From 5a27e46b2282c68b2dc156110059c9ec197d9e7f Mon Sep 17 00:00:00 2001
From: Robert Sachunsky <sachunsky@informatik.uni-leipzig.de>
Date: Thu, 16 Apr 2026 04:56:38 +0200
Subject: [PATCH] =?UTF-8?q?keep=20seps=20over=20artificial=20boundaries=20?=
 =?UTF-8?q?to=20improve=20col=20separation=E2=80=A6?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

(thresholding and decoding with artificial boundary class can
 overwrite existing column separators, which in turn can contribute
 to missing column boundaries; this prioritises seps over boundaries,
 which does not impair separation of instances, as seps will separate
 text/image/etc instances just as well as artificial boundaries)
---
 src/eynollah/eynollah.py       | 12 ++++++++----
 src/eynollah/utils/__init__.py |  8 +++++++-
 2 files changed, 15 insertions(+), 5 deletions(-)

diff --git a/src/eynollah/eynollah.py b/src/eynollah/eynollah.py
index d564f87..180a12a 100644
--- a/src/eynollah/eynollah.py
+++ b/src/eynollah/eynollah.py
@@ -628,6 +628,7 @@ class Eynollah:
             thresholding_for_artificial_class=False,
             threshold_art_class=0.1,
             artificial_class=4,
+            separator_class=0,
     ):
 
         self.logger.debug("enter do_prediction_new_concept (patches=%d)", patches)
@@ -653,7 +654,8 @@ class Eynollah:
                                label=artificial_class,
                                only=True,
                                skeletonize=True,
-                               dilate=3)
+                               dilate=3,
+                               keep=separator_class)
 
             conf_text = resize_image(label_p_pred[:, :, 1], img_h_page, img_w_page)
             return prediction, conf_text
@@ -799,7 +801,8 @@ class Eynollah:
                            label=artificial_class,
                            only=True,
                            skeletonize=True,
-                           dilate=3)
+                           dilate=3,
+                           keep=separator_class)
         gc.collect()
         return prediction, confidence
 
@@ -1071,7 +1074,7 @@ class Eynollah:
         self.logger.debug('enter textline_contours')
 
         kwargs = dict(artificial_class=2,
-                      n_batch_inference=3,
+                      n_batch_inference=1 if self.tables else 3, # avoid OOM
                       thresholding_for_artificial_class=True,
                       threshold_art_class=self.threshold_art_class_textline)
         if use_patches:
@@ -1131,7 +1134,8 @@ class Eynollah:
         #print("inside 2 ", time.time()-t_in)
         kwargs = dict(n_batch_inference=1,
                       thresholding_for_artificial_class=True,
-                      threshold_art_class=self.threshold_art_class_layout)
+                      threshold_art_class=self.threshold_art_class_layout,
+                      separator_class=3)
         if num_col_classifier == 1 or num_col_classifier == 2:
             if img_height_h / img_width_h > 2.5:
                 self.logger.debug("resized to %dx%d for %d cols",
diff --git a/src/eynollah/utils/__init__.py b/src/eynollah/utils/__init__.py
index f8638db..ef27f24 100644
--- a/src/eynollah/utils/__init__.py
+++ b/src/eynollah/utils/__init__.py
@@ -1953,7 +1953,8 @@ def seg_mask_label(segmap:np.ndarray,
                    only:bool=False,
                    label:int=2,
                    skeletonize:bool=False,
-                   dilate:int=0
+                   dilate:int=0,
+                   keep:int=0,
 ) -> None:
     """
     overwrite an existing segmentation map from a binary mask with a given label
@@ -1966,11 +1967,14 @@ def seg_mask_label(segmap:np.ndarray,
         only: whether to suppress the `label` outside `mask`
         skeletonize: whether to transform the mask to its skeleton
         dilate: whether to also apply dilatation after this (convolution with square kernel of given size)
+        keep: if nonzero, a clas label to be kept untouched
 
     Use this to enforce specific confidence thresholds or rules after segmentation.
     """
     if not mask.any():
         return
+    if keep:
+        keepmask = segmap == keep
     if only:
         segmap[segmap == label] = 0
     if skeletonize:
@@ -1982,3 +1986,5 @@ def seg_mask_label(segmap:np.ndarray,
             kernel = np.ones((dilate, dilate), np.uint8)
             mask = cv2.dilate(mask.astype(np.uint8), kernel, iterations=1) > 0
     segmap[mask] = label
+    if keep:
+        segmap[keepmask] = keep