the METS `fileGrp` to delete |
||
208 | Keyword Args: |
||
209 | recursive (boolean): Whether to recursively delete all files in the group |
||
210 | force (boolean): Continue removing even if group or containing files not found in METS |
||
211 | keep_files (boolean): When deleting recursively whether to keep files on disk |
||
212 | page_recursive (boolean): Whether to remove all images referenced in the file |
||
213 | if the file is a PAGE-XML document. |
||
214 | page_same_group (boolean): Remove only images in the same file group as the PAGE-XML. |
||
215 | Has no effect unless ``page_recursive`` is `True`. |
||
216 | """ |
||
217 | if not force and self.overwrite_mode: |
||
218 | force = True |
||
219 | |||
220 | if (not USE.startswith(REGEX_PREFIX)) and (USE not in self.mets.file_groups) and (not force): |
||
221 | raise Exception("No such fileGrp: %s" % USE) |
||
222 | |||
223 | file_dirs = [] |
||
224 | if recursive: |
||
225 | for f in self.mets.find_files(fileGrp=USE): |
||
226 | self.remove_file(f, force=force, keep_file=keep_files, page_recursive=page_recursive, page_same_group=page_same_group) |
||
227 | if f.local_filename: |
||
228 | file_dirs.append(path.dirname(f.local_filename)) |
||
229 | |||
230 | self.mets.remove_file_group(USE, force=force) |
||
231 | |||
232 | # PLEASE NOTE: this only removes directories in the workspace if they are empty |
||
233 | # and named after the fileGrp which is a convention in OCR-D. |
||
234 | with pushd_popd(self.directory): |
||
235 | if Path(USE).is_dir() and not listdir(USE): |
||
236 | Path(USE).rmdir() |
||
237 | if file_dirs: |
||
238 | for file_dir in set(file_dirs): |
||
239 | if Path(file_dir).is_dir() and not listdir(file_dir): |
||
240 | Path(file_dir).rmdir() |
||
241 | |||
242 | |||
243 | def rename_file_group(self, old, new): |
||
244 | """ |
||
245 | Rename a METS `fileGrp`. |
||
246 | |||
247 | Arguments: |
||
248 | old (string): `@USE` of the METS `fileGrp` to rename |
||
249 | new (string): `@USE` of the METS `fileGrp` to rename as |
||
250 | """ |
||
251 | log = getLogger('ocrd.workspace.rename_file_group') |
||
252 | |||
253 | if old not in self.mets.file_groups: |
||
254 | raise ValueError("No such fileGrp: %s" % old) |
||
255 | if new in self.mets.file_groups: |
||
256 | raise ValueError("fileGrp already exists %s" % new) |
||
257 | |||
258 | with pushd_popd(self.directory): |
||
259 | # create workspace dir ``new`` |
||
260 | log.info("mkdir %s" % new) |
||
261 | if not Path(new).is_dir(): |
||
262 | Path(new).mkdir() |
||
263 | url_replacements = {} |
||
264 | log.info("Moving files") |
||
265 | for mets_file in self.mets.find_files(fileGrp=old, local_only=True): |
||
266 | new_url = sub(r'^%s/' % old, '%s/' % new, mets_file.url) |
||
267 | url_replacements[mets_file.url] = new_url |
||
268 | # move file from ``old`` to ``new`` |
||
269 | move(mets_file.url, new_url) |
||
270 | # change the url of ``mets:file`` |
||
271 | mets_file.url = new_url |
||
272 | # change file paths in PAGE-XML imageFilename and filename attributes |
||
273 | for page_file in self.mets.find_files(mimetype=MIMETYPE_PAGE, local_only=True): |
||
274 | log.info("Renaming file references in PAGE-XML %s" % page_file) |
||
275 | pcgts = page_from_file(page_file) |
||
276 | changed = False |
||
277 | for old_url, new_url in url_replacements.items(): |
||
278 | if pcgts.get_Page().imageFilename == old_url: |
||
279 | changed = True |
||
280 | log.info("Rename pc:Page/@imageFilename: %s -> %s" % (old_url, new_url)) |
||
281 | pcgts.get_Page().imageFilename = new_url |
||
282 | for ai in pcgts.get_Page().get_AllAlternativeImages(): |
||
283 | for old_url, new_url in url_replacements.items(): |
||
284 | if ai.filename == old_url: |
||
285 | changed = True |
||
286 | log.info("Rename pc:Page/../AlternativeImage: %s -> %s" % (old_url, new_url)) |
||
287 | ai.filename = new_url |
||
288 | if changed: |
||
289 | log.info("PAGE-XML changed, writing %s" % (page_file.local_filename)) |
||
290 | with open(page_file.local_filename, 'w', encoding='utf-8') as f: |
||
291 | f.write(to_xml(pcgts)) |
||
292 | # change the ``USE`` attribute of the fileGrp |
||
293 | self.mets.rename_file_group(old, new) |
||
294 | # Remove the old dir |
||
295 | log.info("rmdir %s" % old) |
||
296 | if Path(old).is_dir() and not listdir(old): |
||
297 | Path(old).rmdir() |
||
298 | |||
299 | def add_file(self, file_grp, content=None, **kwargs): |
||
300 | """ |
||
301 | Add a file to the :py:class:`ocrd_models.ocrd_mets.OcrdMets` of the workspace. |
||
302 | |||
303 | Arguments: |
||
304 | file_grp (string): `@USE` of the METS `fileGrp` to add to |
||
305 | Keyword Args: |
||
306 | content (string|bytes): optional content to write to the file |
||
307 | in the filesystem |
||
308 | **kwargs: See :py:func:`ocrd_models.ocrd_mets.OcrdMets.add_file` |
||
309 | Returns: |
||
310 | a new :py:class:`ocrd_models.ocrd_file.OcrdFile` |
||
311 | """ |
||
312 | log = getLogger('ocrd.workspace.add_file') |
||
313 | log.debug( |
||
314 | 'outputfile file_grp=%s local_filename=%s content=%s', |
||
315 | file_grp, |
||
316 | kwargs.get('local_filename'), |
||
317 | content is not None) |
||
318 | if 'pageId' not in kwargs: |
||
319 | raise ValueError("workspace.add_file must be passed a 'pageId' kwarg, even if it is None.") |
||
320 | if content is not None and 'local_filename' not in kwargs: |
||
321 | raise Exception("'content' was set but no 'local_filename'") |
||
322 | if self.overwrite_mode: |
||
323 | kwargs['force'] = True |
||
324 | |||
