merge dygraph

MANIFEST.in

-include LICENSE.txt
+include LICENSE
 include README.md
 
-recursive-include ppocr/utils *.txt utility.py logging.py
+recursive-include ppocr/utils *.txt utility.py logging.py network.py
 recursive-include ppocr/data/ *.py
 recursive-include ppocr/postprocess *.py
 recursive-include tools/infer *.py

configs/rec/ch_ppocr_v2.1/rec_chinese_lite_train_distillation_v2.1.yml

@@ -52,9 +52,10 @@ Architecture:
       Neck:
         name: SequenceEncoder
         encoder_type: rnn
-        hidden_size: 48
+        hidden_size: 64
       Head:
         name: CTCHead
+        mid_channels: 96
         fc_decay: 0.00001
     Teacher:
       pretrained:
@@ -71,9 +72,10 @@ Architecture:
       Neck:
         name: SequenceEncoder
         encoder_type: rnn
-        hidden_size: 48
+        hidden_size: 64
       Head:
         name: CTCHead
+        mid_channels: 96
         fc_decay: 0.00001
   
 

configs/table/table_mv3.yml

+Global:
+  use_gpu: true
+  epoch_num: 50
+  log_smooth_window: 20
+  print_batch_step: 5
+  save_model_dir: ./output/table_mv3/
+  save_epoch_step: 5
+  # evaluation is run every 400 iterations after the 0th iteration
+  eval_batch_step: [0, 400]
+  cal_metric_during_train: True
+  pretrained_model: 
+  checkpoints: 
+  save_inference_dir:
+  use_visualdl: False
+  infer_img: doc/imgs_words/ch/word_1.jpg
+  # for data or label process
+  character_dict_path: ppocr/utils/dict/table_structure_dict.txt
+  character_type: en
+  max_text_length: 100
+  max_elem_length: 500
+  max_cell_num: 500
+  infer_mode: False
+  process_total_num: 0
+  process_cut_num: 0
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  clip_norm: 5.0
+  lr:
+    learning_rate: 0.001
+  regularizer:
+    name: 'L2'
+    factor: 0.00000
+
+Architecture:
+  model_type: table
+  algorithm: TableAttn
+  Backbone:
+    name: MobileNetV3
+    scale: 1.0
+    model_name: small
+    disable_se: True
+  Head:
+    name: TableAttentionHead
+    hidden_size: 256
+    l2_decay: 0.00001
+    loc_type: 2
+
+Loss:
+  name: TableAttentionLoss
+  structure_weight: 100.0
+  loc_weight: 10000.0
+
+PostProcess:
+  name: TableLabelDecode
+
+Metric:
+  name: TableMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: PubTabDataSet
+    data_dir: train_data/table/pubtabnet/train/
+    label_file_path: train_data/table/pubtabnet/PubTabNet_2.0.0_train.jsonl
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ResizeTableImage:
+          max_len: 488
+      - TableLabelEncode:
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - PaddingTableImage:
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'structure', 'bbox_list', 'sp_tokens', 'bbox_list_mask']
+  loader:
+    shuffle: True
+    batch_size_per_card: 32
+    drop_last: True
+    num_workers: 1
+
+Eval:
+  dataset:
+    name: PubTabDataSet
+    data_dir: train_data/table/pubtabnet/val/
+    label_file_path: train_data/table/pubtabnet/PubTabNet_2.0.0_val.jsonl
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - ResizeTableImage:
+          max_len: 488
+      - TableLabelEncode:
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [0.485, 0.456, 0.406]
+          std: [0.229, 0.224, 0.225]
+          order: 'hwc'
+      - PaddingTableImage:
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: ['image', 'structure', 'bbox_list', 'sp_tokens', 'bbox_list_mask']
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 16
+    num_workers: 1

