Classic Denoising Algorithm - BM3D

Application Background

The non-local algorithm for image denoising was first proposed at CVPR 2005. Building upon this foundation, the 2007 TIP paper introduced improvements that led to the classic BM3D algorithm, which has become a benchmark in image denoising with excellent performance!

Key Technology

The Non-Local Means (NLM) denoising algorithm reduces noise (zero-mean Gaussian white noise) by estimating the center point of reference blocks through weighted averaging of these self-similar structural blocks. Although NLM achieves good denoising results, its preservation of the original image's structural information remains insufficient. The 2007 TIP paper proposed the 3D Block Matching (BM3D) algorithm based on similarity between image patches. This method not only achieves higher signal-to-noise ratio but also provides excellent visual results, though it comes with relatively high computational complexity. The BM3D implementation typically involves two main stages: hard-thresholding for initial estimate and Wiener filtering for final refinement, utilizing 3D transformations of grouped similar blocks.

Furthermore, with the rapid development of deep learning in recent years, deep learning-based image denoising algorithms have also made significant progress. Deep learning models can learn the distribution patterns of noise from large amounts of image data, thereby more accurately removing noise from images. These deep learning algorithms not only provide better denoising performance but also feature lower computational complexity, making them one of the current research hotspots. Common implementations use convolutional neural networks (CNNs) with architectures like DnCNN or UNet, trained on noisy-clean image pairs using loss functions like MSE or perceptual loss.

In summary, image denoising is an important and challenging problem, and numerous classic algorithms along with deep learning-based approaches have been proposed. These algorithms can achieve excellent denoising results in different application scenarios, but many challenges remain to be addressed. Future research directions include further improving denoising performance, reducing computational complexity, and exploring more application scenarios through optimized algorithm implementations and hybrid approaches combining traditional methods with deep learning techniques.