02546nas a2200277   4500000000100000008004100001653001700042653001900059653001800078653001500096653002600111653002300137653003400160653001800194653002000212653002300232653003500255653001800290653002100308653002300329100001200352245011200364856015800476300001400634520162000648                                       d10aBenchmarking10aCultural value10aDeep learning10aEmbroidery10aHistoric preservation10aHuman civilization10aIntangible cultural heritages10aModel sharing10aMulti-modelling10aPersonnel training10aQuantitative evaluation models10aSocial values10aTextile printing10aTraditional crafts1 aRui Mao00aEmbrNet: Quantitative Evaluation Model for the Inheritance of Embroidery as an Intangible Cultural Heritage  uhttps://www.scopus.com/inward/record.uri?eid=2-s2.0-105010222876&doi=10.1109%2FAINIT65432.2025.11035726&partnerID=40&md5=773c50c55b66d041e44652fb5cebf9eb  a1761-17653 aIntangible Cultural Heritage (ICH), as an essential component of human civilization, carries rich historical, cultural, and social value. Embroidery, as a traditional craft, not only showcases exquisite skills but also embodies profound cultural significance. However, traditional ICH evaluation systems tend to focus on qualitative analysis, lacking systematic quantitative evaluation methods. This limitation restricts the ability to objectively and fairly assess embroidery techniques and their products. To address this issue, we have developed a quantitative evaluation model called EmbrNet, which is based on a multi-model parameter-sharing framework with parameter-level integration and transformation. The model includes three stages: the benchmark training stage, the multi-model sharing stage, and the joint discrimination stage. In the benchmark training stage, we used web scraping technology to collect and label embroidery image datasets for training the base model. In the multi-model sharing stage, features are extracted by leveraging the characteristics of different model architectures, and a parameter converter is used to assign parameters to each model. Finally, in the joint discrimination stage, the parameters of all heterogeneous models are integrated to output a quantitative evaluation result. Experimental results show that EmbrNet significantly outperforms other comparative models in classification tasks, achieving accuracy rates of 91.6\% and 94.3\%, as well as 88.9\% and 90.1\%, proving its powerful ability to handle complex embroidery images and its broad application prospects.