Table image extraction based on traditional intersection method and Tesseract OCR

This paper applies the traditional intersection method to extract the frame table , Mainly for the existence of merge cell Table for , And use tesseract-ocr To do character recognition , The main difficulty is the extraction part of the former , Now I will share some methods with you .

It is suggested that you can first see the extraction methods of other bloggers for simple tables , Look at this one again , Because I omitted a lot of content , Mainly talking about methods , It is not recommended that friends who have never known each other before come to see , Recommend an article ：

python-opencv Form identification _ I want to ask Wentian's blog -CSDN Blog _opencv Form identification

Because I want to eat just right , Earn the expense of graduation trip , Therefore, only part of the content will be shared , Hope to understand , It is also my first time to post an article , If there is something not in place , I would also like to point out that .

---

One 、 Difficulties in table extraction

1. merge cell

I have checked many publications in csdn Blog post on , The traditional method is to use the intersection of frames , Get intersections to extract cells , But most of them are for simple forms , In real office life , We often encounter complex tables with merged cells, as shown in the following figure , Then we need a set of complex methods to extract .

2. The image processing

Image preprocessing is a crucial step , We must reduce all kinds of interference in the image , The purpose is to be able to accurately extract intersections , The whole extraction process , Are carried out according to the intersection coordinates , Image preprocessing can refer to other bloggers , Generally, image binarization is required 、 Graying 、 Denoise 、 Sharpening, etc ......

for instance ： There are some images , The connecting part of the frame line is not connected , The intersection cannot be extracted , Then you can use Opencv Methods in the library , Lengthen the detected frame , So that we can get the intersection , Here's the picture .

This figure has been binarized

The horizontal and vertical lines have been extended
​

Because the image processing part is very important , It will affect the subsequent table structure extraction and character recognition , So you need to constantly debug , To achieve the desired effect , I think this is also the limitation of traditional methods , So the current popular methods are deep learning .

Two 、 Table extraction process

After image preprocessing , That's the main content of this article —— Extract table structure

The whole extraction process is divided into two steps ：

        1. Intersection classification

        2. Extract the table according to the intersection category

---

1. Intersection classification

This is an original picture , There is a case of merging cells , Because this article focuses on methods , Don't talk too much about image processing , So high definition images are used , But we must emphasize the importance of image preprocessing .

（1） After binarization 、 After graying

 （2） Extract horizontal and vertical lines , After corrosion and expansion , So the lines are longer than the original

 （3） Superimpose horizontal and vertical lines , The purpose is to obtain the intersection

 （4） After superposition , Then the intersection is obtained , Marked as black dot

 （5） Classify intersections , First, identify the marker points representing each cell , For example, in the following cell ,p1 Points are marker points , Represents the current cell .

P1 Is the marker

  With the above information , The intersection points in the graph can be classified for the first time , The marking point is red ：

  The current classification is far from enough , Still can't extract table cells well , It needs to be classified again , Keep the red dot , And the color of the bottom and top right points of the table , The remaining points are set to green , Pictured ：

（6） According to the intersection of different color types , Design a set of methods , You can get the area position of each unit .

 （7） The next step is to intercept Each cell is saved in a folder , Record cell position information , Prepare for later table structure restoration and character recognition .

2. Extract the table according to the intersection category

Because this poor student wants to have a meal , Therefore, the core specific methods are not given , If you have interested friends , You can talk in private , I can give the project code , It can also be done through ppt The whole process of the speech , I hope you can scold , To make a long story short , Poverty is the original sin .

3. Actual demonstration

Process interface

After image preprocessing

Complete intersection classification

According to the design algorithm , Get each cell range

The final result of table restoration and character recognition