doc-page-extractor 0.2.4__cp310-cp310-macosx_15_0_arm64.whl

doc_page_extractor/onnxocr/predict_det.py

@@ -0,0 +1,139 @@
+import numpy as np
+from .imaug import transform, create_operators
+from .db_postprocess import DBPostProcess
+from .predict_base import PredictBase
+
+
+class TextDetector(PredictBase):
+  def __init__(self, args):
+    super().__init__()
+    self._args = args
+    self.det_algorithm = args.det_algorithm
+    pre_process_list = [
+      {
+        "DetResizeForTest": {
+          "limit_side_len": args.det_limit_side_len,
+          "limit_type": args.det_limit_type,
+        }
+      },
+      {
+        "NormalizeImage": {
+          "std": [0.229, 0.224, 0.225],
+          "mean": [0.485, 0.456, 0.406],
+          "scale": "1./255.",
+          "order": "hwc",
+        }
+      },
+      {"ToCHWImage": None},
+      {"KeepKeys": {"keep_keys": ["image", "shape"]}},
+    ]
+    postprocess_params = {}
+    postprocess_params["name"] = "DBPostProcess"
+    postprocess_params["thresh"] = args.det_db_thresh
+    postprocess_params["box_thresh"] = args.det_db_box_thresh
+    postprocess_params["max_candidates"] = 1000
+    postprocess_params["unclip_ratio"] = args.det_db_unclip_ratio
+    postprocess_params["use_dilation"] = args.use_dilation
+    postprocess_params["score_mode"] = args.det_db_score_mode
+    postprocess_params["box_type"] = args.det_box_type
+
+    # 实例化预处理操作类
+    self.preprocess_op = create_operators(pre_process_list)
+    # self.postprocess_op = build_post_process(postprocess_params)
+    # 实例化后处理操作类
+    self.postprocess_op = DBPostProcess(**postprocess_params)
+
+    # 初始化模型
+    self._det_onnx_session = None
+    self._det_input_name = None
+    self._det_output_name = None
+
+  @property
+  def det_onnx_session(self):
+    if self._det_onnx_session is None:
+      self._det_onnx_session = self.get_onnx_session(self._args.det_model_dir, self._args.use_gpu)
+    return self._det_onnx_session
+
+  @property
+  def det_input_name(self):
+    if self._det_input_name is None:
+      self._det_input_name = self.get_input_name(self.det_onnx_session)
+    return self._det_input_name
+
+  @property
+  def det_output_name(self):
+    if self._det_output_name is None:
+      self._det_output_name = self.get_output_name(self.det_onnx_session)
+    return self._det_output_name
+
+  def order_points_clockwise(self, pts):
+    rect = np.zeros((4, 2), dtype="float32")
+    s = pts.sum(axis=1)
+    rect[0] = pts[np.argmin(s)]
+    rect[2] = pts[np.argmax(s)]
+    tmp = np.delete(pts, (np.argmin(s), np.argmax(s)), axis=0)
+    diff = np.diff(np.array(tmp), axis=1)
+    rect[1] = tmp[np.argmin(diff)]
+    rect[3] = tmp[np.argmax(diff)]
+    return rect
+
+  def clip_det_res(self, points, img_height, img_width):
+    for pno in range(points.shape[0]):
+      points[pno, 0] = int(min(max(points[pno, 0], 0), img_width - 1))
+      points[pno, 1] = int(min(max(points[pno, 1], 0), img_height - 1))
+    return points
+
+  def filter_tag_det_res(self, dt_boxes, image_shape):
+    img_height, img_width = image_shape[0:2]
+    dt_boxes_new = []
+    for box in dt_boxes:
+      if type(box) is list:
+        box = np.array(box)
+      box = self.order_points_clockwise(box)