325 | with pushd_popd(self.directory): |
||
326 | if 'local_filename' in kwargs: |
||
327 | # If the local filename has folder components, create those folders |
||
328 | local_filename_dir = kwargs['local_filename'].rsplit('/', 1)[0] |
||
329 | if local_filename_dir != kwargs['local_filename'] and not Path(local_filename_dir).is_dir(): |
||
330 | makedirs(local_filename_dir) |
||
331 | if 'url' not in kwargs: |
||
332 | kwargs['url'] = kwargs['local_filename'] |
||
333 | |||
334 | # print(kwargs) |
||
335 | ret = self.mets.add_file(file_grp, **kwargs) |
||
336 | |||
337 | if content is not None: |
||
338 | with open(kwargs['local_filename'], 'wb') as f: |
||
339 | if isinstance(content, str): |
||
340 | content = bytes(content, 'utf-8') |
||
341 | f.write(content) |
||
342 | |||
343 | return ret |
||
344 | |||
345 | def save_mets(self): |
||
346 | """ |
||
347 | Write out the current state of the METS file to the filesystem. |
||
348 | """ |
||
349 | log = getLogger('ocrd.workspace.save_mets') |
||
350 | log.info("Saving mets '%s'", self.mets_target) |
||
351 | if self.automatic_backup: |
||
352 | WorkspaceBackupManager(self).add() |
||
353 | with atomic_write(self.mets_target) as f: |
||
354 | f.write(self.mets.to_xml(xmllint=True).decode('utf-8')) |
||
355 | |||
356 | def resolve_image_exif(self, image_url): |
||
357 | """ |
||
358 | Get the EXIF metadata about an image URL as :py:class:`ocrd_models.ocrd_exif.OcrdExif` |
||
359 | |||
360 | Args: |
||
361 | image_url (string) : `@href` (path or URL) of the METS `file` to inspect |
||
362 | |||
363 | Returns: |
||
364 | :py:class:`ocrd_models.ocrd_exif.OcrdExif` |
||
365 | """ |
||
366 | if not image_url: |
||
367 | # avoid "finding" just any file |
||
368 | raise Exception("Cannot resolve empty image path") |
||
369 | f = next(self.mets.find_files(url=image_url), OcrdFile(None, url=image_url)) |
||
370 | image_filename = self.download_file(f).local_filename |
||
371 | ocrd_exif = exif_from_filename(image_filename) |
||
372 | return ocrd_exif |
||
373 | |||
374 | @deprecated(version='1.0.0', reason="Use workspace.image_from_page and workspace.image_from_segment") |
||
375 | def resolve_image_as_pil(self, image_url, coords=None): |
||
376 | """ |
||
377 | Resolve an image URL to a `PIL.Image`. |
||
378 | |||
379 | Arguments: |
||
380 | image_url (string): `@href` (path or URL) of the METS `file` to retrieve |
||
381 | Keyword Args: |
||
382 | coords (list) : Coordinates of the bounding box to cut from the image |
||
383 | |||
384 | Returns: |
||
385 | Full or cropped `PIL.Image` |
||
386 | |||
387 | """ |
||
388 | return self._resolve_image_as_pil(image_url, coords) |
||
389 | |||
390 | def _resolve_image_as_pil(self, image_url, coords=None): |
||
391 | if not image_url: |
||
392 | # avoid "finding" just any file |
||
393 | raise Exception("Cannot resolve empty image path") |
||
394 | log = getLogger('ocrd.workspace._resolve_image_as_pil') |
||
395 | f = next(self.mets.find_files(url=image_url), OcrdFile(None, url=image_url)) |
||
396 | image_filename = self.download_file(f).local_filename |
||
397 | |||
398 | with pushd_popd(self.directory): |
||
399 | pil_image = Image.open(image_filename) |
||
400 | pil_image.load() # alloc and give up the FD |
||
401 | |||
402 | # Pillow does not properly support higher color depths |
||
403 | # (e.g. 16-bit or 32-bit or floating point grayscale), |
||
404 | # clipping its dynamic range to the lower 8-bit in |
||
405 | # many operations (including paste, putalpha, ImageStat...), |
||
406 | # even including conversion. |
||
407 | # Cf. Pillow#3011 Pillow#3159 Pillow#3838 (still open in 8.0) |
||
408 | # So to be on the safe side, we must re-quantize these |
||
409 | # to 8-bit via numpy (conversion to/from which fortunately |
||
410 | # seems to work reliably): |
||
411 | if (pil_image.mode.startswith('I') or |
||
412 | pil_image.mode.startswith('F')): |
||
413 | arr_image = np.array(pil_image) |
||
414 | if arr_image.dtype.kind == 'i': |
||
415 | # signed integer is *not* trustworthy in this context |
||
416 | # (usually a mistake in the array interface) |
||
417 | log.debug('Casting image "%s" from signed to unsigned', image_url) |
||
418 | arr_image.dtype = np.dtype('u' + arr_image.dtype.name) |
||
419 | if arr_image.dtype.kind == 'u': |
||
420 | # integer needs to be scaled linearly to 8 bit |
||
421 | # of course, an image might actually have some lower range |
||
422 | # (e.g. 10-bit in I;16 or 20-bit in I or 4-bit in L), |
||
423 | # but that would be guessing anyway, so here don't |
||
424 | # make assumptions on _scale_, just reduce _precision_ |
||
425 | log.debug('Reducing image "%s" from depth %d bit to 8 bit', |
||
426 | image_url, arr_image.dtype.itemsize * 8) |
||
427 | arr_image = arr_image >> 8 * (arr_image.dtype.itemsize-1) |
||
428 | arr_image = arr_image.astype(np.uint8) |
||
429 | elif arr_image.dtype.kind == 'f': |
||
430 | # float needs to be scaled from [0,1.0] to [0,255] |
||
431 | log.debug('Reducing image "%s" from floating point to 8 bit', |
||
432 | image_url) |
||
433 | arr_image *= 255 |
||
434 | arr_image = arr_image.astype(np.uint8) |
||
435 | pil_image = Image.fromarray(arr_image) |
||
436 | |||
437 | if coords is None: |
||
438 | return pil_image |
||
439 | |||
440 | # FIXME: remove or replace this by (image_from_polygon+) crop_image ... |
||
441 | log.debug("Converting PIL to OpenCV: %s", image_url) |
||
442 | color_conversion = COLOR_GRAY2BGR if pil_image.mode in ('1', 'L') else COLOR_RGB2BGR |
||