deploy/cpp_infer/src/preprocess_op.cpp

@@ -47,16 +47,13 @@ void Normalize::Run(cv::Mat *im, const std::vector<float> &mean,
     e /= 255.0;
   }
   (*im).convertTo(*im, CV_32FC3, e);
-  for (int h = 0; h < im->rows; h++) {
-    for (int w = 0; w < im->cols; w++) {
-      im->at<cv::Vec3f>(h, w)[0] =
-          (im->at<cv::Vec3f>(h, w)[0] - mean[0]) * scale[0];
-      im->at<cv::Vec3f>(h, w)[1] =
-          (im->at<cv::Vec3f>(h, w)[1] - mean[1]) * scale[1];
-      im->at<cv::Vec3f>(h, w)[2] =
-          (im->at<cv::Vec3f>(h, w)[2] - mean[2]) * scale[2];
-    }
+  std::vector<cv::Mat> bgr_channels(3);
+  cv::split(*im, bgr_channels);
+  for (auto i = 0; i < bgr_channels.size(); i++) {
+    bgr_channels[i].convertTo(bgr_channels[i], CV_32FC1, 1.0 * scale[i],
+                              (0.0 - mean[i]) * scale[i]);
   }
+  cv::merge(bgr_channels, *im);
 }
 
 void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img,

doc/doc_ch/whl.md

@@ -355,3 +355,4 @@ im_show.save('result.jpg')
 | det                     | 前向时使用启动检测                                                                                                                                                                                                   | TRUE                    |
 | rec                     | 前向时是否启动识别                                                                                                                                                                                                   | TRUE                    |
 | cls                     | 前向时是否启动分类 (命令行模式下使用use_angle_cls控制前向是否启动分类)                                                                                                                                                                                                | FALSE                    |
+| show_log                     | 是否打印det和rec等信息                                                                                                                                                                                                | FALSE                    |

doc/doc_en/whl_en.md

@@ -362,3 +362,5 @@ im_show.save('result.jpg')
 | det                     | Enable detction when `ppocr.ocr` func exec                                                                                                                                                                                                   | TRUE                    |
 | rec                     | Enable recognition when `ppocr.ocr` func exec                                                                                                                                                                                                   | TRUE                    |
 | cls                     | Enable classification when `ppocr.ocr` func exec((Use use_angle_cls in command line mode to control whether to start classification in the forward direction)                                                                                                                                                                                                   | FALSE                    |
+| show_log                     | Whether to print log in det and rec 
+                                                                                                                                                                                               | FALSE                    |
\ No newline at end of file

paddleocr.py

@@ -19,17 +19,16 @@ __dir__ = os.path.dirname(__file__)
 sys.path.append(os.path.join(__dir__, ''))
 
 import cv2
+import logging
 import numpy as np
 from pathlib import Path
-import tarfile
-import requests
-from tqdm import tqdm
 
 from tools.infer import predict_system
 from ppocr.utils.logging import get_logger
 
 logger = get_logger()
 from ppocr.utils.utility import check_and_read_gif, get_image_file_list
+from ppocr.utils.network import maybe_download, download_with_progressbar, is_link, confirm_model_dir_url
 from tools.infer.utility import draw_ocr, init_args, str2bool
 
 __all__ = ['PaddleOCR']
@@ -123,50 +122,6 @@ SUPPORT_REC_MODEL = ['CRNN']
 BASE_DIR = os.path.expanduser("~/.paddleocr/")
 
