A Lightweight Backbone Used for Scene Text Recognition
Tianxiang Lan, Dong Yin · 2022 3rd International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE) · 2022
This paper presents AutoMLPMixer which is a lightweight backbone based on the mixer-multi-layer percep-trons(MLPMixer) and automated machine learning. Secondly, in order to deal with parameter surging caused by a large dictionary in Chinese scene text recognition, this paper introduces grouped linear structure to process the embedding layers and prediction layers of the model. Compared with ResNet, the AutoMLPMixer can effectively reduce parameters by 16.4%. By introducing a grouped linear structure to solve the oversized dictionary problem, we can further reduce parameters by 14.89M in scene text recognition. Under coexist condition of both AutoMLPMixer and grouped linear structure, the model image processing speed increased by 29%. The code and models will be made publicly.