+      box = self.clip_det_res(box, img_height, img_width)
+      rect_width = int(np.linalg.norm(box[0] - box[1]))
+      rect_height = int(np.linalg.norm(box[0] - box[3]))
+      if rect_width <= 3 or rect_height <= 3:
+        continue
+      dt_boxes_new.append(box)
+    dt_boxes = np.array(dt_boxes_new)
+    return dt_boxes
+
+  def filter_tag_det_res_only_clip(self, dt_boxes, image_shape):
+    img_height, img_width = image_shape[0:2]
+    dt_boxes_new = []
+    for box in dt_boxes:
+      if type(box) is list:
+        box = np.array(box)
+      box = self.clip_det_res(box, img_height, img_width)
+      dt_boxes_new.append(box)
+    dt_boxes = np.array(dt_boxes_new)
+    return dt_boxes
+
+  def __call__(self, img):
+    ori_im = img.copy()
+    data = {"image": img}
+
+    data = transform(data, self.preprocess_op)
+    img, shape_list = data
+    if img is None:
+      return None, 0
+    img = np.expand_dims(img, axis=0)
+    shape_list = np.expand_dims(shape_list, axis=0)
+    img = img.copy()
+
+    input_feed = self.get_input_feed(self.det_input_name, img)
+    outputs = self.det_onnx_session.run(self.det_output_name, input_feed=input_feed)
+
+    preds = {}
+    preds["maps"] = outputs[0]
+
+    post_result = self.postprocess_op(preds, shape_list)
+    dt_boxes = post_result[0]["points"]
+
+    if self._args.det_box_type == "poly":
+      dt_boxes = self.filter_tag_det_res_only_clip(dt_boxes, ori_im.shape)
+    else:
+      dt_boxes = self.filter_tag_det_res(dt_boxes, ori_im.shape)
+
+    return dt_boxes

doc_page_extractor/onnxocr/predict_rec.py

@@ -0,0 +1,344 @@
+import cv2
+import numpy as np
+import math
+from PIL import Image
+
+
+from .rec_postprocess import CTCLabelDecode
+from .predict_base import PredictBase
+
+
+class TextRecognizer(PredictBase):
+  def __init__(self, args):
+    super().__init__()
+    self._args = args
+    self.rec_image_shape = args.rec_image_shape
+    self.rec_batch_num = args.rec_batch_num
+    self.rec_algorithm = args.rec_algorithm
+    self.postprocess_op = CTCLabelDecode(
+      character_dict_path=args.rec_char_dict_path,
+      use_space_char=args.use_space_char,
+    )
+
+    # 初始化模型
+    self._rec_onnx_session = None
+    self._rec_input_name = None
+    self._rec_output_name = None
+
+  @property
+  def rec_onnx_session(self):
+    if self._rec_onnx_session is None:
+      self._rec_onnx_session = self.get_onnx_session(
+        self._args.rec_model_dir, self._args.use_gpu
+      )
+    return self._rec_onnx_session
+
+  @property
+  def rec_input_name(self):
+    if self._rec_input_name is None:
+      self._rec_input_name = self.get_input_name(self.rec_onnx_session)
+    return self._rec_input_name
+
+  @property
+  def rec_output_name(self):
+    if self._rec_output_name is None:
+      self._rec_output_name = self.get_output_name(self.rec_onnx_session)
+    return self._rec_output_name
+
+  def resize_norm_img(self, img, max_wh_ratio):
+    imgC, imgH, imgW = self.rec_image_shape
+    if self.rec_algorithm == "NRTR" or self.rec_algorithm == "ViTSTR":
+      img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+      # return padding_im
+      image_pil = Image.fromarray(np.uint8(img))
+      if self.rec_algorithm == "ViTSTR":
+        img = image_pil.resize([imgW, imgH], Image.Resampling.BICUBIC)
