C test calls the paddlesharp module to recognize pictures and words

  PaddleSharp yes PaddleInference C API Of C# Packaged version , Support in Windows(x64)、NVIDIA GPU and Linux(Ubuntu-20.04 x64) Use in .PaddleShare It mainly includes PaddleOCR and PaddleDetection, among PaddleOCR Support 14 Words in different languages , Support the recognition of text with multiple rotation angles ,PaddleDetection Support PPYolo Check the model and PicoDet Model （ About PaddleSharp See references for a detailed introduction of 1）.
   Can pass NuGet Manager Download PaddleSharp Related to the package , At first, I thought the package name was based on PaddleSharp start , It turns out that NuGet I can't find anything in the manager , stay GitHub Of PaddleSharp Home page , Lists PaddleSharp Related package list （ As shown in the figure below ）, All of them Sdcb start , These bags are in NuGet You can search and install in the manager .

   This paper mainly tests C# call PaddleSharp Of PaddleOCR Carry out picture and character recognition , reference 2 Of PaddleOCR The instructions for use list the packages to be installed , stay VS Install the following packages in （ At first, I didn't want to install so many , Just install Sdcb.PaddleOCR, Although the program can be compiled through , It can also run , However, in the process of recognizing the picture and text, it will always be prompted that some of the items in the drop-down package list are missing , Finally, you need to recognize pictures and words normally , You still need to pack all the following bags ）.

Sdcb.PaddleInference
Sdcb.PaddleInference.runtime.win64.mkl
Sdcb.PaddleOCR
Sdcb.PaddleOCR.KnownModels
OpenCvSharp4
OpenCvSharp4.runtime.win

   Refer to references 2 Function test the sample code in （ Please check the detailed code in the literature , I won't post it here ）, The code sequence is mainly to load the model -> Loading pictures -> Identifying text -> Show results . Here we mainly talk about and use PaddleOCRSharp The difference between , Because the tests are simple , What I said is relatively simple .
  PaddleOCRSharp and PaddleSharp Are written in other languages ocr Functional C# Packaged version , by comparison PaddleOCRSharp The encapsulation is simpler . adopt NuGet install PaddleOCRSharp, The dependent files will be automatically downloaded and associated to the project , and PaddleSharp You need to manually install all required packages .
  PaddleOCRSharp Execute the required model, download it synchronously when installing the package and associate it with the project , and PaddleSharp, From the test code , It defaults from C:\Users\HP\AppData\Roaming\paddleocr-models\ppocr-v2 Read the required model in the folder , If it doesn't exist , Downloaded from the Internet in real time ,OCRModel The download path of various models is recorded in the object （ In the process of actual implementation , Due to the average download speed , Directly use Xunlei to manually download and copy to the corresponding folder ）.

OCRModel model = KnownOCRModel.PPOcrV2;
await model.EnsureAll();

   And when recognizing pictures and words ,PaddleOCRSharp It encapsulates , Just specify the image path in the program 、 Picture byte array or Image Object can , and PaddleSharp Functions that recognize pictures ,PaddleOcrAll.Run, You need to convert the picture to the specified Mat object .
  PaddleSharp The results of recognition include two types ,PaddleOcrResult.Text Combine all the recognized characters into one string ,PaddleOcrResult.Regions Save each recognized text block in .
   The text block type is PaddleOcrResultRegion, among Rect（ The type is RotatedRect） Save rectangle position information containing text （ The rectangle may be non horizontal or vertical ）.RotatedRect The definition of class is shown in the figure below , among BoundingRect Function returns a rectangle containing a rotation rectangle . Take the picture below as an example PaddleOcrResultRegion.Rect Is a red rectangle , and BoundingRect（） Returns a purple rectangle . meanwhile RotatedRect Class Points Function returns the position information of the four corners of the rotating rectangle , You can draw a rotating rectangle containing text by drawing lines , It's simpler .


   Finally, the recognition effect of the test image is listed （ Use the sample code in the references , The text blocks are basically circled out , But some of the contents are not correctly identified ）.

reference
[1]https://github.com/sdcb/PaddleSharp
[2]https://github.com/sdcb/PaddleSharp/blob/master/docs/ocr.md

Test picture source
[1]https://baijiahao.baidu.com/s?id=1676249081188736513&wfr=spider&for=pc
[2]https://www.xjauto.net/html/cheping/202105/69508.html
[3]https://item.jd.com/10041422363991.html?openbpab=dontwritecookie&uabt=92_16_21&cu=true&utm_source=image.baidu.com&utm_medium=tuiguang&utm_campaign=t_1003608409_&utm_term=1a1dd3a2f0b24eaab2325b16b12d3d8d
[4]http://news.sohu.com/a/524033869_121124363