Scholarlink

<p>Ancient Chinese murals possess profound historical and artistic value. Their digital preservation relies on deep learning-based automated object detection, which serves as a crucial step for subsequent restoration and archaeological analysis. However, mural images frequently suffer from severe environmental degradation, including weathering, blurring, and peeling. While existing detection frameworks perform well on clear or pre-restored data, their performance drops significantly when processing low-quality raw murals. Due to the lack of mechanisms to handle complex and irregular degradations, traditional spatial convolutions struggle to separate blurred targets from noisy backgrounds. To this end, we propose SDEM-Net, a novel spectral and deformation-enhanced object detection framework tailored specifically for severely degraded murals. Built upon the YOLO architecture, this framework introduces the Frequency Amplitude Modulation combined with C2f (FAMC2f) module, which uses the Fast Fourier Transform to conduct adaptive amplitude modulation in the frequency domain, thereby effectively extracting globally invariant features and suppressing noise. Furthermore, we designed a Deformable Asymmetric Band Convolution (DAB-Conv) module in the backbone network to dynamically capture multi-scale spatial information and non-rigid object deformations via a multi-branch structure. Extensive evaluations demonstrate that the proposed method outperforms the compared baselines in both detection accuracy and robustness. Specifically, on the custom MURAL dataset, SDEM-Net achieves a 2.2% improvement in mAP50-95 over YOLOv12-L, and a 3.1% improvement in mAP50 over the strongest competing baseline DKR-YOLO. Furthermore, on the public RITK benchmark dataset, SDEM-Net outperforms DKR-YOLO by 1.3% and 2.8% in mAP50-95 and mAP50, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SDEM-Net:spectral and deformable enhanced framework for mural object detection

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Yuge Ouyang,
  • Liujue Zhang

摘要

Ancient Chinese murals possess profound historical and artistic value. Their digital preservation relies on deep learning-based automated object detection, which serves as a crucial step for subsequent restoration and archaeological analysis. However, mural images frequently suffer from severe environmental degradation, including weathering, blurring, and peeling. While existing detection frameworks perform well on clear or pre-restored data, their performance drops significantly when processing low-quality raw murals. Due to the lack of mechanisms to handle complex and irregular degradations, traditional spatial convolutions struggle to separate blurred targets from noisy backgrounds. To this end, we propose SDEM-Net, a novel spectral and deformation-enhanced object detection framework tailored specifically for severely degraded murals. Built upon the YOLO architecture, this framework introduces the Frequency Amplitude Modulation combined with C2f (FAMC2f) module, which uses the Fast Fourier Transform to conduct adaptive amplitude modulation in the frequency domain, thereby effectively extracting globally invariant features and suppressing noise. Furthermore, we designed a Deformable Asymmetric Band Convolution (DAB-Conv) module in the backbone network to dynamically capture multi-scale spatial information and non-rigid object deformations via a multi-branch structure. Extensive evaluations demonstrate that the proposed method outperforms the compared baselines in both detection accuracy and robustness. Specifically, on the custom MURAL dataset, SDEM-Net achieves a 2.2% improvement in mAP50-95 over YOLOv12-L, and a 3.1% improvement in mAP50 over the strongest competing baseline DKR-YOLO. Furthermore, on the public RITK benchmark dataset, SDEM-Net outperforms DKR-YOLO by 1.3% and 2.8% in mAP50-95 and mAP50, respectively.