oas_ocr.py 3.3 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697
  1. import functools
  2. import cv2
  3. import numpy as np
  4. from module.base.utils import color_similarity_2d, load_image
  5. # from module.ocr.ocr import Ocr
  6. from module.ocr.base_ocr import BaseCor
  7. def apply_mask(image, mask):
  8. image16 = image.astype(np.uint16)
  9. mask16 = mask.astype(np.uint16)
  10. mask16 = cv2.merge([mask16, mask16, mask16])
  11. image16 = cv2.multiply(image16, mask16)
  12. # cv2.multiply(image16, mask16, dst=image16)
  13. image16 = cv2.convertScaleAbs(image16, alpha=1 / 255)
  14. # cv2.convertScaleAbs(image16, alpha=1 / 255, dst=image16)
  15. # Image.fromarray(image16.astype(np.uint8)).show()
  16. return image16.astype(np.uint8)
  17. class VerticalText(BaseCor):
  18. @staticmethod
  19. def rotate_image(image):
  20. height, width = image.shape[0:2]
  21. if height * 1.0 / width >= 1.5:
  22. image = np.rot90(image)
  23. return image
  24. return image
  25. def detect_and_ocr(self, *args, **kwargs):
  26. if getattr(self.model, "is_proxy", False):
  27. params = {"drop_score": 0.1, "box_thresh": 0.2, "vertical": True}
  28. for key in list(params.keys()):
  29. if key in kwargs:
  30. params.pop(key)
  31. return self.model.detect_and_ocr(*args, **params, **kwargs)
  32. # Try hard to lower TextSystem.box_thresh
  33. backup = self.model.text_detector.box_thresh
  34. # Patch text_recognizer
  35. text_recognizer = self.model.text_recognizer
  36. # Lower drop_score
  37. lower_score = functools.partial(self.model.detect_and_ocr, drop_score=0.1)
  38. detect_and_ocr = self.model.detect_and_ocr
  39. def vertical_text_recognizer(img_crop_list):
  40. img_crop_list = [VerticalText.rotate_image(i) for i in img_crop_list]
  41. result = text_recognizer(img_crop_list)
  42. return result
  43. self.model.text_detector.box_thresh = 0.2
  44. self.model.text_recognizer = vertical_text_recognizer
  45. self.model.detect_and_ocr = lower_score
  46. try:
  47. result = super().detect_and_ocr(*args, **kwargs)
  48. finally:
  49. self.model.text_detector.box_thresh = backup
  50. self.model.text_recognizer = text_recognizer
  51. self.model.detect_and_ocr = detect_and_ocr
  52. return result
  53. class StoneOcr(VerticalText):
  54. def pre_process(self, image):
  55. yuv = cv2.cvtColor(image, cv2.COLOR_RGB2YUV)
  56. _, u, _ = cv2.split(yuv)
  57. cv2.subtract(128, u, dst=u)
  58. cv2.multiply(u, 8, dst=u)
  59. color = color_similarity_2d(image, color=(234, 213, 181))
  60. hsv = cv2.cvtColor(image, cv2.COLOR_RGB2HSV)
  61. _, _, v = cv2.split(hsv)
  62. image = apply_mask(image, u)
  63. image = apply_mask(image, color)
  64. image = apply_mask(image, v)
  65. cv2.convertScaleAbs(image, alpha=3, dst=image)
  66. cv2.subtract((255, 255, 255, 0), image, dst=image)
  67. # from PIL import Image
  68. # Image.fromarray(image.astype(np.uint8)).show()
  69. return image
  70. if __name__ == '__main__':
  71. from tasks.SixRealms.assets import SixRealmsAssets
  72. file = r'C:\Users\Ryland\Desktop\Desktop\20.png'
  73. image = load_image(file)
  74. ocr = StoneOcr(roi=(0,0,1280,720), area=(0,0,1280,720), mode="Full", method="Default", keyword="", name="ocr_map")
  75. results = ocr.detect_and_ocr(image)
  76. for r in results:
  77. print(r.box, r.ocr_text)