ignore spurious text in clipping detection

microqa/engine.py

@@ -1,3 +1,4 @@
+import re
 from sys import stdout
 from typing import Optional
 
@@ -243,23 +244,22 @@ def compute_text_margin_px(
         Integer pixel count if text is present; otherwise `None`.
     """
 
-    word_margins_all_directions = (
+    block_margins = [
-        np.sort(
+        np.array(
-            np.concat(
+            [
-                [
+                block.x0,
-                    np.array(
+                block.y0,
-                        [
+                im_size[0] - block.x1,
-                            block.x0,
+                im_size[1] - block.y1,
-                            block.y0,
+            ]
-                            im_size[0] - block.x1,
-                            im_size[1] - block.y1,
-                        ]
-                    )
-                    for block in text_blocks
-                ]
-            ).astype(np.int_)
         )
-        if len(text_blocks) > 0
+        for block in text_blocks
+        # Exclude text without clear alphanumeric substance.
+        if len(re.sub(r"[^a-zA-Z0-9]", "", block.text)) > 1
+    ]
+    word_margins_all_directions = (
+        np.sort(np.concat(block_margins).astype(np.int_))
+        if len(block_margins) > 0
         else np.array([])
     )
     # Skip the n closest words to the edge, to help ignore stray OCR artifacts.