add vqa_ser to ppstructure predict pipeline

ppstructure/README.md

@@ -153,7 +153,7 @@ wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_in
 wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar && tar xf en_ppocr_mobile_v2.0_table_structure_infer.tar
 cd ..
 
-python3 predict_system.py --det_model_dir=inference/ch_ppocr_mobile_v2.0_det_infer --rec_model_dir=inference/ch_ppocr_mobile_v2.0_rec_infer --table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer --image_dir=../doc/table/1.png --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt --table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt --rec_char_type=ch --output=../output/table --vis_font_path=../doc/fonts/simfang.ttf
+python3 predict_system.py --det_model_dir=inference/ch_ppocr_mobile_v2.0_det_infer --rec_model_dir=inference/ch_ppocr_mobile_v2.0_rec_infer --table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer --image_dir=../doc/table/1.png --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt --table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt --output=../output/table --vis_font_path=../doc/fonts/simfang.ttf
 ```
 After running, each image will have a directory with the same name under the directory specified in the output field. Each table in the picture will be stored as an excel and figure area will be cropped and saved, the excel and image file name will be the coordinates of the table in the image.
 

ppstructure/README_ch.md

@@ -8,11 +8,37 @@ PP-Structure是一个可用于复杂文档结构分析和处理的OCR工具包
 - 支持表格区域进行结构化分析，最终结果输出Excel文件
 - 支持python whl包和命令行两种方式，简单易用
 - 支持版面分析和表格结构化两类任务自定义训练
+- 支持文档关键信息提取-SER和RE任务
+
 
 ## 1. 效果展示
 
+### 1.1 版面分析和表格识别
+
 <img src="../doc/table/ppstructure.GIF" width="100%"/>
 
+### 1.2 VQA
+
+* SER
+
+![](./vqa/images/result_ser/zh_val_0_ser.jpg) | ![](./vqa/images/result_ser/zh_val_42_ser.jpg)
+---|---
+
+图中不同颜色的框表示不同的类别，对于XFUN数据集，有`QUESTION`, `ANSWER`, `HEADER` 3种类别
+
+* 深紫色：HEADER
+* 浅紫色：QUESTION
+* 军绿色：ANSWER
+
+在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
+
+* RE
+
+![](./vqa/images/result_re/zh_val_21_re.jpg) | ![](./vqa/images/result_re/zh_val_40_re.jpg)
+---|---
+
+
+图中红色框表示问题，蓝色框表示答案，问题和答案之间使用绿色线连接。在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
 
 
 ## 2. 安装
@@ -33,18 +59,24 @@ python3 -m pip install paddlepaddle-gpu==2.1.1 -i https://mirror.baidu.com/pypi/
 ```
 更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
 
-- **(2) 安装 Layout-Parser**
+- **(2) 安装依赖 **
 
 ```bash
+# 版面分析所需 Layout-Parser
 pip3 install -U https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
+
+# VQA所需 PaddleNLP
+git clone https://github.com/PaddlePaddle/PaddleNLP -b develop
+cd PaddleNLP
+pip3 install -e .
 ```
 
 ### 2.2 安装PaddleOCR（包含PP-OCR和PP-Structure）
 
 - **（1) PIP快速安装PaddleOCR whl包（仅预测）**
 
 ```bash
-pip install "paddleocr>=2.2" # 推荐使用2.2+版本
+pip3 install "paddleocr>=2.2" # 推荐使用2.2+版本
 ```
 
 - **（2) 完整克隆PaddleOCR源码（预测+训练）**
@@ -63,12 +95,14 @@ git clone https://gitee.com/paddlepaddle/PaddleOCR
 
 ### 3.1 命令行使用（默认参数，极简）
 
+* 版面分析+表格识别
 ```bash
 paddleocr --image_dir=../doc/table/1.png --type=structure
 ```
 