443 | pil_as_np_array = np.array(pil_image).astype('uint8') if pil_image.mode == '1' else np.array(pil_image) |
||
444 | cv2_image = cvtColor(pil_as_np_array, color_conversion) |
||
445 | |||
446 | poly = np.array(coords, np.int32) |
||
447 | log.debug("Cutting region %s from %s", coords, image_url) |
||
448 | region_cut = cv2_image[ |
||
449 | np.min(poly[:, 1]):np.max(poly[:, 1]), |
||
450 | np.min(poly[:, 0]):np.max(poly[:, 0]) |
||
451 | ] |
||
452 | return Image.fromarray(region_cut) |
||
453 | |||
454 | def image_from_page(self, page, page_id, |
||
455 | fill='background', transparency=False, |
||
456 | feature_selector='', feature_filter=''): |
||
457 | """Extract an image for a PAGE-XML page from the workspace. |
||
458 | |||
459 | Args: |
||
460 | page (:py:class:`ocrd_models.ocrd_page.PageType`): a PAGE `PageType` object |
||
461 | page_id (string): its `@ID` in the METS physical `structMap` |
||
462 | Keyword Args: |
||
463 | fill (string): a `PIL` color specifier |
||
464 | transparency (boolean): whether to add an alpha channel for masking |
||
465 | feature_selector (string): a comma-separated list of `@comments` classes |
||
466 | feature_filter (string): a comma-separated list of `@comments` classes |
||
467 | |||
468 | Extract a `PIL.Image` from ``page``, either from its `AlternativeImage` |
||
469 | (if it exists), or from its `@imageFilename` (otherwise). Also crop it, |
||
470 | if a `Border` exists, and rotate it, if any `@orientation` angle is |
||
471 | annotated. |
||
472 | |||
473 | If ``feature_selector`` and/or ``feature_filter`` is given, then |
||
474 | select/filter among the `@imageFilename` image and the available |
||
475 | AlternativeImages the last one which contains all of the selected, |
||
476 | but none of the filtered features (i.e. `@comments` classes), or |
||
477 | raise an error. |
||
478 | |||
479 | (Required and produced features need not be in the same order, so |
||
480 | ``feature_selector`` is merely a mask specifying Boolean AND, and |
||
481 | ``feature_filter`` is merely a mask specifying Boolean OR.) |
||
482 | |||
483 | If the chosen image does not have the feature `"cropped"` yet, but |
||
484 | a `Border` exists, and unless `"cropped"` is being filtered, then crop it. |
||
485 | Likewise, if the chosen image does not have the feature `"deskewed"` yet, |
||
486 | but an `@orientation` angle is annotated, and unless `"deskewed"` is being |
||
487 | filtered, then rotate it. (However, if `@orientation` is above the |
||
488 | [-45°,45°] interval, then apply as much transposition as possible first, |
||
489 | unless `"rotated-90"` / `"rotated-180"` / `"rotated-270"` is being filtered.) |
||
490 | |||
491 | Cropping uses a polygon mask (not just the bounding box rectangle). |
||
492 | Areas outside the polygon will be filled according to ``fill``: |
||
493 | |||
494 | - if `"background"` (the default), |
||
495 | then fill with the median color of the image; |
||
496 | - otherwise, use the given color, e.g. `"white"` or `(255,255,255)`. |
||
497 | |||
498 | Moreover, if ``transparency`` is true, and unless the image already |
||
499 | has an alpha channel, then add an alpha channel which is fully opaque |
||
500 | before cropping and rotating. (Thus, unexposed/masked areas will be |
||
501 | transparent afterwards for consumers that can interpret alpha channels). |
||
502 | |||
503 | Returns: |
||
504 | a tuple of |
||
505 | * the extracted `PIL.Image`, |
||
506 | * a `dict` with information about the extracted image: |
||
507 | - `"transform"`: a `Numpy` array with an affine transform which |
||
508 | converts from absolute coordinates to those relative to the image, |
||
509 | i.e. after cropping to the page's border / bounding box (if any) |
||
510 | and deskewing with the page's orientation angle (if any) |
||
511 | - `"angle"`: the rotation/reflection angle applied to the image so far, |
||
512 | - `"features"`: the `AlternativeImage` `@comments` for the image, i.e. |
||
513 | names of all applied operations that lead up to this result, |
||
514 | * an :py:class:`ocrd_models.ocrd_exif.OcrdExif` instance associated with |
||
515 | the original image. |
||
516 | |||
517 | (The first two can be used to annotate a new `AlternativeImage`, |
||
518 | or be passed down with :py:meth:`image_from_segment`.) |
||
519 | |||
520 | Examples: |
||
521 | |||
522 | * get a raw (colored) but already deskewed and cropped image:: |
||
523 | |||
524 | page_image, page_coords, page_image_info = workspace.image_from_page( |
||
525 | page, page_id, |
||
526 | feature_selector='deskewed,cropped', |
||
527 | feature_filter='binarized,grayscale_normalized') |
||
528 | """ |
||
529 | log = getLogger('ocrd.workspace.image_from_page') |
||
530 | page_image_info = self.resolve_image_exif(page.imageFilename) |
||
531 | page_image = self._resolve_image_as_pil(page.imageFilename) |
||
532 | page_coords = dict() |
||
533 | # use identity as initial affine coordinate transform: |
||
534 | page_coords['transform'] = np.eye(3) |
||
535 | # interim bbox (updated with each change to the transform): |
||
536 | page_bbox = [0, 0, page_image.width, page_image.height] |
||
537 | page_xywh = {'x': 0, 'y': 0, |
||
538 | 'w': page_image.width, 'h': page_image.height} |
||
539 | |||
540 | border = page.get_Border() |
||
541 | # page angle: PAGE @orientation is defined clockwise, |
||
542 | # whereas PIL/ndimage rotation is in mathematical direction: |
||
543 | page_coords['angle'] = -(page.get_orientation() or 0) |
||
544 | # map angle from (-180,180] to [0,360], and partition into multiples of 90; |
