Transforms complex documents like PDFs and Office docs into LLM-ready markdown/JSON for your Agentic workflows.
Python ai4sciencedocument-analysisdocxextract-datalayout-analysisocrparserpdfpdf-converterpdf-extractor-llmpdf-extractor-pretrainpdf-extractor-ragpdf-parserpptxpythonxlsx
fix(ocr): skip degenerate text crops before recognition
myhloli0dbefb400eaf
diff --git a/mineru/model/ocr/image.py b/mineru/model/ocr/image.pyindex 7b5cf03d..0f3e64e5 100644--- a/mineru/model/ocr/image.py+++ b/mineru/model/ocr/image.py@@ -145,6 +145,11 @@ def resize_text_recognition_image(def get_rotate_crop_image_for_text_rec(img: np.ndarray, points: np.ndarray) -> np.ndarray | None:"""裁剪四边形文本区域并应用文字识别方向规则。"""+ # 对边长度截断不足一个像素的退化框没有可识别裁图,避免送入共享 warp 的零尺寸目标。+ crop_width = max(np.linalg.norm(points[0] - points[1]), np.linalg.norm(points[2] - points[3]))+ crop_height = max(np.linalg.norm(points[0] - points[3]), np.linalg.norm(points[1] - points[2]))+ if crop_width < 1 or crop_height < 1:+ return Nonereturn rotate_vertical_crop_if_needed(get_rotate_crop_image(img, points))diff --git a/mineru/model/ocr/pytorch_paddle.py b/mineru/model/ocr/pytorch_paddle.pyindex 5d214c4f..836be732 100644--- a/mineru/model/ocr/pytorch_paddle.py+++ b/mineru/model/ocr/pytorch_paddle.py@@ -228,7 +228,7 @@ class PytorchPaddleOCR(TextSystem):def __call__(self, img: np.ndarray | None, mfd_res: list[dict[str, Any]] | None = None) -> tuple[list[np.ndarray] | None, list[tuple[str, float]] | None]:- """执行检测与识别,跳过无效印章裁片时同步移除检测框,避免文字错配。"""+ """执行检测与识别,跳过无效裁片时同步移除检测框,避免文字错配。"""if img is None:logger.debug("no valid image provided")return None, None@@ -262,10 +262,17 @@ class PytorchPaddleOCR(TextSystem):# Standard text OCR rotates tall crops before recognition.# Seal OCR keeps its dedicated poly-crop path above.+ crop_boxes = []for bno in range(len(dt_boxes)):tmp_box = copy.deepcopy(dt_boxes[bno])img_crop = get_rotate_crop_image_for_text_rec(ori_im, tmp_box)+ if img_crop is None:+ continueimg_crop_list.append(img_crop)+ crop_boxes.append(dt_boxes[bno])+ dt_boxes = crop_boxes+ if not img_crop_list:+ return [], []rec_res, elapse = self.text_recognizer(img_crop_list)# logger.debug("rec_res num : {}, elapsed : {}".format(len(rec_res), elapse))diff --git a/mineru/model/ocr/results.py b/mineru/model/ocr/results.pyindex 512ecffe..fc4045d7 100644--- a/mineru/model/ocr/results.py+++ b/mineru/model/ocr/results.py@@ -60,11 +60,16 @@ def get_ocr_result_list(else:p1, p2, p3, p4 = box_ocr_restext, score = "", 1+ # 公式遮罩拆框可能产生零宽或窄残片;沿用既有过滤条件,先过滤再裁图。+ if (p3[0] - p1[0]) < OcrConfidence.min_width:+ continueif ocr_enable:if ori_im is None:ori_im = bgr_image.copy()tmp_box = copy.deepcopy(np.array([p1, p2, p3, p4]).astype("float32"))img_crop = get_rotate_crop_image_for_text_rec(ori_im, tmp_box)+ if img_crop is None:+ continueneed_ocr_rec = Truepoly = [p1, p2, p3, p4]diff --git a/tests/unittest/test_ocr_text_crop_degenerate.py b/tests/unittest/test_ocr_text_crop_degenerate.pynew file mode 100644index 00000000..c5a0e5cc--- /dev/null+++ b/tests/unittest/test_ocr_text_crop_degenerate.py@@ -0,0 +1,63 @@+"""正文 OCR 在公式遮罩拆框之后的零尺寸裁图回归。"""++from __future__ import annotations++from unittest.mock import Mock++import numpy as np+import pytest++from mineru.model.ocr.geometry import update_det_boxes+from mineru.model.ocr.image import get_rotate_crop_image_for_text_rec+from mineru.model.ocr.results import get_ocr_result_list+++ "points",+ (+ [[4, 4], [4, 4], [4, 12], [4, 12]],+ [[4, 4], [16, 4], [16, 4], [4, 4]],+ [[4, 4], [16, 6], [16, 6.5], [4, 4.5]],+ ),+)+def test_zero_pixel_text_crop_is_discarded(points: list[list[float]]) -> None:+ """零宽、零高及倾斜亚像素高度不能进入透视采样或后续 OCR rec。"""+ image = np.zeros((24, 24, 3), dtype=np.uint8)+ assert get_rotate_crop_image_for_text_rec(image, np.asarray(points, dtype=np.float32)) is None+ assert get_ocr_result_list([points], [0, 0, 0, 0, 24, 24, 24, 24], True, image) == []+++def test_formula_split_zero_width_fragment_preserves_valid_text() -> None:+ """公式遮罩留下闭区间端点时丢弃零宽残片,合法正文的坐标和裁图保持原样。"""+ image = np.arange(24 * 24 * 3, dtype=np.uint8).reshape(24, 24, 3)+ boxes = update_det_boxes(+ [[[4, 4], [20, 4], [20, 12], [4, 12]]],+ [{"bbox": [5, 4, 10, 12]}],+ )+ assert boxes[0][0][0] == boxes[0][2][0] == 4+ output = get_ocr_result_list(boxes, [0, 0, 0, 0, 24, 24, 24, 24], True, image)+ assert len(output) == 1 and output[0]["bbox"] == (11, 4, 20, 12)+ assert output[0]["text"] == "" and output[0]["score"] == 1.0+ np.testing.assert_array_equal(output[0]["np_img"], image[4:12, 11:20])++[email protected]("mask_right", (10, 20))+def test_torch_ocr_discards_invalid_crops_without_text_misalignment(mask_right: int) -> None:+ """Torch 整体 OCR 同步过滤零宽框和裁图,全部被过滤时不调用识别器。"""+ from mineru.model.ocr.pytorch_paddle import PytorchPaddleOCR++ model = object.__new__(PytorchPaddleOCR)+ model.is_seal = False+ model.enable_merge_det_boxes = False+ model.drop_score = 0.5+ model.text_detector = Mock(return_value=(np.asarray([[[4, 4], [20, 4], [20, 12], [4, 12]]]), 0.0))+ model.text_recognizer = Mock(return_value=([("valid", 0.99)], 0.0))+ image = np.zeros((24, 24, 3), dtype=np.uint8)+ boxes, texts = model(image, mfd_res=[{"bbox": [5, 4, mask_right, 12]}])+ if mask_right == 20:+ assert boxes == texts == []+ model.text_recognizer.assert_not_called()+ else:+ assert len(boxes) == 1 and texts == [("valid", 0.99)]+ np.testing.assert_array_equal(boxes[0], [[11, 4], [20, 4], [20, 12], [11, 12]])+ assert len(model.text_recognizer.call_args.args[0]) == 1