From c43db703add42e6d47a97abd8876a2c0135ea995 Mon Sep 17 00:00:00 2001 From: Yizhan Date: Tue, 11 Nov 2025 22:20:52 -0800 Subject: [PATCH] =?UTF-8?q?fix:=20=E4=BF=AE=E5=A4=8DTableRecognitionPipeli?= =?UTF-8?q?neV2=E4=B8=ADrec=5Ftexts=E5=92=8Crec=5Fscores=E9=95=BF=E5=BA=A6?= =?UTF-8?q?=E4=B8=8D=E4=B8=80=E8=87=B4=E7=9A=84=E9=97=AE=E9=A2=98?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 问题描述： 当OCR文本框跨越多个表格单元格需要拆分时，split_ocr_bboxes_by_table_cells方法 只更新了rec_texts列表，但没有同步更新rec_scores列表，导致两者长度不一致。 修复内容： - 在split_ocr_bboxes_by_table_cells方法中添加对rec_scores的处理 - 拆分OCR框时，为每个拆分后的框提取并保存对应的置信度分数 - 未拆分时，保留原始置信度分数 - 确保rec_texts和rec_scores始终保持相同长度 修复文件： - paddlex/inference/pipelines/table_recognition/pipeline_v2.py Fixes: OCR#17051 --- .../table_recognition/pipeline_v2.py | 24 +++++++++++++++++-- 1 file changed, 22 insertions(+), 2 deletions(-) diff --git a/paddlex/inference/pipelines/table_recognition/pipeline_v2.py b/paddlex/inference/pipelines/table_recognition/pipeline_v2.py index c7615d9a34..c9a510be18 100644 --- a/paddlex/inference/pipelines/table_recognition/pipeline_v2.py +++ b/paddlex/inference/pipelines/table_recognition/pipeline_v2.py @@ -663,16 +663,24 @@ def split_box_by_cells(ocr_box, cell_indices, cells): else: ocr_det_results = overall_ocr_res["rec_boxes"] ocr_texts = overall_ocr_res["rec_texts"] + # Get rec_scores if it exists + ocr_scores = overall_ocr_res.get("rec_scores", []) + if ocr_scores is not None and hasattr(ocr_scores, "tolist"): + ocr_scores = ocr_scores.tolist() + elif ocr_scores is None: + ocr_scores = [] # Make copies to modify new_boxes = [] new_texts = [] + new_scores = [] # Process each OCR box i = 0 while i < len(ocr_det_results): ocr_box = ocr_det_results[i] text = ocr_texts[i] + score = ocr_scores[i] if i < len(ocr_scores) else None # Find cells that significantly overlap with this OCR box overlapping_cells = get_overlapping_cells(ocr_box, cells_det_results) # Check if we need to split (spans >= k cells) @@ -683,6 +691,7 @@ def split_box_by_cells(ocr_box, cell_indices, cells): ) # Process each split box split_texts = [] + split_scores = [] for box in split_boxes: x1, y1, x2, y2 = int(box[0]), int(box[1]), int(box[2]), int(box[3]) if y2 - y1 > 1 and x2 - x1 > 1: @@ -698,21 +707,32 @@ def split_box_by_cells(ocr_box, cell_indices, cells): ] # Assumes "rec_texts" contains a single string else: result = "" + # Extract the score from the OCR result + if "rec_score" in ocr_result: + result_score = ocr_result["rec_score"] + else: + result_score = score if score is not None else 0.0 else: result = "" + result_score = score if score is not None else 0.0 split_texts.append(result) - # Add split boxes and texts to results + split_scores.append(result_score) + # Add split boxes, texts, and scores to results new_boxes.extend(split_boxes) new_texts.extend(split_texts) + new_scores.extend(split_scores) else: - # Keep original box and text + # Keep original box, text, and score new_boxes.append(ocr_box) new_texts.append(text) + new_scores.append(score if score is not None else 0.0) i += 1 # Update the results dictionary overall_ocr_res["rec_boxes"] = new_boxes overall_ocr_res["rec_texts"] = new_texts + if "rec_scores" in overall_ocr_res or len(new_scores) > 0: + overall_ocr_res["rec_scores"] = new_scores return overall_ocr_res