CHANG Yangyang, HU Jingfang, CHENG Xien, NIE Yu

(School of Information Engineering, Jingdezhen Ceramic University, Jingdezhen 333403, Jiangxi, China)

Extended abstract:

[Background and purposes] Weakly textured ceramic micrographs pose a major challenge for computer-vision-based identification, as conventional SIFT matching often suffers from poor precision and instability under scale, rotation and illumination variations. To address these issues, this study is aimed to propose a high-precision identification method for weakly textured ceramic micrographs based on an improved SIFT framework. The method integrates multi-scale edge analysis, adaptive regional statistics and hierarchical geometric verification to achieve robust and accurate feature matching. The purpose of this work is to develop a reliable, non-destructive and generalizable approach for ceramic authentication at microscopic scale, providing technical support for digital heritage protection and artwork appraisal.

[Methods] High-resolution SEM images of the ceramics were captured using a "ChaoyanB011" digital microscope (2048×1536 resolution). A micro-ceramic dataset containing 10,000 images from 20 ceramic specimens was constructed. The algorithm begins with SIFT keypoint detection and enhances it through multi-scale edge analysis to suppress redundant boundaries, while preserving discriminative edges. Subsequently, adaptive regional statistics are employed to evaluate feature quality based on local mean, variance and entropy, enabling dynamic thresholding for stable keypoint selection. Finally, a multi-level geometric verification scheme, combining local affine consistency and global RANSAC estimation, ensures high matching reliability. The proposed method was compared with SIFT+BF, SIFT+NNDR, SIFT+RANSAC, ORB+NNDR and AdaLAM, under five experimental scenarios, including true-to-true, fake-to-true, scale variation, rotation and illumination change.

[Results] It is demonstrated that both the precision and robustness are substaintially improved. In true-to-true experiments, the method consistently achieved precision values above 0.9916 with RMSE below 0.0468. In fake-to-true tests, no false matches occurred. Compared with classical SIFT and its variants, the proposed method achieved an average precision improvement of about 48.89%. Under extreme conditions, such as 90° rotation and 4× scaling, the precision reached 1.0, while RMSE dropped to 0.0036. When tested under seven illumination levels, the RMSE remained nearly constant (≈0.001), with negligible standard errors, confirming strong robustness to brightness fluctuations. Although the computation time per image pair (≈3.3 s at 2048×1536) is slightly higher than that of the traditional methods, the improvement in reliability and precision makes it suitable for offline authentication and future real-time optimization.

[Conclusions] A high-precision and robust identification method is presented, which is tailored for weakly textured ceramic micrographs. By combining multi-scale edge extraction, adaptive regional evaluation and multi-layer geometric verification, the method effectively overcomes the limitations of traditional SIFT in low-texture environments. The results validate its superior accuracy, stability and generalization across different imaging conditions. Beyond technical advances, this method provides an important auxiliary tool for the scientific authentication of ceramic artworks, supporting both academic research and practical appraisal. With further optimization, such as GPU acceleration and coarse-to-fine keypoint matching, the approach holds strong potential for real-time or large-scale ceramic identification systems.

Key words: microscopic image; ceramic identification; SIFT; image matching; robustness