new version 2.5 (#539)

* Added simple multilingual textual layer correctness classification based on letter percentage calculation (`textual_layer_classifier=letter`).
* Added a new parameter `textual_layer_classifier = [simple, ml (default), letter]`.
* Remove parameter `fast_textual_layer_detection`. Now it is a `textual_layer_classifier=simple`.
* Fix bug with `table_type=table_wo_external_bounds` (fixed cv2.BoundingRect).
* Some refactoring `TableRecognition`.
* Added parameter `table_type` and `TableRecognition` info into documentation.

Author: oksidgy

VERSION

@@ -1 +1 @@
-2.4
+2.5

dedoc/api/api_args.py

@@ -26,8 +26,7 @@ class QueryParameters:
     # pdf handling
     pdf_with_text_layer: str = Form("auto_tabby", enum=["true", "false", "auto", "auto_tabby", "tabby", "bad_encoding"],
                                     description="Extract text from a text layer of PDF or using OCR methods for image-like documents")
-    fast_textual_layer_detection: str = Form("false", enum=["true", "false"],
-                                             description="Use non-ML solution to detect textual layer. Much faster but less accurate.")
+    textual_layer_classifier: str = Form("ml", enum=["ml", "simple", "letter"], description="Type of classifier for PDF textual layer detection")
     each_page_textual_layer_detection: str = Form("false", enum=["true", "false"], description="Detect textual layer on each page. Slower but more accurate.")
     language: str = Form("rus+eng", description="Recognition language ('rus+eng', 'rus', 'eng', 'fra', 'spa')")
     pages: str = Form(":", description='Page numbers range for reading PDF or images, "left:right" means read pages from left to right')

dedoc/api/web/index.html

@@ -13,6 +13,7 @@
         details > summary {font-style: italic; cursor: pointer; display: list-item;}
         .child.max {padding-left: 5px; flex: 1}
         .parent {display: flex}
+        details { padding-left: 24px;}
     </style>
 </head>
 
@@ -100,7 +101,7 @@ <h4>Attachments handling</h4>
 
         <div class="parameters">
             <h4>PDF handling</h4>
-            <details><summary>pdf_with_text_layer, need_pdf_table_analysis, fast_textual_layer_detection, each_page_textual_layer_detection, language, pages, is_one_column_document, document_orientation, need_header_footer_analysis, need_binarization, need_gost_frame_analysis</summary>
+            <details><summary>pdf_with_text_layer, need_pdf_table_analysis, textual_layer_classifier, each_page_textual_layer_detection, language, pages, is_one_column_document, document_orientation, need_header_footer_analysis, need_binarization, need_gost_frame_analysis</summary>
                 <br>
                 <p>
                     <label>
@@ -116,7 +117,13 @@ <h4>PDF handling</h4>
                 </p>
 
                 <p>
-                    <label><input name="fast_textual_layer_detection" type="checkbox" value="true"> fast_textual_layer_detection</label>
+                    <label>
+                        <select name="textual_layer_classifier">
+                            <option value="ml">ml</option>
+                            <option value="simple">simple</option>
+                            <option value="letter">letter</option>
+                        </select> textual_layer_classifier
+                    </label>
                 </p>
 
                 <p>
@@ -136,15 +143,18 @@ <h4>PDF handling</h4>
                     </label>
                 </p>
 
