Is it amazing to extract text from pictures? Try three steps to realize OCR!

 

About OCR

OCR（Optical Character Recognition, Optical character recognition ） In short, it is a kind of image 、 A system that scans documents or converts text in natural scenes captured by cameras into digital machine encoded text , Through extraction and transformation, it is more convenient to store and search these text information digitally , Reduce input time , And reduce manual search 、 The pain of checking . The bank account number mentioned at the beginning of this article 、 Automatic extraction and input of express address information , Is its typical application .

With the rapid development of deep learning technology , The deep learning technology based on neural network is used to realize OCR It has stronger robustness 、 More accurate 、 Easy to use and other features .

So next we will focus on how to use deep learning technology to achieve OCR. I hope that through today's content introduction , You can roughly understand what is OCR, And how to use your personal computer or notebook CPU, Fast implementation OCR The reasoning of , Realize handwritten digit recognition based on deep learning .

Use deep learning technology to achieve OCR

In this course , We will provide a deep learning model to realize OCR Simple Demo, And use Intel open source tool suite OpenVINO To optimize and accelerate the performance of this model . Just use the source code provided in this course , And learn the next three simple steps , You can use one very conveniently Jupyter Notebook Page based implementation MNIST Handwritten numeral recognition of such open source handwritten numeral data set .

stay Demo in , Considering that we need to judge each number of pictures and make sure that the handwritten digits in each picture are numbers 0~9 this 10 What kind of class , So here OCR Demo We need to build and train a neural network model that can realize image classification , This model can be used for 0~9 this 10 The probability that each of the categories returns a category , Then the number represented by the category with the greatest probability will be determined as the handwritten number finally recognized in our picture .

Three steps of course operation

Next, let's take a look at the specific code .

Step one ： Building environment . We need to install OpenVINO Development kit and corresponding Python tool kit .

Step two ： Build and train the neural network model . Here we only need a few lines of code like the above figure , Can build a simple neural network model .

Besides , We also need to define the output of the final model , We can add a layer as shown in the figure above softmax Layer to get the probability of each category , Selecting the number represented by the category with the greatest probability will be the final result of image recognition and model training . Next , You can see that the whole model training is already running , The running speed should not be underestimated .

Step three ： utilize OpenVINO Provided model optimizer ,Model Optimizer (mo) Optimize the whole neural network model . The whole optimization process runs , Here's the picture .

stay mo After the run is over , We will get the model file saved in intermediate format , Namely xml Document and bin file , These two models save the model structure of the file and the weight of the model .

Next , We can use it OpenVINO To reason , The reasoning code is also quite simple and convenient , Whole OCR Implementation of the complete code , You can refer to here （https://www.kaggle.com/code/raymondlo84/mnist-with-openvino-and-tensorflow-on-kaggle） To download .

Last , Let's take a look at the effect of the whole model ！ Let's print some tests MNIST Handwritten digit recognition results on open source datasets .

You can see that the recognition effect is quite amazing , The accuracy can even reach 99% above , What are you waiting for ？ Come on, according to the source code we provide , Facing the course video and Nono Try it together ！

Full code download address ：https://www.kaggle.com/code/raymondlo84/mnist-with-openvino-and-tensorflow-on-kaggle