||
545 | # but avoid unnecessary large remainders, i.e. split symmetrically: |
||
546 | orientation = (page_coords['angle'] + 45) % 360 |
||
547 | orientation = orientation - (orientation % 90) |
||
548 | skew = (page_coords['angle'] % 360) - orientation |
||
549 | skew = 180 - (180 - skew) % 360 # map to [-45,45] |
||
550 | page_coords['angle'] = 0 # nothing applied yet (depends on filters) |
||
551 | log.debug("page '%s' has %s orientation=%d skew=%.2f", |
||
552 | page_id, "border," if border else "", orientation, skew) |
||
553 | |||
554 | # initialize AlternativeImage@comments classes as empty: |
||
555 | page_coords['features'] = '' |
||
556 | alternative_image = None |
||
557 | alternative_images = page.get_AlternativeImage() |
||
558 | View Code Duplication | if alternative_images: |
|
|
|||
559 | # (e.g. from page-level cropping, binarization, deskewing or despeckling) |
||
560 | if feature_selector or feature_filter: |
||
561 | alternative_image = None |
||
562 | # search from the end, because by convention we always append, |
||
563 | # and among multiple satisfactory images we want the most recent: |
||
564 | for alternative_image in reversed(alternative_images): |
||
565 | features = alternative_image.get_comments() |
||
566 | if not features: |
||
567 | log.warning("AlternativeImage %d for page '%s' does not have any feature attributes", |
||
568 | alternative_images.index(alternative_image) + 1, page_id) |
||
569 | features = '' |
||
570 | if (all(feature in features |
||
571 | for feature in feature_selector.split(',') if feature) and |
||
572 | not any(feature in features |
||
573 | for feature in feature_filter.split(',') if feature)): |
||
574 | break |
||
575 | else: |
||
576 | alternative_image = None |
||
577 | else: |
||
578 | alternative_image = alternative_images[-1] |
||
579 | features = alternative_image.get_comments() |
||
580 | if not features: |
||
581 | log.warning("AlternativeImage %d for page '%s' does not have any feature attributes", |
||
582 | alternative_images.index(alternative_image) + 1, page_id) |
||
583 | features = '' |
||
584 | if alternative_image: |
||
585 | log.debug("Using AlternativeImage %d (%s) for page '%s'", |
||
586 | alternative_images.index(alternative_image) + 1, |
||
587 | features, page_id) |
||
588 | page_image = self._resolve_image_as_pil(alternative_image.get_filename()) |
||
589 | page_coords['features'] = features |
||
590 | |||
591 | # adjust the coord transformation to the steps applied on the image, |
||
592 | # and apply steps on the existing image in case it is missing there, |
||
593 | # but traverse all steps (crop/reflect/rotate) in a particular order: |
||
594 | # - existing image features take priority (in the order annotated), |
||
595 | # - next is cropping (if necessary but not already applied), |
||
596 | # - next is reflection (if necessary but not already applied), |
||
597 | # - next is rotation (if necessary but not already applied). |
||
598 | # This helps deal with arbitrary workflows (e.g. crop then deskew, |
||
599 | # or deskew then crop), regardless of where images are generated. |
||
600 | alternative_image_features = page_coords['features'].split(',') |
||
601 | for duplicate_feature in set([feature for feature in alternative_image_features |
||
602 | # features relevant in reconstructing coordinates: |
||
603 | if (feature in ['cropped', 'deskewed', 'rotated-90', |
||
604 | 'rotated-180', 'rotated-270'] and |
||
605 | alternative_image_features.count(feature) > 1)]): |
||
606 | log.error("Duplicate feature %s in AlternativeImage for page '%s'", |
||
607 | duplicate_feature, page_id) |
||
608 | for i, feature in enumerate(alternative_image_features + |
||
609 | (['cropped'] |
||
610 | if (border and |
||
611 | not 'cropped' in alternative_image_features and |
||
612 | not 'cropped' in feature_filter.split(',')) |
||
613 | else []) + |
||
614 | (['rotated-%d' % orientation] |
||
615 | if (orientation and |
||
616 | not 'rotated-%d' % orientation in alternative_image_features and |
||
617 | not 'rotated-%d' % orientation in feature_filter.split(',')) |
||
618 | else []) + |
||
619 | (['deskewed'] |
||
620 | if (skew and |
||
621 | not 'deskewed' in alternative_image_features and |
||
622 | not 'deskewed' in feature_filter.split(',')) |
||
623 | else []) + |
||
624 | # not a feature to be added, but merely as a fallback position |
||
625 | # to always enter loop at i == len(alternative_image_features) |
||
626 | ['_check']): |
||
627 | # image geometry vs feature consistency can only be checked |
||
628 | # after all features on the existing AlternativeImage have |
||
629 | # been adjusted for in the transform, and when there is a mismatch, |
||
630 | # additional steps applied here would only repeat the respective |
||
631 | # error message; so we only check once at the boundary between |
||
632 | # existing and new features |
||
633 | # FIXME we should check/enforce consistency when _adding_ AlternativeImage |
||
634 | if (i == len(alternative_image_features) and |
||
635 | not (page_xywh['w'] - 2 < page_image.width < page_xywh['w'] + 2 and |
||
636 | page_xywh['h'] - 2 < page_image.height < page_xywh['h'] + 2)): |
||
637 | log.error('page "%s" image (%s; %dx%d) has not been cropped properly (%dx%d)', |
||
638 | page_id, page_coords['features'], |
||
639 | page_image.width, page_image.height, |
||
640 | page_xywh['w'], page_xywh['h']) |
||
641 | name = "%s for page '%s'" % ("AlternativeImage" if alternative_image |
||