-                <details><summary>need_pdf_table_analysis</summary>
+                <details><summary>need_pdf_table_analysis, table_type</summary>
                 <br>
                     <p>
                     <label>
                         <input type="hidden" name="need_pdf_table_analysis" value="false">
                         <input type="checkbox" name="need_pdf_table_analysis" value="true" checked> need_pdf_table_analysis</label>
                     </p>
+                    <p>
+                        <label>table_type <input name="table_type" type="text" size="20" value=""></label>
+                    </p>
                 </details>
-
+                <br>
                 <p>
                     <label>pages <input name="pages" type="text" size="8" value=":"></label>
                 </p>

dedoc/readers/pdf_reader/data_classes/tables/scantable.py

@@ -10,6 +10,10 @@
 
 
 class ScanTable(Table):
+    """
+    Utility class for storing recognized tables from document images. The class
+    :class:`~dedoc.readers.pdf_reader.pdf_image_reader.table_recognizer.table_recognizer.TableRecognizer` works with this class.
+    """
     def __init__(self, page_number: int, cells: List[List[CellWithMeta]], bbox: BBox, order: int = -1) -> None:
 
         super().__init__(cells, TableMetadata(page_id=page_number))

dedoc/readers/pdf_reader/data_classes/tables/table_tree.py

@@ -1,6 +1,7 @@
 from collections import namedtuple
 from typing import List, Optional
 
+import numpy as np
 from dedocutils.data_structures import BBox
 from numpy import ndarray
 
@@ -79,7 +80,7 @@ def parse_contours_to_tree(contours: List, hierarchy: List, *, config: dict) ->
         if len(contours) == 0:
             return table_tree
 
-        bbox = [cv2.boundingRect(c) for c in contours[0]][0]   # [x_begin, y_begin, width, height]
+        bbox = cv2.boundingRect(contours[0].astype(np.int32))   # [x_begin, y_begin, width, height]
         table_tree.cell_box = BBox(x_top_left=bbox[0], y_top_left=bbox[1], width=bbox[2], height=bbox[3])
 
         table_tree = table_tree.__build_childs(table_tree, hierarchy, contours)
@@ -101,7 +102,7 @@ def __build_childs(self, cur: "TableTree", hierarchy: List, contours: List) -> "
         list_childs = []
         for i, h in enumerate(hierarchy[0]):
             if h[3] == cur.id_contours:
-                bbox = cv2.boundingRect(contours[i])  # [x_begin, y_begin, width, height]
+                bbox = cv2.boundingRect(contours[i].astype(np.int32))  # [x_begin, y_begin, width, height]
                 # Эвристика №1 на ячейку
                 if bbox[2] < self.min_w_cell or bbox[3] < self.min_h_cell:
                     if self.config.get("debug_mode", False):

dedoc/readers/pdf_reader/data_classes/tables/table_type.py

@@ -1,4 +1,61 @@
 class TableTypeAdditionalOptions:
+    """
+    Setting up the table recognizer. The value of the parameter specifies the type of tables recognized when processed by
+    class :class:`~dedoc.readers.pdf_reader.pdf_image_reader.table_recognizer.table_recognizer.TableRecognizer`.
+
+    * Parameter `table_type=wo_external_bounds` - recognize tables without external bounds;
+
+    Example of a table of type `wo_external_bounds`::
+
+       text   | text | text
+      --------+------+------
+       text   | text | text
+      --------+------+------
+       text   | text | text
+      --------+------+------
+       text   | text | text
+
+
+    * Parameter `table_type=one_cell_table` - if a document contains a bounding box with text, it will be considered a table;
+
+    Example of a page with a table of type `one_cell_table`::
+
+         _________________________
+         Header of document
+         text text text +------+
+         text           | text |  <--- it is a table
+                        +------+
+         ________________________
+
+    * Parameter `table_type=split_last_column` - specified parameter for the merged last column of the table;
+
+    Example of a table of type `split_last_column`::
+
+         +--------+------+-------+
+         | text   | text | text1 |
+         +--------+------+       |
+         | text0  | text | text2 |
+         |        | -----|       |
+         |        | text | text3 |
+         +--------+------+       |
+         | text   | text | text4 |
+         +--------+------+-------+
+                     |
+                 Recognition
+                    |
+                    V
+         +--------+------+-------+
+         | text   | text | text1 |
+         +--------+------+-------|
+         | text0  | text | text2 |
+         |--------+ -----+------ |
+         | text0  | text | text3 |
+         +--------+------+------ |
+         | text   | text | text4 |
+         +--------+------+-------+
+
+    """
+
     def __init__(self) -> None:
         self.table_wo_external_bounds = "wo_external_bounds"
         self.detect_one_cell_table = "one_cell_table"

