The paper explores a deep learning-based approach for denoising panoramic X-ray images. The key highlights are:
The authors propose integrating Daubechies (Db2) and Haar wavelets within a U-Net neural network architecture to enhance the denoising capabilities.
The Db2 wavelet is used in the initial layers to capture fine-grained details, while the Haar wavelet is employed in the later layers to provide a more coarse-grained representation. This multi-resolution analysis enables effective noise removal while preserving essential image features.
Experiments are conducted on standard datasets (Set12, Set14, BSD68) as well as a large proprietary dataset of over 500,000 panoramic X-ray images. The proposed method outperforms previous techniques like U-Net and Haar wavelet-based approaches in terms of PSNR and SSIM metrics.
The authors discuss the importance of careful wavelet selection and integration within the neural network architecture to balance the trade-off between noise reduction and preservation of image details. The flexibility in adjusting the step size for different wavelet types is highlighted as a key factor in the optimization process.
The results demonstrate that the integration of Db2 and Haar wavelets in the U-Net architecture significantly improves the denoising performance, especially for panoramic X-ray images, which are often plagued by noise and artifacts.
Sang ngôn ngữ khác
từ nội dung nguồn
arxiv.org
Thông tin chi tiết chính được chắt lọc từ
by Masoud Shahr... lúc arxiv.org 04-02-2024
https://arxiv.org/pdf/2302.10306.pdfYêu cầu sâu hơn