 ### 3.2 Python脚本使用（自定义参数，灵活）
 
+* 版面分析+表格识别
 ```python
 import os
 import cv2
@@ -98,6 +132,7 @@ im_show.save('result.jpg')
 ### 3.3 返回结果说明
 PP-Structure的返回结果为一个dict组成的list，示例如下
 
+* 版面分析+表格识别
 ```shell
 [
  { 'type': 'Text',
@@ -130,7 +165,7 @@ dict 里各个字段说明如下
 运行完成后，每张图片会在`output`字段指定的目录下有一个同名目录，图片里的每个表格会存储为一个excel，图片区域会被裁剪之后保存下来，excel文件和图片名名为表格在图片里的坐标。
 
 
-## 4. PP-Structure Pipeline介绍
+## 4. PP-Structure 版面分析+表格识别 Pipeline介绍
 
 ![pipeline](../doc/table/pipeline.jpg)
 
@@ -148,6 +183,8 @@ dict 里各个字段说明如下
 
 使用如下命令即可完成预测引擎的推理
 
+* 版面分析+表格识别
+
 ```python
 cd ppstructure
 
@@ -161,9 +198,24 @@ wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_in
 wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar && tar xf en_ppocr_mobile_v2.0_table_structure_infer.tar
 cd ..
 
-python3 predict_system.py --det_model_dir=inference/ch_ppocr_mobile_v2.0_det_infer --rec_model_dir=inference/ch_ppocr_mobile_v2.0_rec_infer --table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer --image_dir=../doc/table/1.png --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt --table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt --rec_char_type=ch --output=../output/table --vis_font_path=../doc/fonts/simfang.ttf
+python3 predict_system.py --det_model_dir=inference/ch_ppocr_mobile_v2.0_det_infer --rec_model_dir=inference/ch_ppocr_mobile_v2.0_rec_infer --table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer --image_dir=../doc/table/1.png --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt --table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt --output=../output/table --vis_font_path=../doc/fonts/simfang.ttf
 ```
-运行完成后，每张图片会在`output`字段指定的目录下有一个同名目录，图片里的每个表格会存储为一个excel，图片区域会被裁剪之后保存下来，excel文件和图片名名为表格在图片里的坐标。
+运行完成后，每张图片会在`output`字段指定的目录下的`talbe`目录下有一个同名目录，图片里的每个表格会存储为一个excel，图片区域会被裁剪之后保存下来，excel文件和图片名名为表格在图片里的坐标。
+
+* VQA
+
+```python
+cd ppstructure
+
+# 下载模型
+mkdir inference && cd inference
+# 下载SER xfun 模型并解压
+wget https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar && tar xf PP-Layout_v1.0_ser_pretrained.tar
+cd ..
+
+python3 predict_system.py --model_name_or_path=vqa/PP-Layout_v1.0_ser_pretrained/ --mode=vqa --image_dir=vqa/images/input/zh_val_0.jpg --vis_font_path=../doc/fonts/simfang.ttf
+```
+运行完成后，每张图片会在`output`字段指定的目录下的`vqa`目录下存放可视化之后的图片，图片名和输入图片名一致。
 
 **Model List**
 
@@ -185,4 +237,11 @@ OCR和表格识别模型
 |en_ppocr_mobile_v2.0_table_rec|PubLayNet数据集训练的英文表格场景的文字识别|6.9M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_rec_train.tar) |
 |en_ppocr_mobile_v2.0_table_structure|PubLayNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
 
+VQA
+
+|模型名称|模型简介|推理模型大小|下载地址|
+| --- | --- | --- | --- |
+|PP-Layout_v1.0_ser_pretrained|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar) |
+|PP-Layout_v1.0_re_pretrained|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_re_pretrained.tar) |
+
 如需要使用其他模型，可以在 [model_list](../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到`det_model_dir`,`rec_model_dir`,`table_model_dir`三个字段即可。

ppstructure/predict_system.py

@@ -30,90 +30,119 @@
 from ppocr.utils.logging import get_logger
 from tools.infer.predict_system import TextSystem
 from ppstructure.table.predict_table import TableSystem, to_excel
+from ppstructure.vqa.infer_ser_e2e import SerPredictor, draw_ser_results
 from ppstructure.utility import parse_args, draw_structure_result
 