dedoc/readers/pdf_reader/pdf_auto_reader/txtlayer_classifier/__init__.py

@@ -0,0 +1,14 @@
+from typing import Dict
+
+from .abstract_txtlayer_classifier import AbstractTxtlayerClassifier
+from .letter_txtlayer_classifier import LetterTxtlayerClassifier
+from .ml_txtlayer_classifier import MlTxtlayerClassifier
+from .simple_txtlayer_classifier import SimpleTxtlayerClassifier
+
+
+def get_classifiers(config: dict) -> Dict[str, AbstractTxtlayerClassifier]:
+    return {
+        "ml": MlTxtlayerClassifier(config=config),
+        "simple": SimpleTxtlayerClassifier(config=config),
+        "letter": LetterTxtlayerClassifier(config=config)
+    }

dedoc/readers/pdf_reader/pdf_auto_reader/txtlayer_classifier/letter_txtlayer_classifier.py

@@ -0,0 +1,28 @@
+from typing import List
+
+import numpy as np
+
+from dedoc.data_structures.line_with_meta import LineWithMeta
+from dedoc.readers.pdf_reader.pdf_auto_reader.txtlayer_classifier.abstract_txtlayer_classifier import AbstractTxtlayerClassifier
+
+
+class LetterTxtlayerClassifier(AbstractTxtlayerClassifier):
+    """
+    Simple multilingual textual layer correctness classification.
+    Textual layer is considered as correct if percent of letters in the text > 50%.
+    """
+    def __init__(self, *, config: dict) -> None:
+        super().__init__(config=config)
+        self.__symbol_threshold = 0.5
+
+    def predict(self, lines: List[List[LineWithMeta]]) -> np.ndarray:
+        texts = np.array(["".join(line.line for line in line_list) for line_list in lines])
+        result = np.array([bool(text.strip()) for text in texts])
+        ids_for_pred = np.where(result)[0]
+
+        for idx in ids_for_pred:
+            text = texts[idx].replace(".", "").replace("…", "")
+            letters_number = sum(1 for symbol in text if symbol.isalpha())
+            result[idx] = letters_number / max(len(text), 1) > self.__symbol_threshold
+
+        return result

dedoc/readers/pdf_reader/pdf_auto_reader/txtlayer_detector.py

@@ -7,9 +7,8 @@
 import numpy as np
 
 from dedoc.data_structures.unstructured_document import UnstructuredDocument
+from dedoc.readers.pdf_reader.pdf_auto_reader.txtlayer_classifier import get_classifiers
 from dedoc.readers.pdf_reader.pdf_auto_reader.txtlayer_classifier.abstract_txtlayer_classifier import AbstractTxtlayerClassifier
-from dedoc.readers.pdf_reader.pdf_auto_reader.txtlayer_classifier.ml_txtlayer_classifier import MlTxtlayerClassifier
-from dedoc.readers.pdf_reader.pdf_auto_reader.txtlayer_classifier.simple_txtlayer_classifier import SimpleTxtlayerClassifier
 from dedoc.readers.pdf_reader.pdf_auto_reader.txtlayer_result import TxtLayerResult
 from dedoc.readers.pdf_reader.pdf_txtlayer_reader.pdf_tabby_reader import PdfTabbyReader
 from dedoc.utils.parameter_utils import get_bool_parameter, get_param_page_slice
@@ -22,8 +21,7 @@ def __init__(self, pdf_reader: PdfTabbyReader, *, config: dict) -> None:
         self.config = config
         self.logger = config.get("logger", logging.getLogger())
 