 
-def download_with_progressbar(url, save_path):
-    response = requests.get(url, stream=True)
-    total_size_in_bytes = int(response.headers.get('content-length', 0))
-    block_size = 1024  # 1 Kibibyte
-    progress_bar = tqdm(total=total_size_in_bytes, unit='iB', unit_scale=True)
-    with open(save_path, 'wb') as file:
-        for data in response.iter_content(block_size):
-            progress_bar.update(len(data))
-            file.write(data)
-    progress_bar.close()
-    if total_size_in_bytes == 0 or progress_bar.n != total_size_in_bytes:
-        logger.error("Something went wrong while downloading models")
-        sys.exit(0)
-
-
-def maybe_download(model_storage_directory, url):
-    # using custom model
-    tar_file_name_list = [
-        'inference.pdiparams', 'inference.pdiparams.info', 'inference.pdmodel'
-    ]
-    if not os.path.exists(
-            os.path.join(model_storage_directory, 'inference.pdiparams')
-    ) or not os.path.exists(
-            os.path.join(model_storage_directory, 'inference.pdmodel')):
-        tmp_path = os.path.join(model_storage_directory, url.split('/')[-1])
-        print('download {} to {}'.format(url, tmp_path))
-        os.makedirs(model_storage_directory, exist_ok=True)
-        download_with_progressbar(url, tmp_path)
-        with tarfile.open(tmp_path, 'r') as tarObj:
-            for member in tarObj.getmembers():
-                filename = None
-                for tar_file_name in tar_file_name_list:
-                    if tar_file_name in member.name:
-                        filename = tar_file_name
-                if filename is None:
-                    continue
-                file = tarObj.extractfile(member)
-                with open(
-                        os.path.join(model_storage_directory, filename),
-                        'wb') as f:
-                    f.write(file.read())
-        os.remove(tmp_path)
-
-
 def parse_args(mMain=True):
     import argparse
     parser = init_args()
@@ -194,10 +149,12 @@ class PaddleOCR(predict_system.TextSystem):
         args:
             **kwargs: other params show in paddleocr --help
         """
-        postprocess_params = parse_args(mMain=False)
-        postprocess_params.__dict__.update(**kwargs)
-        self.use_angle_cls = postprocess_params.use_angle_cls
-        lang = postprocess_params.lang
+        params = parse_args(mMain=False)
+        params.__dict__.update(**kwargs)
+        if not params.show_log:
+            logger.setLevel(logging.INFO)
+        self.use_angle_cls = params.use_angle_cls
+        lang = params.lang
         latin_lang = [
             'af', 'az', 'bs', 'cs', 'cy', 'da', 'de', 'es', 'et', 'fr', 'ga',
             'hr', 'hu', 'id', 'is', 'it', 'ku', 'la', 'lt', 'lv', 'mi', 'ms',
@@ -229,40 +186,39 @@ class PaddleOCR(predict_system.TextSystem):
         else:
             det_lang = "en"
         use_inner_dict = False
-        if postprocess_params.rec_char_dict_path is None:
+        if params.rec_char_dict_path is None:
             use_inner_dict = True
-            postprocess_params.rec_char_dict_path = model_urls['rec'][lang][
+            params.rec_char_dict_path = model_urls['rec'][lang][
                 'dict_path']
 
         # init model dir
-        if postprocess_params.det_model_dir is None:
-            postprocess_params.det_model_dir = os.path.join(BASE_DIR, VERSION,
-                                                            'det', det_lang)
-        if postprocess_params.rec_model_dir is None:
-            postprocess_params.rec_model_dir = os.path.join(BASE_DIR, VERSION,
-                                                            'rec', lang)
-        if postprocess_params.cls_model_dir is None:
-            postprocess_params.cls_model_dir = os.path.join(BASE_DIR, 'cls')
-        print(postprocess_params)
-        # download model
-        maybe_download(postprocess_params.det_model_dir,
+        params.det_model_dir, det_url = confirm_model_dir_url(params.det_model_dir,
+                                                              os.path.join(BASE_DIR, VERSION, 'det', det_lang),
                                                               model_urls['det'][det_lang])
-        maybe_download(postprocess_params.rec_model_dir,
+        params.rec_model_dir, rec_url = confirm_model_dir_url(params.rec_model_dir,
+                                                              os.path.join(BASE_DIR, VERSION, 'rec', lang),
                                                               model_urls['rec'][lang]['url'])
-        maybe_download(postprocess_params.cls_model_dir, model_urls['cls'])
+        params.cls_model_dir, cls_url = confirm_model_dir_url(params.cls_model_dir,
+                                                              os.path.join(BASE_DIR, VERSION, 'cls'),
+                                                              model_urls['cls'])
+        # download model
+        maybe_download(params.det_model_dir, det_url)
+        maybe_download(params.rec_model_dir, rec_url)
+        maybe_download(params.cls_model_dir, cls_url)
 