+      else:
+        img = image_pil.resize([imgW, imgH], Image.Resampling.LANCZOS)
+      img = np.array(img)
+      norm_img = np.expand_dims(img, -1)
+      norm_img = norm_img.transpose((2, 0, 1))
+      if self.rec_algorithm == "ViTSTR":
+        norm_img = norm_img.astype(np.float32) / 255.0
+      else:
+        norm_img = norm_img.astype(np.float32) / 128.0 - 1.0
+      return norm_img
+    elif self.rec_algorithm == "RFL":
+      img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+      resized_image = cv2.resize(img, (imgW, imgH), interpolation=cv2.INTER_CUBIC)
+      resized_image = resized_image.astype("float32")
+      resized_image = resized_image / 255
+      resized_image = resized_image[np.newaxis, :]
+      resized_image -= 0.5
+      resized_image /= 0.5
+      return resized_image
+
+    assert imgC == img.shape[2]
+    imgW = int((imgH * max_wh_ratio))
+
+    # w = self.rec_onnx_session.get_inputs()[0].shape[3:][0]
+    # w = self.rec_onnx_session.get_inputs()[0].shape[3:][0]
+    # print(w)
+    # if w is not None and w > 0:
+    #   imgW = w
+
+    h, w = img.shape[:2]
+    ratio = w / float(h)
+    if math.ceil(imgH * ratio) > imgW:
+      resized_w = imgW
+    else:
+      resized_w = int(math.ceil(imgH * ratio))
+    if self.rec_algorithm == "RARE":
+      if resized_w > self.rec_image_shape[2]:
+        resized_w = self.rec_image_shape[2]
+      imgW = self.rec_image_shape[2]
+    resized_image = cv2.resize(img, (resized_w, imgH))
+    resized_image = resized_image.astype("float32")
+    resized_image = resized_image.transpose((2, 0, 1)) / 255
+    resized_image -= 0.5
+    resized_image /= 0.5
+    padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
+    padding_im[:, :, 0:resized_w] = resized_image
+    return padding_im
+
+  def resize_norm_img_vl(self, img, image_shape):
+    imgC, imgH, imgW = image_shape
+    img = img[:, :, ::-1]  # bgr2rgb
+    resized_image = cv2.resize(img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
+    resized_image = resized_image.astype("float32")
+    resized_image = resized_image.transpose((2, 0, 1)) / 255
+    return resized_image
+
+  def resize_norm_img_srn(self, img, image_shape):
+    imgC, imgH, imgW = image_shape
+
+    img_black = np.zeros((imgH, imgW))
+    im_hei = img.shape[0]
+    im_wid = img.shape[1]
+
+    if im_wid <= im_hei * 1:
+      img_new = cv2.resize(img, (imgH * 1, imgH))
+    elif im_wid <= im_hei * 2:
+      img_new = cv2.resize(img, (imgH * 2, imgH))
+    elif im_wid <= im_hei * 3:
+      img_new = cv2.resize(img, (imgH * 3, imgH))
+    else:
+      img_new = cv2.resize(img, (imgW, imgH))
+
+    img_np = np.asarray(img_new)
+    img_np = cv2.cvtColor(img_np, cv2.COLOR_BGR2GRAY)
+    img_black[:, 0 : img_np.shape[1]] = img_np
+    img_black = img_black[:, :, np.newaxis]
+
+    row, col, c = img_black.shape
+    c = 1
+
+    return np.reshape(img_black, (c, row, col)).astype(np.float32)
+
+  def srn_other_inputs(self, image_shape, num_heads, max_text_length):
+    imgC, imgH, imgW = image_shape
+    feature_dim = int((imgH / 8) * (imgW / 8))
+
+    encoder_word_pos = (
+      np.array(range(0, feature_dim)).reshape((feature_dim, 1)).astype("int64")
+    )
+    gsrm_word_pos = (
