1. First detect and extract the text area.

2. Then we use radon hough transform to correct the text.

3. Segmentation of single-line text image with projection histogram.

Finally, a single line of OCR.

There are two main ideas for single-line OCR.

The first is the need to split characters.

There are many ways to divide characters. The most common method is to use the extreme point of the projection histogram as the candidate segmentation point, and use classifier+beam search to search the best segmentation point.

After finding the segmentation point, the traditional method is feature engineering+classifier for a single character. The general process is-> gray scale; Binary->; Corrected image->; Feature extraction (various methods such as pca lbp, etc. )->; Quantifiers (quantifiers are generally SVM· ANN Ann, etc.).

Now CNN (Convolutional Neural Network) can largely avoid feature engineering.

The second is that you don't need to split characters.

Another point is end-to-end identification, but only if you need a large number of labeled data sets. This method can output the character sequence continuously without dividing the image.

Multi-label classification can be used for short lengths. Such as license plate, verification code. Here, I tried the multi-label classification of license plates. End-to-end recognition of undivided characters in license plate recognition.

This is the way Google uses to identify street view house numbers.