-        if postprocess_params.det_algorithm not in SUPPORT_DET_MODEL:
+        if params.det_algorithm not in SUPPORT_DET_MODEL:
             logger.error('det_algorithm must in {}'.format(SUPPORT_DET_MODEL))
             sys.exit(0)
-        if postprocess_params.rec_algorithm not in SUPPORT_REC_MODEL:
+        if params.rec_algorithm not in SUPPORT_REC_MODEL:
             logger.error('rec_algorithm must in {}'.format(SUPPORT_REC_MODEL))
             sys.exit(0)
         if use_inner_dict:
-            postprocess_params.rec_char_dict_path = str(
-                Path(__file__).parent / postprocess_params.rec_char_dict_path)
+            params.rec_char_dict_path = str(
+                Path(__file__).parent / params.rec_char_dict_path)
 
+        print(params)
         # init det_model and rec_model
-        super().__init__(postprocess_params)
+        super().__init__(params)
 
     def ocr(self, img, det=True, rec=True, cls=True):
         """
@@ -320,7 +276,7 @@ def main():
     # for cmd
     args = parse_args(mMain=True)
     image_dir = args.image_dir
-    if image_dir.startswith('http'):
+    if is_link(image_dir):
         download_with_progressbar(image_dir, 'tmp.jpg')
         image_file_list = ['tmp.jpg']
     else:

ppocr/data/__init__.py

@@ -35,6 +35,7 @@ from ppocr.data.imaug import transform, create_operators
 from ppocr.data.simple_dataset import SimpleDataSet
 from ppocr.data.lmdb_dataset import LMDBDataSet
 from ppocr.data.pgnet_dataset import PGDataSet
+from ppocr.data.pubtab_dataset import PubTabDataSet
 
 __all__ = ['build_dataloader', 'transform', 'create_operators']
 
@@ -55,7 +56,7 @@ signal.signal(signal.SIGTERM, term_mp)
 def build_dataloader(config, mode, device, logger, seed=None):
     config = copy.deepcopy(config)
 
-    support_dict = ['SimpleDataSet', 'LMDBDataSet', 'PGDataSet']
+    support_dict = ['SimpleDataSet', 'LMDBDataSet', 'PGDataSet', 'PubTabDataSet']
     module_name = config[mode]['dataset']['name']
     assert module_name in support_dict, Exception(
         'DataSet only support {}'.format(support_dict))

ppocr/data/imaug/__init__.py

@@ -30,6 +30,7 @@ from .label_ops import *
 from .east_process import *
 from .sast_process import *
 from .pg_process import *
+from .gen_table_mask import *
 
 
 def transform(data, ops=None):