+      np.array(range(0, max_text_length))
+      .reshape((max_text_length, 1))
+      .astype("int64")
+    )
+
+    gsrm_attn_bias_data = np.ones((1, max_text_length, max_text_length))
+    gsrm_slf_attn_bias1 = np.triu(gsrm_attn_bias_data, 1).reshape(
+      [-1, 1, max_text_length, max_text_length]
+    )
+    gsrm_slf_attn_bias1 = np.tile(gsrm_slf_attn_bias1, [1, num_heads, 1, 1]).astype(
+      "float32"
+    ) * [-1e9]
+
+    gsrm_slf_attn_bias2 = np.tril(gsrm_attn_bias_data, -1).reshape(
+      [-1, 1, max_text_length, max_text_length]
+    )
+    gsrm_slf_attn_bias2 = np.tile(gsrm_slf_attn_bias2, [1, num_heads, 1, 1]).astype(
+      "float32"
+    ) * [-1e9]
+
+    encoder_word_pos = encoder_word_pos[np.newaxis, :]
+    gsrm_word_pos = gsrm_word_pos[np.newaxis, :]
+
+    return [
+      encoder_word_pos,
+      gsrm_word_pos,
+      gsrm_slf_attn_bias1,
+      gsrm_slf_attn_bias2,
+    ]
+
+  def process_image_srn(self, img, image_shape, num_heads, max_text_length):
+    norm_img = self.resize_norm_img_srn(img, image_shape)
+    norm_img = norm_img[np.newaxis, :]
+
+    [encoder_word_pos, gsrm_word_pos, gsrm_slf_attn_bias1, gsrm_slf_attn_bias2] = (
+      self.srn_other_inputs(image_shape, num_heads, max_text_length)
+    )
+
+    gsrm_slf_attn_bias1 = gsrm_slf_attn_bias1.astype(np.float32)
+    gsrm_slf_attn_bias2 = gsrm_slf_attn_bias2.astype(np.float32)
+    encoder_word_pos = encoder_word_pos.astype(np.int64)
+    gsrm_word_pos = gsrm_word_pos.astype(np.int64)
+
+    return (
+      norm_img,
+      encoder_word_pos,
+      gsrm_word_pos,
+      gsrm_slf_attn_bias1,
+      gsrm_slf_attn_bias2,
+    )
+
+  def resize_norm_img_sar(self, img, image_shape, width_downsample_ratio=0.25):
+    imgC, imgH, imgW_min, imgW_max = image_shape
+    h = img.shape[0]
+    w = img.shape[1]
+    valid_ratio = 1.0
+    # make sure new_width is an integral multiple of width_divisor.
+    width_divisor = int(1 / width_downsample_ratio)
+    # resize
+    ratio = w / float(h)
+    resize_w = math.ceil(imgH * ratio)
+    if resize_w % width_divisor != 0:
+      resize_w = round(resize_w / width_divisor) * width_divisor
+    if imgW_min is not None:
+      resize_w = max(imgW_min, resize_w)
+    if imgW_max is not None:
+      valid_ratio = min(1.0, 1.0 * resize_w / imgW_max)
+      resize_w = min(imgW_max, resize_w)
+    resized_image = cv2.resize(img, (resize_w, imgH))
+    resized_image = resized_image.astype("float32")
+    # norm
+    if image_shape[0] == 1:
+      resized_image = resized_image / 255
+      resized_image = resized_image[np.newaxis, :]
+    else:
+      resized_image = resized_image.transpose((2, 0, 1)) / 255
+    resized_image -= 0.5
+    resized_image /= 0.5
+    resize_shape = resized_image.shape
+    padding_im = -1.0 * np.ones((imgC, imgH, imgW_max), dtype=np.float32)
+    padding_im[:, :, 0:resize_w] = resized_image
+    pad_shape = padding_im.shape
+
+    return padding_im, resize_shape, pad_shape, valid_ratio
+
+  def resize_norm_img_spin(self, img):
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+    # return padding_im
+    img = cv2.resize(img, tuple([100, 32]), cv2.INTER_CUBIC)
+    img = np.array(img, np.float32)
+    img = np.expand_dims(img, -1)
+    img = img.transpose((2, 0, 1))
