Enhanced UNet Architecture for Image Segmentation: Improving Efficiency in Understanding the Details of Wood Carving Patterns

Guangtao Zhang¹Wanlin Zhang^1*Huanshu Jiang¹Yuxiang Fan¹Xiwen Peng²

• ¹Southern University of Science and Technology, Shenzhen, China
• ²South China Agricultural University, Guangzhou, Guangdong Province, China

This paper presents ongoing research on digital approaches for preserving and transmitting traditional wooden carvings. By collaborating with a nationally designated Intangible Cultural Heritage (ICH) Inheritor, a master wood carver specializing in lantern making, we explore how to effectively convey the cultural meanings and historical significance of thousands of traditional wood carving patterns that have been digitized. To do so, we establish a fine-grained visualization and information system using these high-resolution digital records. However, the vast number of wooden artifacts with diverse patterns poses a challenge in segmenting and classifying the different parts of the wood carving patterns. To address this challenge, we use an AI-based image segmentation approach, which is particularly effective for understanding image details, especially when sub-regions contain meaningful patterns. In our research, we construct an enhanced UNet architecture (EUNet) to partition images into regions of meaningful patterns and help the users understand these details. Importantly, our improved adaptive algorithm directs our framework to focus on the regions that are more challenging to learn. We implement our proposed method using the Python programming environment and build a dataset to evaluate its performance against other approaches, including the baseline UNet, Attention-UNet, Dense-UNet. Through these experiments, we assess the effectiveness of our method and demonstrate that it achieves superior performance compared to traditional methods, with higher accuracy, recall, F-measure, and precision.