What algorithm is used for OCR text recognition?

The general OCR routine is like this.

1.Detect and extract the text region first

2.Then, the Radon Hough transform and other methods are used to correct the text.

3.Split out a single line of text by projecting a histogram.

Finally, the OCR for a single line

The OCR of a single line consists of two main ideas.

The first is where you need to split the characters.

There are also many methods to segment characters, and the most commonly used is to use the extreme points of the projected histogram as the candidate segmentation points and use the classifier + beam search to search for the best segmentation points.

After searching for a split point, for a single character, the traditional one is feature engineering + classifier. The general process is: grayscale -> binarization -> correcting the image -> extracting features (various methods, such as PCA, LBP, etc.) - classifiers (classifiers are roughly SVM, ann, knn, etc.).

Today's CNNs (Convolutional Neural Networks) can largely eliminate feature engineering.

The second is that there is no need to split the characters.

Another point is end-to-end recognition, but only if you need a large number of labeled datasets. This method makes it possible to output a sequence of characters directly in a continuous without dividing the image.

For short lengths, mutli-label classification can be used. For example, like license plates, captchas. Here I tried a multi-label classification of license plates.

End-to-end (end-to-end) recognition of undivided characters in license plate recognition.

This is the method used by Google to do Street View house number recognition. <>

The reasons for the low accuracy rate of using OCR text recognition software are as follows:

1. The recognized text is not clear enough;

2. Not using professional OCR text recognition software;

3. **The text should not be scribbled.

Basically, the OCR text recognition accuracy is low due to the above three reasons.

The recognition accuracy of OCR technology is instantly able to recognize text and convert almost all types of printed documents, including articles in books and magazines with complex layouts.

If your company needs to identify a large number of text information, it is recommended to use the OCR text recognition system in the future, you can call the specific link, I hope it can help you

This is related to the clarity of the **, if the clarity is high, abbyy is recommended, but this software is too big, you can use the love dot converter (lovedot.cn), this recognition accuracy is very high, and it can output word and excel

First of all, the image mode of scanning is black and white, the resolution is set to 300dpi, and the color level is set to 128, so the recognition rate is more than 90%, and I often use Shangshu.

Normally, the recognition accuracy of OCR text recognition software will not be low, the key is still the ** or other media that are recognized, and the accuracy of OCR text recognition is high, you can try.

The reason for the font is that bold will reduce his recognition.

The best recognition rate is Songti.