+    mean = [127.5]
+    std = [127.5]
+    mean = np.array(mean, dtype=np.float32)
+    std = np.array(std, dtype=np.float32)
+    mean = np.float32(mean.reshape(1, -1))
+    stdinv = 1 / np.float32(std.reshape(1, -1))
+    img -= mean
+    img *= stdinv
+    return img
+
+  def resize_norm_img_svtr(self, img, image_shape):
+    imgC, imgH, imgW = image_shape
+    resized_image = cv2.resize(img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
+    resized_image = resized_image.astype("float32")
+    resized_image = resized_image.transpose((2, 0, 1)) / 255
+    resized_image -= 0.5
+    resized_image /= 0.5
+    return resized_image
+
+  def resize_norm_img_abinet(self, img, image_shape):
+    imgC, imgH, imgW = image_shape
+
+    resized_image = cv2.resize(img, (imgW, imgH), interpolation=cv2.INTER_LINEAR)
+    resized_image = resized_image.astype("float32")
+    resized_image = resized_image / 255.0
+
+    mean = np.array([0.485, 0.456, 0.406])
+    std = np.array([0.229, 0.224, 0.225])
+    resized_image = (resized_image - mean[None, None, ...]) / std[None, None, ...]
+    resized_image = resized_image.transpose((2, 0, 1))
+    resized_image = resized_image.astype("float32")
+
+    return resized_image
+
+  def norm_img_can(self, img, image_shape):
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)  # CAN only predict gray scale image
+
+    # FIXME
+    # if self.inverse:
+    #   img = 255 - img
+
+    if self.rec_image_shape[0] == 1:
+      h, w = img.shape
+      _, imgH, imgW = self.rec_image_shape
+      if h < imgH or w < imgW:
+        padding_h = max(imgH - h, 0)
+        padding_w = max(imgW - w, 0)
+        img_padded = np.pad(
+          img,
+          ((0, padding_h), (0, padding_w)),
+          "constant",
+          constant_values=(255),
+        )
+        img = img_padded
+
+    img = np.expand_dims(img, 0) / 255.0  # h,w,c -> c,h,w
+    img = img.astype("float32")
+
+    return img
+
+  def __call__(self, img_list):
+    img_num = len(img_list)
+    # Calculate the aspect ratio of all text bars
+    width_list = []
+    for img in img_list:
+      width_list.append(img.shape[1] / float(img.shape[0]))
+    # Sorting can speed up the recognition process
+    indices = np.argsort(np.array(width_list))
+    rec_res = [["", 0.0]] * img_num
+    batch_num = self.rec_batch_num
+
+    for beg_img_no in range(0, img_num, batch_num):
+      end_img_no = min(img_num, beg_img_no + batch_num)
+      norm_img_batch = []
+      imgC, imgH, imgW = self.rec_image_shape[:3]
+      max_wh_ratio = imgW / imgH
+      # max_wh_ratio = 0
+      for ino in range(beg_img_no, end_img_no):
+        h, w = img_list[indices[ino]].shape[0:2]
+        wh_ratio = w * 1.0 / h
+        max_wh_ratio = max(max_wh_ratio, wh_ratio)
+      for ino in range(beg_img_no, end_img_no):
+        norm_img = self.resize_norm_img(img_list[indices[ino]], max_wh_ratio)
+        norm_img = norm_img[np.newaxis, :]
+        norm_img_batch.append(norm_img)
+
+      norm_img_batch = np.concatenate(norm_img_batch)
+      norm_img_batch = norm_img_batch.copy()
+
+      # img = img[:, :, ::-1].transpose(2, 0, 1)
+      # img = img[:, :, ::-1]
+      # img = img.transpose(2, 0, 1)
+      # img = img.astype(np.float32)
+      # img = np.expand_dims(img, axis=0)
