Simplified image acquisition model training, image text recognition method and related devices

Abstract

Embodiments of the present invention provide a simplified image acquisition model training, image text recognition method and related devices. The training method includes: obtaining a training data set, including each training scene text image, and each benchmark simplified text image, and the same training scene text image The number of benchmark simplified text images is at least two; use the first encoder to obtain the encoding features of each scene text image, and use the decoder to obtain a predicted simplified text image; randomly select from each benchmark simplified text image corresponding to the training scene text image Obtain the comparison reference text image, simplify the text image and comparison reference text image according to the prediction, obtain the prediction loss, and adjust the parameters of the first encoder and decoder until the prediction loss meets the loss threshold. The simplified image acquisition model training method, image text recognition method and related devices provided by the embodiments of the present invention can ensure the accuracy of image text recognition and reduce the training cost of the model.

Description

technical field [0001] Embodiments of the present invention relate to the field of computers, and in particular to a simplified image acquisition model training method, image text recognition method and related devices. Background technique [0002] Image text recognition refers to the process of identifying character sequences from text images to provide character information for subsequent further processing (such as: topic review, text-to-speech synthesis, etc.). [0003] With the development of technology, image text recognition can be realized through recognition models. However, in order to ensure the recognition effect of the model, it is necessary to use a large amount of data to train the recognition model. Since the text in the image to be recognized may also be handwritten, the training data must not only be large enough, but also satisfy the diversity. In order to realize the judgment of the effect of model training, it is also necessary to manually label the te...

AI Technical Summary

Problems solved by technology

However, in order to ensure the recognition effect of the model, it is necessary to use a large amount of data to train the recognition model. Since the text in the image to be recognized may also be handwritten, the training data must not only be large enough, but also satisfy diversity. In order to realize the judgment of the effect of model training, it is necessary to manually label the text images in the training data, which will consume a lot of manpower and material resources.

Images