Year: 2025
Type: conference-paper
Abstract: Bamboo and silk script characters are challenging to recognize because they look different in many ways. Since they were mostly written by hand, the lines have different thicknesses, sometimes connect, and often have ink smudges. Also, the writing style changed over time and between different writers, leading to many variations of the same character. Over time, the bamboo and silk get squashed and damaged, which can twist and break the characters, making them hard to read. All these things make it very difficult for computers to recognize these old characters. To improve how well computer networks like MobileViT can recognize these scripts, this paper suggests a better network called E-MobileViT. This study aims to make it better at finding and combining important features by using a few new ideas. First, add a better attention system (improved CBAM, I-CBAM) in a specific part of the network (MV2 module). This system automatically decides which parts of the character are most important to look at. Second, put a Multi-scale feature aggregation module (MFA) inside the main building blocks of the network (MobileViT blocks). This helps it see details at different scales, so it doesn't miss small features or get confused by the overall shape. Third, design a way for the network to remember simple features from the beginning and combine them with more complex features later on. This helps it reuse information and learn more effectively. Tests show that this new method greatly improves how accurately the computer can recognize the writing on bamboo and silk.
Authors: Kaikai Liu, Jingqi Song, Zheming Zhao, Helin Liu, Zhijie Qu, Shuin Wang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4413985267