bildedensning

Bildedensning, or image denoising, refers to techniques in digital image processing that aim to remove noise from an image while preserving edges and textures. Noise can arise from sensor limitations, low light, compression, or transmission errors. The goal is to improve visual quality and enable more accurate analysis.

Techniques fall into several groups. Spatial-domain methods include simple filters such as median or Gaussian filters. Transform-domain approaches use representations like wavelets and apply shrinkage or thresholding to reduce noise. Advanced patch-based algorithms, such as non-local means and BM3D, exploit self-similarity in images. Modern approaches increasingly rely on data-driven methods, particularly convolutional neural networks, which learn denoising priors from data and can adapt to different noise levels.

Common noise models include Gaussian noise, Poisson noise in low-light imagery, and multiplicative speckle noise in coherent imaging. Denoising methods often assume a statistical model of the noise and the underlying image, and performance depends on how well these models fit real-world conditions.

Challenges include preserving fine details and textures while removing noise, avoiding artifacts such as over-smoothing or ringing, and adapting to a wide range of image content and noise characteristics. Computational cost is also a consideration for high-resolution images and real-time applications.

Applications span consumer photography, video restoration, medical imaging (MRI, CT, ultrasound), astronomy, and remote sensing. Evaluation typically uses objective metrics like PSNR and SSIM, alongside perceptual assessments. The field has evolved from linear filtering to sparsity- and Bayesian-based methods, culminating in deep learning approaches that offer substantial gains in quality.