+      # print(img.shape)
+      input_feed = self.get_input_feed(self.rec_input_name, norm_img_batch)
+      outputs = self.rec_onnx_session.run(
+        self.rec_output_name, input_feed=input_feed
+      )
+
+      preds = outputs[0]
+
+      rec_result = self.postprocess_op(preds)
+      for rno in range(len(rec_result)):
+        rec_res[indices[beg_img_no + rno]] = rec_result[rno]
+
+    return rec_res

doc_page_extractor/onnxocr/predict_system.py

@@ -0,0 +1,97 @@
+import os
+import cv2
+import copy
+
+from . import predict_det
+from . import predict_cls
+from . import predict_rec
+from .utils import get_rotate_crop_image, get_minarea_rect_crop
+
+class TextSystem:
+  def __init__(self, args):
+    self.text_detector = predict_det.TextDetector(args)
+    self.text_recognizer = predict_rec.TextRecognizer(args)
+    self.use_angle_cls = True
+    self.drop_score = args.drop_score
+    if self.use_angle_cls:
+      self.text_classifier = predict_cls.TextClassifier(args)
+
+    self.args = args
+    self.crop_image_res_index = 0
+
+  def draw_crop_rec_res(self, output_dir, img_crop_list, rec_res):
+    os.makedirs(output_dir, exist_ok=True)
+    bbox_num = len(img_crop_list)
+    for bno in range(bbox_num):
+      cv2.imwrite(
+        os.path.join(
+          output_dir, f"mg_crop_{bno + self.crop_image_res_index}.jpg"
+        ),
+        img_crop_list[bno],
+      )
+
+    self.crop_image_res_index += bbox_num
+
+  def __call__(self, img, cls=True):
+    ori_im = img.copy()
+    # 文字检测
+    dt_boxes = self.text_detector(img)
+
+    if dt_boxes is None:
+      return None, None
+
+    img_crop_list = []
+
+    dt_boxes = sorted_boxes(dt_boxes)
+
+    # 图片裁剪
+    for bno in range(len(dt_boxes)):
+      tmp_box = copy.deepcopy(dt_boxes[bno])
+      if self.args.det_box_type == "quad":
+        img_crop = get_rotate_crop_image(ori_im, tmp_box)
+      else:
+        img_crop = get_minarea_rect_crop(ori_im, tmp_box)
+      img_crop_list.append(img_crop)
+
+    # 方向分类
+    if self.use_angle_cls and cls:
+      img_crop_list, angle_list = self.text_classifier(img_crop_list)
+
+    # 图像识别
+    rec_res = self.text_recognizer(img_crop_list)
+
+    if self.args.save_crop_res:
+      self.draw_crop_rec_res(self.args.crop_res_save_dir, img_crop_list, rec_res)
+    filter_boxes, filter_rec_res = [], []
+    for box, rec_result in zip(dt_boxes, rec_res):
+      text, score = rec_result
+      if score >= self.drop_score:
+        filter_boxes.append(box)
+        filter_rec_res.append(rec_result)
+
+    return filter_boxes, filter_rec_res
+
+
+def sorted_boxes(dt_boxes):
+  """
+  Sort text boxes in order from top to bottom, left to right
+  args:
+    dt_boxes(array):detected text boxes with shape [4, 2]
+  return:
+    sorted boxes(array) with shape [4, 2]
+  """
+  num_boxes = dt_boxes.shape[0]
+  sorted_boxes = sorted(dt_boxes, key=lambda x: (x[0][1], x[0][0]))
+  _boxes = list(sorted_boxes)
+
+  for i in range(num_boxes - 1):
+    for j in range(i, -1, -1):
+      if abs(_boxes[j + 1][0][1] - _boxes[j][0][1]) < 10 and (
+        _boxes[j + 1][0][0] < _boxes[j][0][0]
+      ):
+        tmp = _boxes[j]
+        _boxes[j] = _boxes[j + 1]
+        _boxes[j + 1] = tmp
+      else:
+        break
+  return _boxes