Learning to Follow Frequency View Guidance for Dental CT Images Deblurring

Jiangang Ding, Yiquan Du, Yihui Shan, Lili Pei, Wei Li

Abstract: Factors such as motion blur and scattering noise continuously degrade dental computed tomography images in unpredictable ways. This makes existing natural image-based restoration methods insufficient to tackle this challenge. This work proposes a novel super-resolution pipeline aimed at comprehensively and holistically reconstructing the true pixel-level details of DCTI, named DCT-Diff. Specifically, we formulate DCTI reconstruction as a denoising diffusion process. Using a two-stage pre-training strategy, we create a compact high-frequency prior for the reconstruction process and achieve high-quality reconstruction through a regression-based approach. During the feature extraction and refinement process, we emphasize separating frequency domain representations from ground truth values. This helps suppress ringing and overshoot artifacts generated during imaging. Our method is simple, robust, and demonstrates State-Of-The-Art (SOTA) visual quality and quantitative metrics compared to 12 frameworks in benchmark evaluations.