 logger = get_logger()
 
 
 class OCRSystem(object):
  def __init__(self, args):
- import layoutparser as lp
- # args.det_limit_type = 'resize_long'
- args.drop_score = 0
- if not args.show_log:
- logger.setLevel(logging.INFO)
- self.text_system = TextSystem(args)
- self.table_system = TableSystem(args, self.text_system.text_detector, self.text_system.text_recognizer)
-
- config_path = None
- model_path = None
- if os.path.isdir(args.layout_path_model):
- model_path = args.layout_path_model
- else:
- config_path = args.layout_path_model
- self.table_layout = lp.PaddleDetectionLayoutModel(config_path=config_path,
- model_path=model_path,
- threshold=0.5, enable_mkldnn=args.enable_mkldnn,
- enforce_cpu=not args.use_gpu, thread_num=args.cpu_threads)
- self.use_angle_cls = args.use_angle_cls
- self.drop_score = args.drop_score
+ self.mode = args.mode
+ if self.mode == 'structure':
+ import layoutparser as lp
+ # args.det_limit_type = 'resize_long'
+ args.drop_score = 0
+ if not args.show_log:
+ logger.setLevel(logging.INFO)
+ self.text_system = TextSystem(args)
+ self.table_system = TableSystem(args,
+ self.text_system.text_detector,
+ self.text_system.text_recognizer)
+
+ config_path = None
+ model_path = None
+ if os.path.isdir(args.layout_path_model):
+ model_path = args.layout_path_model
+ else:
+ config_path = args.layout_path_model
+ self.table_layout = lp.PaddleDetectionLayoutModel(
+ config_path=config_path,
+ model_path=model_path,
+ threshold=0.5,
+ enable_mkldnn=args.enable_mkldnn,
+ enforce_cpu=not args.use_gpu,
+ thread_num=args.cpu_threads)
+ self.use_angle_cls = args.use_angle_cls
+ self.drop_score = args.drop_score
+ elif self.mode == 'vqa':
+ self.vqa_engine = SerPredictor(args)
 
  def __call__(self, img):
- ori_im = img.copy()
- layout_res = self.table_layout.detect(img[..., ::-1])
- res_list = []
- for region in layout_res:
- x1, y1, x2, y2 = region.coordinates
- x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
- roi_img = ori_im[y1:y2, x1:x2, :]
- if region.type == 'Table':
- res = self.table_system(roi_img)
- else:
- filter_boxes, filter_rec_res = self.text_system(roi_img)
- filter_boxes = [x + [x1, y1] for x in filter_boxes]
- filter_boxes = [x.reshape(-1).tolist() for x in filter_boxes]
- # remove style char
- style_token = ['<strike>', '<strike>', '<sup>', '</sub>', '<b>', '</b>', '<sub>', '</sup>',
- '<overline>', '</overline>', '<underline>', '</underline>', '<i>', '</i>']
- filter_rec_res_tmp = []
- for rec_res in filter_rec_res:
- rec_str, rec_conf = rec_res
- for token in style_token:
- if token in rec_str:
- rec_str = rec_str.replace(token, '')
- filter_rec_res_tmp.append((rec_str, rec_conf))
- res = (filter_boxes, filter_rec_res_tmp)
- res_list.append({'type': region.type, 'bbox': [x1, y1, x2, y2], 'img': roi_img, 'res': res})
+ if self.mode == 'structure':
+ ori_im = img.copy()
+ layout_res = self.table_layout.detect(img[..., ::-1])
+ res_list = []
+ for region in layout_res:
+ x1, y1, x2, y2 = region.coordinates
+ x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
+ roi_img = ori_im[y1:y2, x1:x2, :]
+ if region.type == 'Table':
+ res = self.table_system(roi_img)
+ else:
+ filter_boxes, filter_rec_res = self.text_system(roi_img)
+ filter_boxes = [x + [x1, y1] for x in filter_boxes]
+ filter_boxes = [
+ x.reshape(-1).tolist() for x in filter_boxes
+ ]
+ # remove style char
+ style_token = [
+ '<strike>', '<strike>', '<sup>', '</sub>', '<b>',
+ '</b>', '<sub>', '</sup>', '<overline>', '</overline>',
+ '<underline>', '</underline>', '<i>', '</i>'
+ ]
+ filter_rec_res_tmp = []
+ for rec_res in filter_rec_res:
+ rec_str, rec_conf = rec_res
+ for token in style_token:
+ if token in rec_str:
+ rec_str = rec_str.replace(token, '')
+ filter_rec_res_tmp.append((rec_str, rec_conf))
+ res = (filter_boxes, filter_rec_res_tmp)
+ res_list.append({
+ 'type': region.type,
+ 'bbox': [x1, y1, x2, y2],
+ 'img': roi_img,
+ 'res': res
+ })
+ elif self.mode == 'vqa':
+ res_list, _ = self.vqa_engine(img)
  return res_list
 
