Model Construction for the Application of Machine Vision in Intangible Cultural Heritage Image Recognition

View at publisher Summarise with VIRAT
Authors: Linlan Chen
Cited by: 0
References: 2

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Intangible cultural heritage images present challenges such as complex fine-grained features (e.g., paper-cutting patterns, shadow puppet color gradations), significant intra-class differences (works by different inheritors of the same intangible cultural heritage project), and scarce labeled data. Traditional machine vision algorithms often suffer from inaccurate feature capture and weak generalization ability. This paper proposes a fine-grained recognition model (Heritage-FGRNet) for intangible cultural heritage images, which improves recognition performance through three core innovations: First, it designs an intangible cultural heritage feature attention module (HFAM) that dynamically weights local image features by combining intangible cultural heritage domain knowledge (e.g., pattern topology, color spectrum) to strengthen the feature expression of key cultural elements; second, it constructs a cross-domain data augmentation strategy (CDAS) to expand labeled samples through line drawing-object image feature mapping and style transfer to alleviate the data scarcity problem; and third, it adopts a contrastive-cross-entropy hybrid loss function (CCEL) to simultaneously optimize inter-class feature separation and intra-class feature concentration. The experiment was conducted using a self-constructed "Intangible Cultural Heritage Image Dataset (HCID)," which contains 8623 images (512×512 resolution) across four categories: paper-cutting, shadow puppetry, New Year paintings, and wood carving. The dataset was tested on an NVIDIA RTX 4090 hardware environment using the PyTorch 2.0 framework. Results show that Heritage-FGRNet achieves an average recognition accuracy of 96.87% on the HCID dataset, representing improvements of 7.55 percentage points and 3.24 percentage points compared to traditional ResNet-50 (89.32%) and ViT-Base (93.63%), respectively. Even in small sample scenarios (≤50 images per class), the model accuracy remains at 92.15%, significantly outperforming the comparison algorithms (lowest 82.47%). Visual verification using attention heatmaps confirms that the model achieves a 94.3% accuracy in capturing key features of intangible cultural heritage (such as paper-cutting patterns and shadow puppet joint structures). This research provides an efficient machine vision solution for the automated identification and digital preservation of intangible cultural heritage (ICH) images.

Fields of Study & Subjects

Advanced Neural Network Applications Advanced Image and Video Retrieval Techniques Image Processing and 3D Reconstruction

Publication details

Journal / source
IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
Pages
1438–1443
Publication date
28 January 2026
Language
EN