A license key information extraction method based on small sample data

Abstract

The technical scheme of the present application provides a license key information extraction method based on small sample data. The present application uses an OCR model to obtain the text and position in the picture. Then, the text, position, picture and other features are fused and input into an improved BERT model to learn the interaction between the features and output the label of each character. The adjacent characters with the same label are combined to form a field with a label. Then, the field position and field label information are input into a DenseCRF model, and the field labels are corrected according to the relative positions and angles between the fields. Since the type and position of the field in the license are relatively fixed, the DenseCRF can achieve a good correction effect. Through language model pre-training, multi-task training, artificial generation of training samples and correction based on position information, the present application realizes accurate extraction of license key information under the condition of small sample data. Only a few artificially labeled images are needed for each license to realize model training.

Images