 
 def save_structure_res(res, save_folder, img_name):
  excel_save_folder = os.path.join(save_folder, img_name)
  os.makedirs(excel_save_folder, exist_ok=True)
  # save res
- with open(os.path.join(excel_save_folder, 'res.txt'), 'w', encoding='utf8') as f:
+ with open(
+ os.path.join(excel_save_folder, 'res.txt'), 'w',
+ encoding='utf8') as f:
  for region in res:
  if region['type'] == 'Table':
- excel_path = os.path.join(excel_save_folder, '{}.xlsx'.format(region['bbox']))
+ excel_path = os.path.join(excel_save_folder,
+ '{}.xlsx'.format(region['bbox']))
  to_excel(region['res'], excel_path)
  if region['type'] == 'Figure':
  roi_img = region['img']
- img_path = os.path.join(excel_save_folder, '{}.jpg'.format(region['bbox']))
+ img_path = os.path.join(excel_save_folder,
+ '{}.jpg'.format(region['bbox']))
  cv2.imwrite(img_path, roi_img)
  else:
  for box, rec_res in zip(region['res'][0], region['res'][1]):
- f.write('{}\t{}\n'.format(np.array(box).reshape(-1).tolist(), rec_res))
+ f.write('{}\t{}\n'.format(
+ np.array(box).reshape(-1).tolist(), rec_res))
 
 
 def main(args):
  image_file_list = get_image_file_list(args.image_dir)
  image_file_list = image_file_list
  image_file_list = image_file_list[args.process_id::args.total_process_num]
- save_folder = args.output
- os.makedirs(save_folder, exist_ok=True)
 
  structure_sys = OCRSystem(args)
  img_num = len(image_file_list)
+ save_folder = os.path.join(args.output, structure_sys.mode)
+ os.makedirs(save_folder, exist_ok=True)
+
  for i, image_file in enumerate(image_file_list):
  logger.info("[{}/{}] {}".format(i, img_num, image_file))
  img, flag = check_and_read_gif(image_file)
@@ -126,10 +155,16 @@ def main(args):
  continue
  starttime = time.time()
  res = structure_sys(img)
- save_structure_res(res, save_folder, img_name)
- draw_img = draw_structure_result(img, res, args.vis_font_path)
- cv2.imwrite(os.path.join(save_folder, img_name, 'show.jpg'), draw_img)
- logger.info('result save to {}'.format(os.path.join(save_folder, img_name)))
+
+ if structure_sys.mode == 'structure':
+ save_structure_res(res, save_folder, img_name)
+ draw_img = draw_structure_result(img, res, args.vis_font_path)
+ img_save_path = os.path.join(save_folder, img_name, 'show.jpg')
+ elif structure_sys.mode == 'vqa':
+ draw_img = draw_ser_results(img, res, args.vis_font_path)
+ img_save_path = os.path.join(save_folder, img_name + '.jpg')
+ cv2.imwrite(img_save_path, draw_img)
+ logger.info('result save to {}'.format(img_save_path))
  elapse = time.time() - starttime
  logger.info("Predict time : {:.3f}s".format(elapse))