ppocr/data/imaug/gen_table_mask.py

+"""
+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import sys
+import six
+import cv2
+import numpy as np
+
+
+class GenTableMask(object):
+    """ gen table mask """
+
+    def __init__(self, shrink_h_max, shrink_w_max, mask_type=0, **kwargs):
+        self.shrink_h_max = 5
+        self.shrink_w_max = 5
+        self.mask_type = mask_type
+        
+    def projection(self, erosion, h, w, spilt_threshold=0):
+        # 水平投影
+        projection_map = np.ones_like(erosion)
+        project_val_array = [0 for _ in range(0, h)]
+
+        for j in range(0, h):
+            for i in range(0, w):
+                if erosion[j, i] == 255:
+                    project_val_array[j] += 1
+        # 根据数组，获取切割点
+        start_idx = 0  # 记录进入字符区的索引
+        end_idx = 0  # 记录进入空白区域的索引
+        in_text = False  # 是否遍历到了字符区内
+        box_list = []
+        for i in range(len(project_val_array)):
+            if in_text == False and project_val_array[i] > spilt_threshold:  # 进入字符区了
+                in_text = True
+                start_idx = i
+            elif project_val_array[i] <= spilt_threshold and in_text == True:  # 进入空白区了
+                end_idx = i
+                in_text = False
+                if end_idx - start_idx <= 2:
+                    continue
+                box_list.append((start_idx, end_idx + 1))
+
+        if in_text:
+            box_list.append((start_idx, h - 1))
+        # 绘制投影直方图
+        for j in range(0, h):
+            for i in range(0, project_val_array[j]):
+                projection_map[j, i] = 0
+        return box_list, projection_map
+
+    def projection_cx(self, box_img):
+        box_gray_img = cv2.cvtColor(box_img, cv2.COLOR_BGR2GRAY)
+        h, w = box_gray_img.shape
+        # 灰度图片进行二值化处理
+        ret, thresh1 = cv2.threshold(box_gray_img, 200, 255, cv2.THRESH_BINARY_INV)
+        # 纵向腐蚀
+        if h < w:
+            kernel = np.ones((2, 1), np.uint8)
+            erode = cv2.erode(thresh1, kernel, iterations=1)
+        else:
+            erode = thresh1
+        # 水平膨胀
+        kernel = np.ones((1, 5), np.uint8)
+        erosion = cv2.dilate(erode, kernel, iterations=1)
+        # 水平投影
+        projection_map = np.ones_like(erosion)
+        project_val_array = [0 for _ in range(0, h)]
+
+        for j in range(0, h):
+            for i in range(0, w):
+                if erosion[j, i] == 255:
+                    project_val_array[j] += 1
+        # 根据数组，获取切割点
+        start_idx = 0  # 记录进入字符区的索引
+        end_idx = 0  # 记录进入空白区域的索引
+        in_text = False  # 是否遍历到了字符区内
+        box_list = []
+        spilt_threshold = 0
+        for i in range(len(project_val_array)):
+            if in_text == False and project_val_array[i] > spilt_threshold:  # 进入字符区了
+                in_text = True
+                start_idx = i
+            elif project_val_array[i] <= spilt_threshold and in_text == True:  # 进入空白区了
+                end_idx = i
+                in_text = False
+                if end_idx - start_idx <= 2:
+                    continue
+                box_list.append((start_idx, end_idx + 1))
+
+        if in_text:
+            box_list.append((start_idx, h - 1))
+        # 绘制投影直方图
+        for j in range(0, h):
+            for i in range(0, project_val_array[j]):
+                projection_map[j, i] = 0
+        split_bbox_list = []
+        if len(box_list) > 1:
+            for i, (h_start, h_end) in enumerate(box_list):
+                if i == 0:
+                    h_start = 0
+                if i == len(box_list):
+                    h_end = h
+                word_img = erosion[h_start:h_end + 1, :]
+                word_h, word_w = word_img.shape
+                w_split_list, w_projection_map = self.projection(word_img.T, word_w, word_h)
+                w_start, w_end = w_split_list[0][0], w_split_list[-1][1]
+                if h_start > 0:
+                    h_start -= 1
+                h_end += 1
+                word_img = box_img[h_start:h_end + 1:, w_start:w_end + 1, :]
+                split_bbox_list.append([w_start, h_start, w_end, h_end])
+        else:
+            split_bbox_list.append([0, 0, w, h])
+        return split_bbox_list
+
+    def shrink_bbox(self, bbox):
+        left, top, right, bottom = bbox
+        sh_h = min(max(int((bottom - top) * 0.1), 1), self.shrink_h_max)
+        sh_w = min(max(int((right - left) * 0.1), 1), self.shrink_w_max)
+        left_new = left + sh_w
+        right_new = right - sh_w
+        top_new = top + sh_h
+        bottom_new = bottom - sh_h
+        if left_new >= right_new:
+            left_new = left
+            right_new = right
+        if top_new >= bottom_new:
+            top_new = top
+            bottom_new = bottom
+        return [left_new, top_new, right_new, bottom_new]
+
+    def __call__(self, data):
+        img = data['image']
+        cells = data['cells']
+        height, width = img.shape[0:2]
+        if self.mask_type == 1:
+            mask_img = np.zeros((height, width), dtype=np.float32)
+        else:
+            mask_img = np.zeros((height, width, 3), dtype=np.float32)
+        cell_num = len(cells)
+        for cno in range(cell_num):
+            if "bbox" in cells[cno]:
+                bbox = cells[cno]['bbox']
+                left, top, right, bottom = bbox
+                box_img = img[top:bottom, left:right, :].copy()
+                split_bbox_list = self.projection_cx(box_img)
+                for sno in range(len(split_bbox_list)):
+                    split_bbox_list[sno][0] += left
+                    split_bbox_list[sno][1] += top
+                    split_bbox_list[sno][2] += left
+                    split_bbox_list[sno][3] += top
+
+                for sno in range(len(split_bbox_list)):
+                    left, top, right, bottom = split_bbox_list[sno]
+                    left, top, right, bottom = self.shrink_bbox([left, top, right, bottom])
+                    if self.mask_type == 1:
+                        mask_img[top:bottom, left:right] = 1.0
+                        data['mask_img'] = mask_img
+                    else:
+                        mask_img[top:bottom, left:right, :] = (255, 255, 255)        
+                        data['image'] = mask_img
+        return data
+
+class ResizeTableImage(object):
+    def __init__(self, max_len, **kwargs):
+        super(ResizeTableImage, self).__init__()
+        self.max_len = max_len
+
+    def get_img_bbox(self, cells):
+        bbox_list = []
+        if len(cells) == 0:
+            return bbox_list
+        cell_num = len(cells)
+        for cno in range(cell_num):
+            if "bbox" in cells[cno]:
+                bbox = cells[cno]['bbox']
+                bbox_list.append(bbox)
+        return bbox_list
+
+    def resize_img_table(self, img, bbox_list, max_len):
+        height, width = img.shape[0:2]
+        ratio = max_len / (max(height, width) * 1.0)
+        resize_h = int(height * ratio)
+        resize_w = int(width * ratio)
+        img_new = cv2.resize(img, (resize_w, resize_h))
+        bbox_list_new = []
+        for bno in range(len(bbox_list)):
+            left, top, right, bottom = bbox_list[bno].copy()
+            left = int(left * ratio)
+            top = int(top * ratio)
+            right = int(right * ratio)
+            bottom = int(bottom * ratio)
+            bbox_list_new.append([left, top, right, bottom])
+        return img_new, bbox_list_new
+    
+    def __call__(self, data):
+        img = data['image']
+        if 'cells' not in data:
+            cells = []
+        else:
+            cells = data['cells']
+        bbox_list = self.get_img_bbox(cells)
+        img_new, bbox_list_new = self.resize_img_table(img, bbox_list, self.max_len)
+        data['image'] = img_new
+        cell_num = len(cells)
+        bno = 0
+        for cno in range(cell_num):
+            if "bbox" in data['cells'][cno]:
+                data['cells'][cno]['bbox'] = bbox_list_new[bno]
+                bno += 1
+        data['max_len'] = self.max_len
+        return data
+
+class PaddingTableImage(object):
+    def __init__(self, **kwargs):
+        super(PaddingTableImage, self).__init__()
+    
+    def __call__(self, data):
+        img = data['image']
+        max_len = data['max_len']
+        padding_img = np.zeros((max_len, max_len, 3), dtype=np.float32)
+        height, width = img.shape[0:2]
+        padding_img[0:height, 0:width, :] = img.copy()
+        data['image'] = padding_img
+        return data
+            
\ No newline at end of file

