drop lines and spans from the provider if we detect bad text [skip ci]

marker/schema/page.py

@@ -3,7 +3,7 @@
 
 from marker.schema.bbox import BboxElement
 from marker.schema.block import Block, Span
-from surya.schema import TextDetectionResult, LayoutResult, OrderResult
+from surya.schema import TextDetectionResult, LayoutResult
 
 
 class Page(BboxElement):
@@ -12,7 +12,7 @@ class Page(BboxElement):
     rotation: Optional[int] = None # Rotation degrees of the page
     text_lines: Optional[TextDetectionResult] = None
     layout: Optional[LayoutResult] = None
-    order: Optional[OrderResult] = None
+    order: Optional[Any] = None
     ocr_method: Optional[str] = None # One of "surya" or "tesseract"
     char_blocks: Optional[List[Dict]] = None # Blocks with character-level data from pdftext
     images: Optional[List[Any]] = None # Images to save along with the page, need Any to avoid pydantic error

marker/v2/converters/pdf.py

@@ -1,3 +1,5 @@
+from marker.v2.providers.pdf import PdfProvider
+
 import tempfile
 from typing import List, Optional
 
@@ -10,7 +12,6 @@
 from marker.v2.converters import BaseConverter
 from marker.v2.processors.equation import EquationProcessor
 from marker.v2.processors.table import TableProcessor
-from marker.v2.providers.pdf import PdfProvider
 from marker.v2.models import setup_layout_model, setup_texify_model, setup_recognition_model, setup_table_rec_model, \
     setup_detection_model
 
@@ -36,8 +37,8 @@ def __call__(self, filepath: str, page_range: List[int] | None = None):
         equation_processor(document)
 
         # TODO: re-enable once we add OCR method
-        #table_processor = TableProcessor(self.detection_model, self.recognition_model, self.table_rec_model)
-        #table_processor(document)
+        # table_processor = TableProcessor(self.detection_model, self.recognition_model, self.table_rec_model)
+        # table_processor(document)
 
         rendered = document.render()
         return rendered
@@ -55,5 +56,3 @@ def __call__(self, filepath: str, page_range: List[int] | None = None):
         rendered = converter(temp_pdf.name)
 
         print(rendered)
-
-

marker/v2/providers/pdf.py

@@ -1,12 +1,12 @@
 import functools
-from typing import Dict, List, Tuple
-from typing import Dict, List, Optional
+from typing import Dict, List, Optional, Tuple
 
 import pypdfium2 as pdfium
 from pdftext.extraction import dictionary_output
 from PIL import Image
 from pydantic import BaseModel
 
+from marker.ocr.heuristics import detect_bad_ocr
 from marker.v2.providers import BaseProvider
 from marker.v2.schema.polygon import PolygonBox
 from marker.v2.schema.text.line import Line, Span
@@ -110,7 +110,22 @@ def setup(self):
                         )
                     )
                     line_spans.append(spans)
-            self.page_lines[page_id] = (lines, line_spans)
+            if self.check_line_spans(line_spans):
+                self.page_lines[page_id] = (lines, line_spans)
+
+    def check_line_spans(self, line_spans_list: List[List[Span]]) -> bool:
+        if not len(sum(line_spans_list, [])):
+            return False
+        text = ""
+        for line_spans in line_spans_list:
+            for span in line_spans:
+                text = text + " " + span.text
+            text = text + "\n"
+        if len(text.strip()) == 0:
+            return False
+        if detect_bad_ocr(text):
+            return False
+        return True
 
     @ functools.lru_cache(maxsize=None)
     def get_image(self, idx: int, dpi: int) -> Image.Image: