Skip to content

📄 Awesome OCR multiple programing languages toolkits based on ONNX Runtime, OpenVINO, MNN, PaddlePaddle, TensorRT and PyTorch.

Synchronize PR #151 and set it to openvino and paddle versions.

SWHLb2349ae4e662
diff --git a/python/rapidocr_onnxruntime/main.py b/python/rapidocr_onnxruntime/main.py
index 82e9175..feb63e2 100644
--- a/python/rapidocr_onnxruntime/main.py
+++ b/python/rapidocr_onnxruntime/main.py
@@ -2,7 +2,6 @@
# @Author: SWHL
# @Contact: [email protected]
import copy
-import logging
from pathlib import Path
from typing import List, Optional, Tuple, Union
@@ -114,8 +113,8 @@ class RapidOCR:
use_limit_ratio = w / h > self.width_height_ratio
if h <= self.min_height or use_limit_ratio:
- new_h = w
- padding_h = int((new_h - h) / 2)
+ new_h = max(int(w / self.width_height_ratio), self.min_height) * 2
+ padding_h = int(abs(new_h - h) / 2)
block_img = cv2.copyMakeBorder(
img, padding_h, padding_h, 0, 0, cv2.BORDER_CONSTANT, value=(0, 0, 0)
)
diff --git a/python/rapidocr_openvino/main.py b/python/rapidocr_openvino/main.py
index 269b952..feb63e2 100644
--- a/python/rapidocr_openvino/main.py
+++ b/python/rapidocr_openvino/main.py
@@ -2,7 +2,6 @@
# @Author: SWHL
# @Contact: [email protected]
import copy
-import logging
from pathlib import Path
from typing import List, Optional, Tuple, Union
@@ -83,6 +82,7 @@ class RapidOCR:
det_elapse, cls_elapse, rec_elapse = 0.0, 0.0, 0.0
if use_det:
+ img, padding_h = self.maybe_add_letterbox(img)
dt_boxes, det_elapse = self.auto_text_det(img)
if dt_boxes is None:
return None, None
@@ -95,28 +95,36 @@ class RapidOCR:
if use_rec:
rec_res, rec_elapse = self.text_rec(img)
+ if dt_boxes is not None and padding_h > 0:
+ for box in dt_boxes:
+ box[:, 1] -= padding_h
+
ocr_res = self.get_final_res(
dt_boxes, cls_res, rec_res, det_elapse, cls_elapse, rec_elapse
)
return ocr_res
- def auto_text_det(
- self,
- img: np.ndarray,
- ) -> Tuple[Optional[np.ndarray], float, Optional[List[np.ndarray]]]:
+ def maybe_add_letterbox(self, img: np.ndarray) -> Tuple[np.ndarray, int]:
h, w = img.shape[:2]
+
if self.width_height_ratio == -1:
use_limit_ratio = False
else:
use_limit_ratio = w / h > self.width_height_ratio
if h <= self.min_height or use_limit_ratio:
- logging.warning(
- "Because the aspect ratio of the current image exceeds the limit (min_height or width_height_ratio), the program will skip the detection step."
+ new_h = max(int(w / self.width_height_ratio), self.min_height) * 2
+ padding_h = int(abs(new_h - h) / 2)
+ block_img = cv2.copyMakeBorder(
+ img, padding_h, padding_h, 0, 0, cv2.BORDER_CONSTANT, value=(0, 0, 0)
)
- dt_boxes = self.get_boxes_img_without_det(h, w)
- return dt_boxes, 0.0
+ return block_img, padding_h
+ return img, 0
+ def auto_text_det(
+ self,
+ img: np.ndarray,
+ ) -> Tuple[Optional[np.ndarray], float, Optional[List[np.ndarray]]]:
dt_boxes, det_elapse = self.text_det(img)
if dt_boxes is None or len(dt_boxes) < 1:
return None, 0.0
@@ -124,12 +132,6 @@ class RapidOCR:
dt_boxes = self.sorted_boxes(dt_boxes)
return dt_boxes, det_elapse
- def get_boxes_img_without_det(self, h, w):
- x0, y0, x1, y1 = 0, 0, w, h
- dt_boxes = np.float32([[x0, y0], [x1, y0], [x1, y1], [x0, y1]])
- dt_boxes = dt_boxes[np.newaxis, ...]
