Myancanda OCR method based on knowledge distillation

Abstract

The invention relates to a Myanmao OCR method based on knowledge distillation. The method comprises the following steps: constructing a Myanmao image data set adapting to task requirements; adding noise to a generated image, and carrying out regularity labeling on a label code corresponding to the image; constructing a model architecture of a student network and a teacher network using a deep convolutional neural network and a recurrent neural network framework; based on a knowledge distillation method, training a teacher network and a student network in an ensemble learning mode, and by coupling of a teacher integrated sub-network and the student network, aligning local character image features corresponding to a single receptive field in the student network and overall character image features in the teacher network. Therefore, the acquisition of the local features in the long-sequence character image is enhanced, and the student network can efficiently identify the Myanhuang language complex scene text image by calling the deployment model on the server. The recognition of the Myanmai image text is realized, the recognition accuracy is high, and the effect is good.

Description

technical field [0001] The invention relates to a Burmese OCR method based on knowledge distillation, belonging to the technical field of natural language processing. Background technique [0002] Burmese text has Zawgyi-One, Myanmar Three and other font encodings. In order to avoid the problem of garbled Chinese and Burmese text content on the Internet, most Burmese text content is presented in the form of pictures. This has brought great difficulties to the research on Burmese-oriented natural language processing, machine translation, and information retrieval. Although the method combined with deep learning has achieved very impressive results in Chinese and English image text recognition tasks, due to the particularity of Burmese characters, as far as I know, there is no relevant research on Burmese OCR research. Burmese OCR research has very important theoretical and practical application value. Burmese language is different from Chinese or English. In a receptive fie...

AI Technical Summary

Problems solved by technology

[0003] The invention provides a Burmese OCR method based on knowledge distillation, which solves the problem that complex characters in a Burmese image that are nested and combined with multiple characters in a receptive field are difficult to extract and recognize; the invention constructs a teacher network using a CNN+RNN framework and The model architecture in which the student network is trained in an integrated learning manner. During the training process, the sub-network integrated by the teacher is coupled with the student network to realize the local character image features corresponding to a single receptive field in the student network and the overall character image in the teacher network. Alignment of features to enhance the acquisition of local features in long sequence character images

Images