Application of Deep Learning-Based Image Processing Technology in Historical Architectural Symbol Recognition
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
Historical architectural symbols are a core component of architectural cultural heritage, but their recognition faces challenges such as scarcity of small sample data, local incompleteness, and difficulty fusing local and global features. Traditional SIFT and SVM methods, as well as existing deep learning models (CNNs are strong locally but weak globally, while Transformers are strong globally but weak locally and sensitive to small samples), struggle to effectively address these challenges. To address this, this paper proposes a CNN-Transformer Few Shot recognition algorithm with an attention mechanism (CNN-TransAtt-Few Shot): This algorithm uses an improved CNN with dilated convolutions and adaptive residual blocks to extract local texture from symbols. It then dynamically fuses global semantic features from the Transformer through a cross-attention module. This algorithm, combined with a prototype network with noise suppression weights, achieves small-shot classification, thus constructing an end-to-end recognition framework. Experiments were conducted using the HBOD public dataset (5,400 images of 30 categories of Western architectural symbols) and a proprietary dataset of Ming and Qing architectural symbols (20 categories, 2,800 images, including 15% incomplete samples), comparing algorithms such as ResNet50 and ViT-B/16. Results showed that the 10-way 1-shot task accuracy reached 92.3%, an improvement of 8.7 and 4.2 percentage points over ResNet50 and ViT-B/16, respectively. The 5-shot task accuracy reached 95.8%, with an F1-Score of 95.7%. Even with a 30% incompleteness rate, the accuracy remained at 89.1%, a 7.6 percentage point improvement over ViT-B/16. The algorithm achieved a single-image inference time of 28ms and a parameter count of 42M, balancing accuracy and deployment efficiency, providing technical support for the digital preservation of architectural culture.
Fields of Study & Subjects
Advanced Neural Network Applications
Domain Adaptation and Few-Shot Learning
Advanced Image and Video Retrieval Techniques
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 1449–1454
- Publication date
- 28 January 2026
- Language
- EN