642 | else "original image", page_id) |
||
643 | # adjust transform to feature, and ensure feature is applied to image |
||
644 | if feature == 'cropped': |
||
645 | page_image, page_coords, page_xywh = _crop( |
||
646 | log, name, border, page_image, page_coords, |
||
647 | fill=fill, transparency=transparency) |
||
648 | elif feature == 'rotated-%d' % orientation: |
||
649 | page_image, page_coords, page_xywh = _reflect( |
||
650 | log, name, orientation, page_image, page_coords, page_xywh) |
||
651 | elif feature == 'deskewed': |
||
652 | page_image, page_coords, page_xywh = _rotate( |
||
653 | log, name, skew, border, page_image, page_coords, page_xywh, |
||
654 | fill=fill, transparency=transparency) |
||
655 | |||
656 | # verify constraints again: |
||
657 | if not all(feature in page_coords['features'] |
||
658 | for feature in feature_selector.split(',') if feature): |
||
659 | raise Exception('Found no AlternativeImage that satisfies all requirements ' + |
||
660 | 'selector="%s" in page "%s"' % ( |
||
661 | feature_selector, page_id)) |
||
662 | if any(feature in page_coords['features'] |
||
663 | for feature in feature_filter.split(',') if feature): |
||
664 | raise Exception('Found no AlternativeImage that satisfies all requirements ' + |
||
665 | 'filter="%s" in page "%s"' % ( |
||
666 | feature_filter, page_id)) |
||
667 | page_image.format = 'PNG' # workaround for tesserocr#194 |
||
668 | return page_image, page_coords, page_image_info |
||
669 | |||
670 | def image_from_segment(self, segment, parent_image, parent_coords, |
||
671 | fill='background', transparency=False, |
||
672 | feature_selector='', feature_filter=''): |
||
673 | """Extract an image for a PAGE-XML hierarchy segment from its parent's image. |
||
674 | |||
675 | Args: |
||
676 | segment (object): a PAGE segment object |
||
677 | (i.e. :py:class:`ocrd_models.ocrd_page.TextRegionType` |
||
678 | or :py:class:`ocrd_models.ocrd_page.TextLineType` |
||
679 | or :py:class:`ocrd_models.ocrd_page.WordType` |
||
680 | or :py:class:`ocrd_models.ocrd_page.GlyphType`) |
||
681 | parent_image (PIL.Image): image of the segment's parent |
||
682 | parent_coords (dict): a `dict` with information about ``parent_image``: |
||
683 | - `"transform"`: a `Numpy` array with an affine transform which |
||
684 | converts from absolute coordinates to those relative to the image, |
||
685 | i.e. after applying all operations (starting with the original image) |
||
686 | - `"angle"`: the rotation/reflection angle applied to the image so far, |
||
687 | - `"features"`: the `AlternativeImage` `@comments` for the image, i.e. |
||
688 | names of all operations that lead up to this result, and |
||
689 | Keyword Args: |
||
690 | fill (string): a `PIL` color specifier |
||
691 | transparency (boolean): whether to add an alpha channel for masking |
||
692 | feature_selector (string): a comma-separated list of `@comments` classes |
||
693 | feature_filter (string): a comma-separated list of `@comments` classes |
||
694 | |||
695 | Extract a `PIL.Image` from ``segment``, either from `AlternativeImage` |
||
696 | (if it exists), or producing a new image via cropping from ``parent_image`` |
||
697 | (otherwise). Pass in ``parent_image`` and ``parent_coords`` from the result |
||
698 | of the next higher-level of this function or from :py:meth:`image_from_page`. |
||
699 | |||
700 | If ``feature_selector`` and/or ``feature_filter`` is given, then |
||
701 | select/filter among the cropped ``parent_image`` and the available |
||
702 | `AlternativeImage`s the last one which contains all of the selected, |
||
703 | but none of the filtered features (i.e. `@comments` classes), or |
||
704 | raise an error. |
||
705 | |||
706 | (Required and produced features need not be in the same order, so |
||
707 | ``feature_selector`` is merely a mask specifying Boolean AND, and |
||
708 | ``feature_filter`` is merely a mask specifying Boolean OR.) |
||
709 | |||
710 | Cropping uses a polygon mask (not just the bounding box rectangle). |
||
711 | Areas outside the polygon will be filled according to ``fill``: |
||
712 | |||
713 | - if `"background"` (the default), |
||
714 | then fill with the median color of the image; |
||
715 | - otherwise, use the given color, e.g. `"white"` or `(255,255,255)`. |
||
716 | |||
717 | Moreover, if ``transparency`` is true, and unless the image already |
||
718 | has an alpha channel, then add an alpha channel which is fully opaque |
||
719 | before cropping and rotating. (Thus, unexposed/masked areas will be |
||
720 | transparent afterwards for consumers that can interpret alpha channels). |
||
721 | |||
722 | When cropping, compensate any `@orientation` angle annotated for the |
||
723 | parent (from parent-level deskewing) by rotating the segment coordinates |
||
724 | in an inverse transformation (i.e. translation to center, then passive |
||
725 | rotation, and translation back). |
||
726 | |||
727 | Regardless, if any `@orientation` angle is annotated for the segment |
||
728 | (from segment-level deskewing), and the chosen image does not have |
||
729 | the feature `"deskewed"` yet, and unless `"deskewed"` is being filtered, |
||
730 | then rotate it - compensating for any previous `"angle"`. (However, |
||
731 | if `@orientation` is above the [-45°,45°] interval, then apply as much |
||
732 | transposition as possible first, unless `"rotated-90"` / `"rotated-180"` / |
||
733 | `"rotated-270"` is being filtered.) |
||
734 | |||
735 | Returns: |
||
736 | a tuple of |
||
737 | * the extracted `PIL.Image`, |