ppocr/data/imaug/label_ops.py

@@ -351,3 +351,162 @@ class SRNLabelEncode(BaseRecLabelEncode):
             assert False, "Unsupport type %s in get_beg_end_flag_idx" \
                           % beg_or_end
         return idx
+
+class TableLabelEncode(object):
+    """ Convert between text-label and text-index """
+    def __init__(self, 
+        max_text_length,
+        max_elem_length,
+        max_cell_num,
+        character_dict_path,
+        span_weight = 1.0, 
+        **kwargs):
+        self.max_text_length = max_text_length
+        self.max_elem_length = max_elem_length
+        self.max_cell_num = max_cell_num
+        list_character, list_elem = self.load_char_elem_dict(character_dict_path)
+        list_character = self.add_special_char(list_character)
+        list_elem = self.add_special_char(list_elem)
+        self.dict_character = {}
+        for i, char in enumerate(list_character):
+            self.dict_character[char] = i
+        self.dict_elem = {}
+        for i, elem in enumerate(list_elem):
+            self.dict_elem[elem] = i
+        self.span_weight = span_weight
+    
+    def load_char_elem_dict(self, character_dict_path):
+        list_character = []
+        list_elem = []
+        with open(character_dict_path, "rb") as fin:
+            lines = fin.readlines()
+            substr = lines[0].decode('utf-8').strip("\n").split("\t")
+            character_num = int(substr[0])
+            elem_num = int(substr[1])
+            for cno in range(1, 1+character_num):
+                character = lines[cno].decode('utf-8').strip("\n")
+                list_character.append(character)
+            for eno in range(1+character_num, 1+character_num+elem_num):
+                elem = lines[eno].decode('utf-8').strip("\n")
+                list_elem.append(elem)
+        return list_character, list_elem
+        
+    def add_special_char(self, list_character):
+        self.beg_str = "sos"
+        self.end_str = "eos"
+        list_character = [self.beg_str] + list_character + [self.end_str]
+        return list_character
+    
+    def get_span_idx_list(self):
+        span_idx_list = []
+        for elem in self.dict_elem:
+            if 'span' in elem:
+                span_idx_list.append(self.dict_elem[elem])
+        return span_idx_list
+        
+    def __call__(self, data):
+        cells = data['cells']
+        structure = data['structure']['tokens']
+        structure = self.encode(structure, 'elem')
+        if structure is None:
+            return None
+        elem_num = len(structure)
+        structure = [0] + structure + [len(self.dict_elem) - 1]
+        structure = structure + [0] * (self.max_elem_length + 2 - len(structure))
+        structure = np.array(structure)
+        data['structure'] = structure
+        elem_char_idx1 = self.dict_elem['<td>']
+        elem_char_idx2 = self.dict_elem['<td']
+        span_idx_list = self.get_span_idx_list()
+        td_idx_list = np.logical_or(structure == elem_char_idx1, structure == elem_char_idx2)
+        td_idx_list = np.where(td_idx_list)[0]
+        
+        structure_mask = np.ones((self.max_elem_length + 2, 1), dtype=np.float32)
+        bbox_list = np.zeros((self.max_elem_length + 2, 4), dtype=np.float32)
+        bbox_list_mask = np.zeros((self.max_elem_length + 2, 1), dtype=np.float32)
+        img_height, img_width, img_ch = data['image'].shape
+        if len(span_idx_list) > 0:
+            span_weight = len(td_idx_list) * 1.0 / len(span_idx_list)
+            span_weight = min(max(span_weight, 1.0), self.span_weight)
+        for cno in range(len(cells)):
+            if 'bbox' in cells[cno]:
+                bbox = cells[cno]['bbox'].copy()
+                bbox[0] = bbox[0] * 1.0 / img_width
+                bbox[1] = bbox[1] * 1.0 / img_height
+                bbox[2] = bbox[2] * 1.0 / img_width
+                bbox[3] = bbox[3] * 1.0 / img_height
+                td_idx = td_idx_list[cno]
+                bbox_list[td_idx] = bbox
+                bbox_list_mask[td_idx] = 1.0
+                cand_span_idx = td_idx + 1
+                if cand_span_idx < (self.max_elem_length + 2):
+                    if structure[cand_span_idx] in span_idx_list:
+                        structure_mask[cand_span_idx] = span_weight
+
+        data['bbox_list'] = bbox_list
+        data['bbox_list_mask'] = bbox_list_mask
+        data['structure_mask'] = structure_mask
+        char_beg_idx = self.get_beg_end_flag_idx('beg', 'char')
+        char_end_idx = self.get_beg_end_flag_idx('end', 'char')
+        elem_beg_idx = self.get_beg_end_flag_idx('beg', 'elem')
+        elem_end_idx = self.get_beg_end_flag_idx('end', 'elem')
+        data['sp_tokens'] = np.array([char_beg_idx, char_end_idx, elem_beg_idx, 
+            elem_end_idx, elem_char_idx1, elem_char_idx2, self.max_text_length, 
+            self.max_elem_length, self.max_cell_num, elem_num])
+        return data
+
+    def encode(self, text, char_or_elem):
+        """convert text-label into text-index.
+        """
+        if char_or_elem == "char":
+            max_len = self.max_text_length
+            current_dict = self.dict_character
+        else:
+            max_len = self.max_elem_length
+            current_dict = self.dict_elem
+        if len(text) > max_len:
+            return None
+        if len(text) == 0:
+            if char_or_elem == "char":
+                return [self.dict_character['space']]
+            else:
+                return None
+        text_list = []
+        for char in text:
+            if char not in current_dict:
+                return None
+            text_list.append(current_dict[char])
+        if len(text_list) == 0:
+            if char_or_elem == "char":
+                return [self.dict_character['space']]
+            else:
+                return None
+        return text_list
+
+    def get_ignored_tokens(self, char_or_elem):
+        beg_idx = self.get_beg_end_flag_idx("beg", char_or_elem)
+        end_idx = self.get_beg_end_flag_idx("end", char_or_elem)
+        return [beg_idx, end_idx]
+
+    def get_beg_end_flag_idx(self, beg_or_end, char_or_elem):
+        if char_or_elem == "char":
+            if beg_or_end == "beg":
+                idx = np.array(self.dict_character[self.beg_str])
+            elif beg_or_end == "end":
+                idx = np.array(self.dict_character[self.end_str])
+            else:
+                assert False, "Unsupport type %s in get_beg_end_flag_idx of char" \
+                              % beg_or_end
+        elif char_or_elem == "elem":
+            if beg_or_end == "beg":
+                idx = np.array(self.dict_elem[self.beg_str])
+            elif beg_or_end == "end":
+                idx = np.array(self.dict_elem[self.end_str])
+            else:
+                assert False, "Unsupport type %s in get_beg_end_flag_idx of elem" \
+                              % beg_or_end  
+        else:
+            assert False, "Unsupport type %s in char_or_elem" \
+                              % char_or_elem 
+        return idx
+    
\ No newline at end of file

ppocr/data/imaug/operators.py

@@ -163,7 +163,7 @@ class DetResizeForTest(object):
             img, (ratio_h, ratio_w)
         """
         limit_side_len = self.limit_side_len
-        h, w, _ = img.shape
+        h, w, c = img.shape
 
         # limit the max side
         if self.limit_type == 'max':
@@ -174,7 +174,7 @@ class DetResizeForTest(object):
                     ratio = float(limit_side_len) / w
             else:
                 ratio = 1.
-        else:
+        elif self.limit_type == 'min':
             if min(h, w) < limit_side_len:
                 if h < w:
                     ratio = float(limit_side_len) / h
@@ -182,6 +182,10 @@ class DetResizeForTest(object):
                     ratio = float(limit_side_len) / w
             else:
                 ratio = 1.
+        elif self.limit_type == 'resize_long':
+            ratio = float(limit_side_len) / max(h,w)
+        else:
+            raise Exception('not support limit type, image ')
         resize_h = int(h * ratio)
         resize_w = int(w * ratio)