- return dt_boxes
-
def get_crop_img_list(self, img, dt_boxes):
def get_rotate_crop_image(img, points):
img_crop_width = int(
@@ -241,7 +243,10 @@ def main():
use_cls = not args.no_cls
use_rec = not args.no_rec
result, elapse_list = ocr_engine(
- args.img_path, use_det=use_det, use_cls=use_cls, use_rec=use_rec
+ args.img_path,
+ use_det=use_det,
+ use_cls=use_cls,
+ use_rec=use_rec,
)
print(result)
diff --git a/python/rapidocr_paddle/main.py b/python/rapidocr_paddle/main.py
index 269b952..feb63e2 100644
--- a/python/rapidocr_paddle/main.py
+++ b/python/rapidocr_paddle/main.py
@@ -2,7 +2,6 @@
# @Author: SWHL
# @Contact: [email protected]
import copy
-import logging
from pathlib import Path
from typing import List, Optional, Tuple, Union
@@ -83,6 +82,7 @@ class RapidOCR:
det_elapse, cls_elapse, rec_elapse = 0.0, 0.0, 0.0
if use_det:
+ img, padding_h = self.maybe_add_letterbox(img)
dt_boxes, det_elapse = self.auto_text_det(img)
if dt_boxes is None:
return None, None
@@ -95,28 +95,36 @@ class RapidOCR:
if use_rec:
rec_res, rec_elapse = self.text_rec(img)
+ if dt_boxes is not None and padding_h > 0:
+ for box in dt_boxes:
+ box[:, 1] -= padding_h
+
ocr_res = self.get_final_res(
dt_boxes, cls_res, rec_res, det_elapse, cls_elapse, rec_elapse
)
return ocr_res
- def auto_text_det(
- self,
- img: np.ndarray,
- ) -> Tuple[Optional[np.ndarray], float, Optional[List[np.ndarray]]]:
+ def maybe_add_letterbox(self, img: np.ndarray) -> Tuple[np.ndarray, int]:
h, w = img.shape[:2]
+
if self.width_height_ratio == -1:
use_limit_ratio = False
else:
use_limit_ratio = w / h > self.width_height_ratio
if h <= self.min_height or use_limit_ratio:
- logging.warning(
- "Because the aspect ratio of the current image exceeds the limit (min_height or width_height_ratio), the program will skip the detection step."
+ new_h = max(int(w / self.width_height_ratio), self.min_height) * 2
+ padding_h = int(abs(new_h - h) / 2)
+ block_img = cv2.copyMakeBorder(
+ img, padding_h, padding_h, 0, 0, cv2.BORDER_CONSTANT, value=(0, 0, 0)
)
- dt_boxes = self.get_boxes_img_without_det(h, w)
- return dt_boxes, 0.0
+ return block_img, padding_h
+ return img, 0
+ def auto_text_det(
+ self,
+ img: np.ndarray,
+ ) -> Tuple[Optional[np.ndarray], float, Optional[List[np.ndarray]]]:
dt_boxes, det_elapse = self.text_det(img)
if dt_boxes is None or len(dt_boxes) < 1:
return None, 0.0
@@ -124,12 +132,6 @@ class RapidOCR:
dt_boxes = self.sorted_boxes(dt_boxes)
return dt_boxes, det_elapse
- def get_boxes_img_without_det(self, h, w):
- x0, y0, x1, y1 = 0, 0, w, h
- dt_boxes = np.float32([[x0, y0], [x1, y0], [x1, y1], [x0, y1]])
- dt_boxes = dt_boxes[np.newaxis, ...]