||
738 | * a `dict` with information about the extracted image: |
||
739 | - `"transform"`: a `Numpy` array with an affine transform which |
||
740 | converts from absolute coordinates to those relative to the image, |
||
741 | i.e. after applying all parent operations, and then cropping to |
||
742 | the segment's bounding box, and deskewing with the segment's |
||
743 | orientation angle (if any) |
||
744 | - `"angle"`: the rotation/reflection angle applied to the image so far, |
||
745 | - `"features"`: the `AlternativeImage` `@comments` for the image, i.e. |
||
746 | names of all applied operations that lead up to this result. |
||
747 | |||
748 | (These can be used to create a new `AlternativeImage`, or passed down |
||
749 | for :py:meth:`image_from_segment` calls on lower hierarchy levels.) |
||
750 | |||
751 | Examples: |
||
752 | |||
753 | * get a raw (colored) but already deskewed and cropped image:: |
||
754 | |||
755 | image, xywh = workspace.image_from_segment(region, |
||
756 | page_image, page_xywh, |
||
757 | feature_selector='deskewed,cropped', |
||
758 | feature_filter='binarized,grayscale_normalized') |
||
759 | """ |
||
760 | log = getLogger('ocrd.workspace.image_from_segment') |
||
761 | # note: We should mask overlapping neighbouring segments here, |
||
762 | # but finding the right clipping rules can be difficult if operating |
||
763 | # on the raw (non-binary) image data alone: for each intersection, it |
||
764 | # must be decided which one of either segment or neighbour to assign, |
||
765 | # e.g. an ImageRegion which properly contains our TextRegion should be |
||
766 | # completely ignored, but an ImageRegion which is properly contained |
||
767 | # in our TextRegion should be completely masked, while partial overlap |
||
768 | # may be more difficult to decide. On the other hand, on the binary image, |
||
769 | # we can use connected component analysis to mask foreground areas which |
||
770 | # originate in the neighbouring regions. But that would introduce either |
||
771 | # the assumption that the input has already been binarized, or a dependency |
||
772 | # on some ad-hoc binarization method. Thus, it is preferable to use |
||
773 | # a dedicated processor for this (which produces clipped AlternativeImage |
||
774 | # or reduced polygon coordinates). |
||
775 | segment_image, segment_coords, segment_xywh = _crop( |
||
776 | log, "parent image for segment '%s'" % segment.id, |
||
777 | segment, parent_image, parent_coords, |
||
778 | fill=fill, transparency=transparency) |
||
779 | |||
780 | # Semantics of missing @orientation at region level could be either |
||
781 | # - inherited from page level: same as line or word level (no @orientation), |
||
782 | # - zero (unrotate page angle): different from line or word level (because |
||
783 | # otherwise deskewing would never have an effect on lines and words) |
||
784 | # The PAGE specification is silent here (but does generally not concern itself |
||
785 | # much with AlternativeImage coordinate consistency). |
||
786 | # Since our (generateDS-backed) ocrd_page supports the zero/none distinction, |
||
787 | # we choose the former (i.e. None is inheritance). |
||
788 | if 'orientation' in segment.__dict__ and segment.get_orientation() is not None: |
||
789 | # region angle: PAGE @orientation is defined clockwise, |
||
790 | # whereas PIL/ndimage rotation is in mathematical direction: |
||
791 | angle = -segment.get_orientation() |
||
792 | # @orientation is always absolute; if higher levels |
||
793 | # have already rotated, then we must compensate: |
||
794 | angle -= parent_coords['angle'] |
||
795 | # map angle from (-180,180] to [0,360], and partition into multiples of 90; |
||
796 | # but avoid unnecessary large remainders, i.e. split symmetrically: |
||
797 | orientation = (angle + 45) % 360 |
||
798 | orientation = orientation - (orientation % 90) |
||
799 | skew = (angle % 360) - orientation |
||
800 | skew = 180 - (180 - skew) % 360 # map to [-45,45] |
||
801 | log.debug("segment '%s' has orientation=%d skew=%.2f", |
||
802 | segment.id, orientation, skew) |
||
803 | else: |
||
804 | orientation = 0 |
||
805 | skew = 0 |
||
806 | segment_coords['angle'] = parent_coords['angle'] # nothing applied yet (depends on filters) |
||
807 | |||
808 | # initialize AlternativeImage@comments classes from parent, except |
||
809 | # for those operations that can apply on multiple hierarchy levels: |
||
810 | segment_coords['features'] = ','.join( |
||
811 | [feature for feature in parent_coords['features'].split(',') |
||
812 | if feature in ['binarized', 'grayscale_normalized', |
||
813 | 'despeckled', 'dewarped']]) |
||
814 | |||
815 | alternative_image = None |
||
816 | alternative_images = segment.get_AlternativeImage() |
||
817 | View Code Duplication | if alternative_images: |
|
818 | # (e.g. from segment-level cropping, binarization, deskewing or despeckling) |
||
819 | if feature_selector or feature_filter: |
||
820 | alternative_image = None |
||
821 | # search from the end, because by convention we always append, |
||
822 | # and among multiple satisfactory images we want the most recent: |
||
823 | for alternative_image in reversed(alternative_images): |
||
824 | features = alternative_image.get_comments() |
||
825 | if not features: |
||
826 | log.warning("AlternativeImage %d for segment '%s' does not have any feature attributes", |
||
827 | alternative_images.index(alternative_image) + 1, segment.id) |
||
828 | features = '' |
||
829 | if (all(feature in features |
