Lightweight character recognition model design method, system and device and medium

Abstract

The invention discloses a lightweight character recognition model design method, system and device and a medium. The method comprises the following steps: selecting a reference model: adopting a text line recognition model based on a convolutional recurrent neural network as the reference model; network structure search: searching a backbone network suitable for a character recognition task by adopting a ProxylessNAS network structure search algorithm, and using a LayeNorm layer as a normalization layer of a feature sequence; knowledge distillation: improving the performance of the lightweight model by adopting a feature-based knowledge distillation method, assigning a regression device weight in the distillation method based on SVD decomposition, and performing dimensionality reduction on features extracted by the teacher model; and distillation auxiliary network structure searching: adding a distillation learning auxiliary searching process in a network searching process. According to the method, the knowledge distillation and the network structure search model are organically combined, the problem that the calculation amount and the storage amount of an existing method are too large is solved, the character recognition model can be deployed on mobile terminal equipment, and the method can be widely applied to the technical field of artificial intelligence.

Description

Technical field [0001] The present invention relates to the field of pattern recognition and artificial intelligence technology, in particular to a lightweight text recognition model design method, system, apparatus and medium. Background [0002] Words provide an important information resource for human understanding of the outside world, and in recent years, artificial intelligence technology has continued to develop, and how to make machines learn to read words and understand words has become a hot topic of concern in academia and industry. As an important step in the process of text from image to digitization, text recognition has an important impact on the performance of the entire system. [0003] At present, with the development of big data and the improvement of hardware technologies such as GPUs, scene text detection algorithms based on deep learning are emerging in an endless stream, which has a certain degree of performance improvement compared with traditional image p...

AI Technical Summary

Problems solved by technology

However, methods based on deep learning usually require a large amount of computing resources and storage space, which limits the application of text recognition algorithms to a certain extent, especially on various mobile terminals and edge devices.

Images