计算机技术与发展2026，Vol.36Issue(9)：45-53,9.DOI:10.20165/j.cnki.ISSN1673-629X.2026.0094
基于MSR-OverLoCK的唐卡图像识别
Tangka Image Recognition Based on MSR-OverLoCK
摘要
Abstract
As an important carrier of Tibetan traditional painting art,Thangka exhibits visual characteristics such as complex composition,rich details,and slight inter-class differences,posing significant challenges to automatic recognition.To address this issue,we propose an improved OverLoCK network model for refined recognition of Thangka images.Based on the global-local collaborative framework of OverLoCK,the model incorporates three key improvements.Firstly,a Multi-Scale Gated Attention(MSGA)module is introduced into the Base-Net,which captures multi-scale feature information through a multi-branch parallel convolution structure and achieves adaptive feature fusion combined with the SE attention mechanism,enhancing the network's ability to extract multi-scale texture details.Second-ly,the ConvFFN in the Base-Net is replaced with the MK-ConvGLU module,improving the model's perception of local details.Finally,an Aggregated Attention Feature Enhancement(AAFE)module is integrated into the Focus-Net to realize accurate localization and fea-ture enhancement of key regions.Comparative experiments on the Thangka dataset show that compared with the baseline OverLoCK model,MSR-OverLoCK achieves an accuracy increase from 93.23%to 96.85%and an F1 improvement from 92.40%to 96.11%,out-performing mainstream models such as ResNet-50,ConvNeXt-T,and Swin Transformer significantly.These results verify the effective-ness of the proposed improvement strategies,providing a feasible technical solution for the intelligent recognition of Thangka images and the digital protection of cultural heritage.关键词
唐卡图像识别/MSR-OverLoCK/多尺度特征融合/注意力机制/文化遗产数字化保护Key words
Thangka image recognition/MSR-OverLoCK/multi-scale feature fusion/attention mechanism/digital preservation of cultur-al heritage分类
信息技术与安全科学引用本文复制引用
王雪晴,史伟,白科..基于MSR-OverLoCK的唐卡图像识别[J].计算机技术与发展,2026,36(9):45-53,9.基金项目
国家自然科学基金项目(62166030) （62166030）