||
830 | for feature in feature_selector.split(',') if feature) and |
||
831 | not any(feature in features |
||
832 | for feature in feature_filter.split(',') if feature)): |
||
833 | break |
||
834 | else: |
||
835 | alternative_image = None |
||
836 | else: |
||
837 | alternative_image = alternative_images[-1] |
||
838 | features = alternative_image.get_comments() |
||
839 | if not features: |
||
840 | log.warning("AlternativeImage %d for segment '%s' does not have any feature attributes", |
||
841 | alternative_images.index(alternative_image) + 1, segment.id) |
||
842 | features = '' |
||
843 | if alternative_image: |
||
844 | log.debug("Using AlternativeImage %d (%s) for segment '%s'", |
||
845 | alternative_images.index(alternative_image) + 1, |
||
846 | features, segment.id) |
||
847 | segment_image = self._resolve_image_as_pil(alternative_image.get_filename()) |
||
848 | segment_coords['features'] = features |
||
849 | |||
850 | alternative_image_features = segment_coords['features'].split(',') |
||
851 | for duplicate_feature in set([feature for feature in alternative_image_features |
||
852 | # features relevant in reconstructing coordinates: |
||
853 | if (feature in ['deskewed', 'rotated-90', |
||
854 | 'rotated-180', 'rotated-270'] and |
||
855 | alternative_image_features.count(feature) > 1)]): |
||
856 | log.error("Duplicate feature %s in AlternativeImage for segment '%s'", |
||
857 | duplicate_feature, segment.id) |
||
858 | for i, feature in enumerate(alternative_image_features + |
||
859 | (['rotated-%d' % orientation] |
||
860 | if (orientation and |
||
861 | not 'rotated-%d' % orientation in alternative_image_features and |
||
862 | not 'rotated-%d' % orientation in feature_filter.split(',')) |
||
863 | else []) + |
||
864 | (['deskewed'] |
||
865 | if (skew and |
||
866 | not 'deskewed' in alternative_image_features and |
||
867 | not 'deskewed' in feature_filter.split(',')) |
||
868 | else []) + |
||
869 | # not a feature to be added, but merely as a fallback position |
||
870 | # to always enter loop at i == len(alternative_image_features) |
||
871 | ['_check']): |
||
872 | # image geometry vs feature consistency can only be checked |
||
873 | # after all features on the existing AlternativeImage have |
||
874 | # been adjusted for in the transform, and when there is a mismatch, |
||
875 | # additional steps applied here would only repeat the respective |
||
876 | # error message; so we only check once at the boundary between |
||
877 | # existing and new features |
||
878 | # FIXME we should enforce consistency here (i.e. split into transposition |
||
879 | # and minimal rotation, rotation always reshapes, rescaling never happens) |
||
880 | # FIXME: inconsistency currently unavoidable with line-level dewarping (which increases height) |
||
881 | if (i == len(alternative_image_features) and |
||
882 | not (segment_xywh['w'] - 2 < segment_image.width < segment_xywh['w'] + 2 and |
||
883 | segment_xywh['h'] - 2 < segment_image.height < segment_xywh['h'] + 2)): |
||
884 | log.error('segment "%s" image (%s; %dx%d) has not been cropped properly (%dx%d)', |
||
885 | segment.id, segment_coords['features'], |
||
886 | segment_image.width, segment_image.height, |
||
887 | segment_xywh['w'], segment_xywh['h']) |
||
888 | name = "%s for segment '%s'" % ("AlternativeImage" if alternative_image |
||
889 | else "parent image", segment.id) |
||
890 | # adjust transform to feature, and ensure feature is applied to image |
||
891 | if feature == 'rotated-%d' % orientation: |
||
892 | segment_image, segment_coords, segment_xywh = _reflect( |
||
893 | log, name, orientation, segment_image, segment_coords, segment_xywh) |
||
894 | elif feature == 'deskewed': |
||
895 | segment_image, segment_coords, segment_xywh = _rotate( |
||
896 | log, name, skew, segment, segment_image, segment_coords, segment_xywh, |
||
897 | fill=fill, transparency=transparency) |
||
898 | |||
899 | # verify constraints again: |
||
900 | if not all(feature in segment_coords['features'] |
||
901 | for feature in feature_selector.split(',') if feature): |
||
902 | raise Exception('Found no AlternativeImage that satisfies all requirements' + |
||
903 | 'selector="%s" in segment "%s"' % ( |
||
904 | feature_selector, segment.id)) |
||
905 | if any(feature in segment_coords['features'] |
||
906 | for feature in feature_filter.split(',') if feature): |
||
907 | raise Exception('Found no AlternativeImage that satisfies all requirements ' + |
||
908 | 'filter="%s" in segment "%s"' % ( |
||
909 | feature_filter, segment.id)) |
||
910 | segment_image.format = 'PNG' # workaround for tesserocr#194 |
||
911 | return segment_image, segment_coords |
||
912 | |||
913 | # pylint: disable=redefined-builtin |
||
914 | def save_image_file(self, image, |
||
915 | file_id, |
||
916 | file_grp, |
||
917 | page_id=None, |
||
918 | mimetype='image/png', |
||
919 | force=False): |
||
920 | """Store an image in the filesystem and reference it as new file in the METS. |
||
921 | |||
922 | Args: |
||
923 | image (PIL.Image): derived image to save |
||
924 | file_id (string): `@ID` of the METS `file` to use |
||
925 | file_grp (string): `@USE` of the METS `fileGrp` to use |
||
926 | Keyword Args: |
||
927 | page_id (string): `@ID` in the METS physical `structMap` to use |
||
928 | mimetype (string): MIME type of the image format to serialize as |
||
929 | force (boolean): whether to replace any existing `file` with that `@ID` |
||
930 | |||
931 | Serialize the image into the filesystem, and add a `file` for it in the METS. |
