[論文レビュー] A Comparison of Image Denoising Methods

Zhaoming Kong, Fangxi Deng+4 more|arXiv (Cornell University)|Apr 18, 2023

Image and Signal Denoising Methods被引用数 8

ひとこと要約

この論文は、伝統的な手法と深層ニューラルネットを含む幅広い画像ノイズ除去法を合成データと実データで比較し、新しいベンチマークデータセットを導入します。

ABSTRACT

The advancement of imaging devices and countless images generated everyday pose an increasingly high demand on image denoising, which still remains a challenging task in terms of both effectiveness and efficiency. To improve denoising quality, numerous denoising techniques and approaches have been proposed in the past decades, including different transforms, regularization terms, algebraic representations and especially advanced deep neural network (DNN) architectures. Despite their sophistication, many methods may fail to achieve desirable results for simultaneous noise removal and fine detail preservation. In this paper, to investigate the applicability of existing denoising techniques, we compare a variety of denoising methods on both synthetic and real-world datasets for different applications. We also introduce a new dataset for benchmarking, and the evaluations are performed from four different perspectives including quantitative metrics, visual effects, human ratings and computational cost. Our experiments demonstrate: (i) the effectiveness and efficiency of representative traditional denoisers for various denoising tasks, (ii) a simple matrix-based algorithm may be able to produce similar results compared with its tensor counterparts, and (iii) the notable achievements of DNN models, which exhibit impressive generalization ability and show state-of-the-art performance on various datasets. In spite of the progress in recent years, we discuss shortcomings and possible extensions of existing techniques. Datasets, code and results are made publicly available and will be continuously updated at https://github.com/ZhaomingKong/Denoising-Comparison.

研究の動機と目的

• Assess the effectiveness of traditional denoisers vs. DNN-based methods across diverse datasets and applications.
• Introduce real-world benchmarking datasets for image, video, MSI/HSI, and MRI denoising tasks.
• Evaluate methods using quantitative metrics, visual effects, human ratings, and computational cost.
• Provide observations on generalizability and practical applicability of denoising techniques.

提案手法

• Review and categorize denoisers into traditional and DNN-based approaches.
• Use grouping-collaborative filtering-aggregation framework to describe traditional denoisers.
• Discuss NLSS priors and various algebraic representations (matrix, tensor, SVD-based, low-rank, etc.).
• Examine DNN training strategies (supervised, self-supervised, unsupervised) and their impact on performance.
• Introduce a new real-world dataset and benchmarks for multiple denoising tasks.
• Provide comparisons across synthetic and real-world experiments with multiple data modalities.

実験結果

リサーチクエスチョン

• RQ1 How do traditional denoisers compare with DNN-based denoisers across image, video, MSI/HSI, and MRI denoising tasks?
• RQ2 What are the roles of NLSS priors, transforms, and tensor representations in denoising performance?
• RQ3 How well do pretrained or dataset-specific DNN models generalize to new datasets and applications?

主な発見

• DNN methods show impressive improvements on several real-world datasets but pretrained models may not generalize well across datasets.
• Traditional denoisers like the BM3D family remain competitive in multiple denoising tasks.
• Modified SVD (M-SVD) can achieve competitive results relative to tensor-based methods.
• Certain DNN models such as FCCF, RVRT, and FastDVDNet achieve strong results in real-world image and video denoising tasks.
• Video denoising methods like FastDVDNet, FloRNN, and RVRT demonstrate outstanding performance in terms of effectiveness and efficiency.