-        self.ml_txtlayer_classifier = MlTxtlayerClassifier(config=config)
-        self.simple_txtlayer_classifier = SimpleTxtlayerClassifier(config=config)
+        self.classifiers = get_classifiers(config=config)
         self.pdf_reader = pdf_reader
 
     def detect_txtlayer(self, path: str, parameters: dict) -> List[TxtLayerResult]:
@@ -34,10 +32,11 @@ def detect_txtlayer(self, path: str, parameters: dict) -> List[TxtLayerResult]:
         :param parameters: parameters for the txtlayer classifier
         :return: information about a textual layer in the PDF document
         """
-        if get_bool_parameter(parameters, "fast_textual_layer_detection", False):
-            txtlayer_classifier = self.simple_txtlayer_classifier
-        else:
-            txtlayer_classifier = self.ml_txtlayer_classifier
+        classifier_name = str(parameters.get("textual_layer_classifier", "ml")).lower()
+        txtlayer_classifier = self.classifiers.get(classifier_name)
+        if txtlayer_classifier is None:
+            raise ValueError(f"Unknown textual layer classifier `{classifier_name}`")
+
         classify_each_page = get_bool_parameter(parameters, "each_page_textual_layer_detection", False)
         detect_function = self.__classify_each_page if classify_each_page else self.__classify_all_pages
         try:
@@ -110,7 +109,10 @@ def __classify_each_page(self, path: str, parameters: dict, txtlayer_classifier:
         prev_idx = 0
         for transition_idx in transitions:
             chunk_lines = list(chain.from_iterable(lines_for_predict[prev_idx:transition_idx]))
-            chunk_document = UnstructuredDocument(lines=chunk_lines, tables=document.tables, attachments=document.attachments)
+            if is_correct:
+                chunk_document = UnstructuredDocument(lines=chunk_lines, tables=document.tables, attachments=document.attachments)
+            else:
+                chunk_document = None
             chunk_result = TxtLayerResult(start=prev_idx + fisrt_page_id + 1, end=transition_idx + fisrt_page_id, correct=is_correct, document=chunk_document)
             result.append(chunk_result)
             is_correct = not is_correct

dedoc/readers/pdf_reader/pdf_image_reader/pdf_image_reader.py

@@ -53,6 +53,7 @@ def __init__(self, *, config: Optional[dict] = None) -> None:
                                                                           config=self.config)
         self.binarizer = AdaptiveBinarizer()
         self.ocr = OCRLineExtractor(config=self.config)
+        self.page_number = None
 
     def read(self, file_path: str, parameters: Optional[dict] = None) -> UnstructuredDocument:
         return super().read(file_path, parameters)
@@ -68,6 +69,7 @@ def _process_one_page(self,
         from dedoc.utils.parameter_utils import get_path_param
 
         #  --- Step 1: correct orientation and detect column count ---
+        self.page_number = page_number
         rotated_image, is_one_column_document, angle = self._detect_column_count_and_orientation(image, parameters)
         if self.config.get("debug_mode", False):
             self.logger.info(f"Angle page rotation = {angle}")
@@ -105,7 +107,6 @@ def _detect_column_count_and_orientation(self, image: ndarray, parameters: Param
         Return: rotated_image and indicator if the page is one-column
         """
         import os
-        from datetime import datetime
         import cv2
         from dedoc.utils.parameter_utils import get_path_param
 
@@ -124,7 +125,7 @@ def _detect_column_count_and_orientation(self, image: ndarray, parameters: Param
 
         if self.config.get("debug_mode", False):
             debug_dir = get_path_param(self.config, "path_debug")
-            img_path = os.path.join(debug_dir, f"{datetime.now().strftime('%H-%M-%S')}_result_orientation.jpg")
+            img_path = os.path.join(debug_dir, f"page-{self.page_number}_result_orientation.jpg")
             self.logger.info(f"Save image to {img_path}")
             cv2.imwrite(img_path, rotated_image)