||
932 | Use a filename extension based on ``mimetype``. |
||
933 | |||
934 | Returns: |
||
935 | The (absolute) path of the created file. |
||
936 | """ |
||
937 | log = getLogger('ocrd.workspace.save_image_file') |
||
938 | if not force and self.overwrite_mode: |
||
939 | force = True |
||
940 | image_bytes = io.BytesIO() |
||
941 | image.save(image_bytes, format=MIME_TO_PIL[mimetype]) |
||
942 | file_path = str(Path(file_grp, '%s%s' % (file_id, MIME_TO_EXT[mimetype]))) |
||
943 | out = self.add_file( |
||
944 | ID=file_id, |
||
945 | file_grp=file_grp, |
||
946 | pageId=page_id, |
||
947 | local_filename=file_path, |
||
948 | mimetype=mimetype, |
||
949 | content=image_bytes.getvalue(), |
||
950 | force=force) |
||
951 | log.info('created file ID: %s, file_grp: %s, path: %s', |
||
952 | file_id, file_grp, out.local_filename) |
||
953 | return file_path |
||
954 | |||
955 | def _crop(log, name, segment, parent_image, parent_coords, **kwargs): |
||
956 | segment_coords = parent_coords.copy() |
||
957 | # get polygon outline of segment relative to parent image: |
||
958 | segment_polygon = coordinates_of_segment(segment, parent_image, parent_coords) |
||
959 | # get relative bounding box: |
||
960 | segment_bbox = bbox_from_polygon(segment_polygon) |
||
961 | # get size of the segment in the parent image after cropping |
||
962 | # (i.e. possibly different from size before rotation at the parent, but |
||
963 | # also possibly different from size after rotation below/AlternativeImage): |
||
964 | segment_xywh = xywh_from_bbox(*segment_bbox) |
||
965 | # crop, if (still) necessary: |
||
966 | if (not isinstance(segment, BorderType) or # always crop below page level |
||
967 | not 'cropped' in parent_coords['features']): |
||
968 | if isinstance(segment, BorderType): |
||
969 | log.info("Cropping %s", name) |
||
970 | segment_coords['features'] += ',cropped' |
||
971 | # create a mask from the segment polygon: |
||
972 | segment_image = image_from_polygon(parent_image, segment_polygon, **kwargs) |
||
973 | # crop to bbox: |
||
974 | segment_image = crop_image(segment_image, box=segment_bbox) |
||
975 | else: |
||
976 | segment_image = parent_image |
||
977 | # subtract offset from parent in affine coordinate transform: |
||
978 | # (consistent with image cropping) |
||
979 | segment_coords['transform'] = shift_coordinates( |
||
980 | parent_coords['transform'], |
||
981 | np.array([-segment_bbox[0], |
||
982 | -segment_bbox[1]])) |
||
983 | return segment_image, segment_coords, segment_xywh |
||
984 | |||
985 | def _reflect(log, name, orientation, segment_image, segment_coords, segment_xywh): |
||
986 | # Transpose in affine coordinate transform: |
||
987 | # (consistent with image transposition or AlternativeImage below) |
||
988 | transposition = { |
||
989 | 90: Image.ROTATE_90, |
||
990 | 180: Image.ROTATE_180, |
||
991 | 270: Image.ROTATE_270 |
||
992 | }.get(orientation) # no default |
||
993 | segment_coords['transform'] = transpose_coordinates( |
||
994 | segment_coords['transform'], transposition, |
||
995 | np.array([0.5 * segment_xywh['w'], |
||
996 | 0.5 * segment_xywh['h']])) |
||
997 | segment_xywh['w'], segment_xywh['h'] = adjust_canvas_to_transposition( |
||
998 | [segment_xywh['w'], segment_xywh['h']], transposition) |
||
999 | segment_coords['angle'] += orientation |
||
1000 | # transpose, if (still) necessary: |
||
1001 | if not 'rotated-%d' % orientation in segment_coords['features']: |
||
1002 | log.info("Transposing %s by %d°", name, orientation) |
||
1003 | segment_image = transpose_image(segment_image, transposition) |
||
1004 | segment_coords['features'] += ',rotated-%d' % orientation |
||
1005 | return segment_image, segment_coords, segment_xywh |
||
1006 | |||
1007 | def _rotate(log, name, skew, segment, segment_image, segment_coords, segment_xywh, **kwargs): |
||
1008 | # Rotate around center in affine coordinate transform: |
||
1009 | # (consistent with image rotation or AlternativeImage below) |
||
1010 | segment_coords['transform'] = rotate_coordinates( |
||
1011 | segment_coords['transform'], skew, |
||
1012 | np.array([0.5 * segment_xywh['w'], |
||
1013 | 0.5 * segment_xywh['h']])) |
||
1014 | segment_xywh['w'], segment_xywh['h'] = adjust_canvas_to_rotation( |
||
1015 | [segment_xywh['w'], segment_xywh['h']], skew) |
||
1016 | segment_coords['angle'] += skew |
||
1017 | # deskew, if (still) necessary: |
||
1018 | if not 'deskewed' in segment_coords['features']: |
||
1019 | log.info("Rotating %s by %.2f°", name, skew) |
||
1020 | segment_image = rotate_image(segment_image, skew, **kwargs) |
||
1021 | segment_coords['features'] += ',deskewed' |
||
1022 | if (segment and |
||
1023 | (not isinstance(segment, BorderType) or # always crop below page level |
||
1024 | 'cropped' in segment_coords['features'])): |
||
1025 | # re-crop to new bbox (which may deviate |
||
1026 | # if segment polygon was not a rectangle) |
||
1027 | segment_image, segment_coords, segment_xywh = _crop( |
||
1028 | log, name, segment, segment_image, segment_coords, |
||
1029 | **kwargs) |
||
1030 | elif (segment and |
||
1031 | (not isinstance(segment, BorderType) or # always crop below page level |
||
1032 | 'cropped' in segment_coords['features'])): |
||
1033 | # only shift coordinates as if re-cropping |
||
1034 | _, segment_coords, segment_xywh = _crop( |
||
1035 | log, name, segment, segment_image, segment_coords, |
||
1036 | **kwargs) |
||
1037 | return segment_image, segment_coords, segment_xywh |
||
1038 |