- return dt_boxes
-
def get_crop_img_list(self, img, dt_boxes):
def get_rotate_crop_image(img, points):
img_crop_width = int(
@@ -241,7 +243,10 @@ def main():
use_cls = not args.no_cls
use_rec = not args.no_rec
result, elapse_list = ocr_engine(
- args.img_path, use_det=use_det, use_cls=use_cls, use_rec=use_rec
+ args.img_path,
+ use_det=use_det,
+ use_cls=use_cls,
+ use_rec=use_rec,
)
print(result)
diff --git a/python/tests/test_ort.py b/python/tests/test_ort.py
index 247bd9e..3e70514 100644
--- a/python/tests/test_ort.py
+++ b/python/tests/test_ort.py
@@ -26,7 +26,7 @@ package_name = "rapidocr_onnxruntime"
(
"test_letterbox_like.jpg",
2,
- "A::取决于所使用的执行提供者,它可能没有完全支持模型中的所有操作。回落到CPU操作可能会导致性能速度的下降。此外,即使一个操作是由CUDAexecution",
+ "A::取决于所使用的执行提供者,它可能没有完全支持模型中的所有操作。回落到CPU操作可能会导致性能速度的下降。此外,即使一个操作是由CUDAeXecution",
),
("test_without_det.jpg", 1, "在中国作家协会第三届儿童文学"),
],
diff --git a/python/tests/test_paddle.py b/python/tests/test_paddle.py
index af344ce..c370001 100644
--- a/python/tests/test_paddle.py
+++ b/python/tests/test_paddle.py
@@ -20,11 +20,23 @@ img_path = tests_dir / "ch_en_num.jpg"
package_name = "rapidocr_paddle"
-def test_without_det():
- img_path = tests_dir / "test_without_det.png"
+ "img_name,gt_len,gt_first_len",
+ [
+ (
+ "test_letterbox_like.jpg",
+ 2,
+ "A::取决于所使用的执行提供者,它可能没有完全支持模型中的所有操作。回落到CPU操作可能会导致性能速度的下降。此外,即使一个操作是由CUDAeXecution",
+ ),
+ ("test_without_det.jpg", 1, "在中国作家协会第三届儿童文学"),
+ ],
+)
+def test_letterbox_like(img_name, gt_len, gt_first_len):
+ img_path = tests_dir / img_name
result, _ = engine(img_path)
- assert result[0][1] == "の持、持场所无。"
- assert len(result) == 1
+
+ assert len(result) == gt_len
+ assert result[0][1] == gt_first_len
def test_only_det():
@@ -114,7 +126,7 @@ def test_input_parameters():
def test_input_det_parameters():
with pytest.raises(FileNotFoundError) as exc_info:
- engine = RapidOCR(det_model_path="1")
+ engine = RapidOCR(det_model_path="1.onnx")
result, _ = engine(img_path)
raise FileNotFoundError()
assert exc_info.type is FileNotFoundError
@@ -122,7 +134,7 @@ def test_input_det_parameters():
def test_input_cls_parameters():
with pytest.raises(FileNotFoundError) as exc_info:
- engine = RapidOCR(cls_model_path="1")
+ engine = RapidOCR(cls_model_path="1.onnx")
result, _ = engine(img_path)
raise FileNotFoundError()
assert exc_info.type is FileNotFoundError
@@ -130,7 +142,7 @@ def test_input_cls_parameters():
def test_input_rec_parameters():
with pytest.raises(FileNotFoundError) as exc_info:
- engine = RapidOCR(rec_model_path="1")
+ engine = RapidOCR(rec_model_path="1.onnx")
result, _ = engine(img_path)
raise FileNotFoundError()
assert exc_info.type is FileNotFoundError
@@ -142,7 +154,7 @@ def test_input_three_ndim_two_channel():
result, _ = engine(image_array)
assert len(result) == 1
- assert result[0][1] == " TREND PLOT REPORT"
+ assert result[0][1] == "TREND PLOT REPORT"
def test_input_three_ndim_one_channel():
diff --git a/python/tests/test_vino.py b/python/tests/test_vino.py
index 5f5f619..db8e2a0 100644
--- a/python/tests/test_vino.py
+++ b/python/tests/test_vino.py
@@ -20,11 +20,23 @@ img_path = tests_dir / "ch_en_num.jpg"
package_name = "rapidocr_openvino"
-def test_without_det():
- img_path = tests_dir / "test_without_det.png"
+ "img_name,gt_len,gt_first_len",
+ [
+ (
+ "test_letterbox_like.jpg",
+ 2,
+ "A::取决于所使用的执行提供者,它可能没有完全支持模型中的所有操作。回落到CPU操作可能会导致性能速度的下降。此外,即使一个操作是由CUDAeXecution",
+ ),
+ ("test_without_det.jpg", 1, "在中国作家协会第三届儿童文学"),
+ ],
+)
+def test_letterbox_like(img_name, gt_len, gt_first_len):
+ img_path = tests_dir / img_name
result, _ = engine(img_path)
- assert result[0][1] == "の持、持场所无。"
- assert len(result) == 1
+
+ assert len(result) == gt_len
+ assert result[0][1] == gt_first_len
def test_only_det():
@@ -142,7 +154,7 @@ def test_input_three_ndim_two_channel():
result, _ = engine(image_array)
assert len(result) == 1
- assert result[0][1] == " TREND PLOT REPORT"
+ assert result[0][1] == "TREND PLOT REPORT"
def test_input_three_ndim_one_channel():