Cut To Size: A Lightweight Design for STR via Channel Pruning

In natural scenes, text full of rich and prominent information contributes to understand the scene. Scene text recognition (STR) is commercially viable in the wave of big data, which comes to be the center of attention. The existing STR methods based on deep learning have make significant progress, but for the complexity and size of model, it is difficult to be deployed in embedded devices. Therefore, a lightweight design for STR via synergistic pruning is proposed. To score the value of channels, it evaluates not only the convolution layer weights, but also the BN(Batch Normalization) layer scaling factors jointly. It can delete redundant channels which slightly influences the ability of the model. Meanwhile, some strategies are presented for fine-tuning model performance or contracting the model size. All models are trained and tested on public datasets. The experimental results on CCPD2019 well prove the effectiveness of this work, which can cut 87.60% params of CRNN and 78.31% params of Rosetta, respectively, and only reduce a little accuracy.