OCR实践-Table-Transformer

前言

书接上文

OCR实践---PaddleOCR

Table-Transformer 与 PubTables-1M

table-transformer，来自微软，基于Detr，在PubTables1M 数据集上进行训练，模型是在提出数据集同时的工作，

paper PubTables-1M: Towards comprehensive table extraction from unstructured documents，发表在2022年的 CVPR

数据来自 PubMed PMCOA 数据库的一百万个文章表格

PubTables-1M 针对表格处理一共有三个任务（所以table transformer 也能做到）

表格检测（表格定位）TD
表格结构识别（行、列、spanning cell，grid cell， text cell）TSR
表格分析（表头 cell，projected row header cell) FA

table-transformer

是第一个将 detr 用于表格处理任务的模型，没有使用任何特别的定制模块，简称为 TATR

we apply the Detection Transformer (DETR) $2$ for the first time to the tasks of TD, TSR, and FA, and demonstrate how with PubTables-1M all three tasks can be addressed with a transformer-based object detection framework without any special customization for these tasks.

有关模型详细的权重、指标信息可以通过论文和 Github仓库可以进一步了解

https://arxiv.org/abs/2110.00061

https://github.com/microsoft/table-transformer

官方也在HuggingFace 上提供了各个模型权重

https://huggingface.co/collections/microsoft/table-transformer-6564528e330b667bb267502e

各个模型的版本和区别信息如下

官方提示，microsoft/table-transformer-structure-recognition-v1.1-all 是最好的结构识别模型

实践代码

如有问题，需要帮助，欢迎留言、私信或加群交流【群号：392784757】

表格检测 TD

通过以下设置，可以加速下载以及保存模型到当前文件夹下

python 复制代码

import os
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
os.environ['HF_HUB_CACHE'] = './hf_models/'
os.environ['TRANSFORMERS_CACHE'] = './hf_models'
os.environ['HF_HOME'] = './hf_models'

打开文件

python 复制代码

table_img_path = './table.jpg'
image = Image.open(table_img_path).convert("RGB")
file_name = table_img_path.split('/')[-1].split('.')[0]

加载模型

python 复制代码

image_processor = AutoImageProcessor.from_pretrained("microsoft/table-transformer-detection")
model = TableTransformerForObjectDetection.from_pretrained("microsoft/table-transformer-detection")

模型推理与后处理

python 复制代码

inputs = image_processor(images=image, return_tensors="pt")
outputs = model(**inputs)

target_sizes = torch.tensor([image.size[::-1]])
results = image_processor.post_process_object_detection(outputs, threshold=0.9, target_sizes=target_sizes)[0]

结果解析

python 复制代码

i = 0
for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
    box = [round(i, 2) for i in box.tolist()]
    print(
        f"Detected {model.config.id2label[label.item()]} with confidence "
        f"{round(score.item(), 3)} at location {box}"
    )
 
    region = image.crop(box) #检测
    region.save(f'./{file_name}_{i}.jpg')
    i += 1

表格结构识别 TSR

打开图片与模型加载

python 复制代码

from transformers import DetrFeatureExtractor
feature_extractor = DetrFeatureExtractor()


file_path = "./locate_table.jpg"
image = Image.open(file_path).convert("RGB")


encoding = feature_extractor(image, return_tensors="pt")
model = TableTransformerForObjectDetection.from_pretrained("microsoft/table-transformer-structure-recognition-v1.1-all")
print(model.config.id2label)
# {0: 'table', 1: 'table column', 2: 'table row', 3: 'table column header', 4: 'table projected row header', 5: 'table spanning cell'}

模型推理与后处理

python 复制代码

with torch.no_grad():
    outputs = model(**encoding)
    
target_sizes = [image.size[::-1]]
results = feature_extractor.post_process_object_detection(outputs, threshold=0.6, target_sizes=target_sizes)[0]
# print(results)

结果解析

header

python 复制代码

headers_box_list = [results['boxes'][i].tolist() for i in range(len(results['boxes'])) if results['labels'][i].item()==3] 
crop_image = image.crop(headers_box_list[0]) 
crop_image.save('header.png')

column

python 复制代码

columns_box_list = [results['boxes'][i].tolist() for i in range(len(results['boxes'])) if results['labels'][i].item()==1] 
print(len(columns_box_list))

row

python 复制代码

rows_box_list = [results['boxes'][i].tolist() for i in range(len(results['boxes'])) if results['labels'][i].item()==2] 
print(len(rows_box_list))

cell

python 复制代码

cell_draw_image = image.copy()
cell_draw = ImageDraw.Draw(cell_draw_image)

# col row inserction
for col in columns_box_list:
  for row in rows_box_list:
    cell = intersection(col,row) # 自行定义 
    if cell is not None:
       cell_draw.rectangle(cell, outline="red", width=3)

cell_draw_image.save("cells.png")

效果

效果还不错

感谢

感谢以下文章提供的灵感与代码参考

$表格检测与识别入门 - My Github Blog$ (https://percent4.github.io/表格检测与识别入门/#表格结构识别
表格检测与识别的初次尝试