Lightweight Text Page Spotter for Chinese Handwriting Document
Qiyan Zhao, Jiuze Li, Lanying Liang, Yuefeng Liu, Zheng Li, Hongyu Shen · 2024
Text page spotting is crucial for the digitization of documents. However, lightweight spotters and available datasets for Chinese handwriting documents remain limited. This paper proposes a lightweight text page spotter, utilizing lightweight ShuffleNetV2 as the backbone. The feature representation capabilities of the backbone are enhanced through a feature enhancement mechanism incorporating both multi-scale and multi-dimensional enhancements. Additionally, a text correction module comprising Mask Correction and thin-plate-spline (TPS) transformation is introduced to mitigate non-target text pixel interference and differentiate adj acent text in complex backgrounds. Finally, An iterative magnitude pruning algorithm is employed to further compress the model. The experiment results on self-built Chinese handwriting documents are demonstrated, achieving an accurate rate (AR) of 88.91 %, a correct rate (CR) of 89.36%, and a text detection F1 score of 95.15. After model compression, the number of parameters is reduced by 86.58%, with AR and CR decreasing by only 2.83% and 2.18%, respectively.