Summary of Digital Image Enhancement and Noise Suppression
Digital Image Enhancement & Noise Suppression Guide
Introduction
Image denoising is the process of estimating a clean image from a noisy observation. The goal is to remove or reduce unwanted random variations (noise) while preserving meaningful structures such as edges and textures. Denoising is essential in photography, remote sensing, surveillance, and scientific imaging where noise corrupts measurements.
Definition: Noise is any unwanted random variation in pixel values that does not belong to the original scene.
1. Noise models and estimation goals
A noise model describes how noise is added to an original image. The common additive noise model is:
$$g_{i,k} = f_{i,k} + \mathcal{N}(\mu,\sigma)$$
where $g_{i,k}$ is the observed pixel, $f_{i,k}$ is the original pixel, and $\mathcal{N}(\mu,\sigma)$ is a noise random variable (often Gaussian). Denoising aims to estimate $f_{i,k}$ (denoted $\hat{f}{i,k}$) from $g{i,k}$.
Definition: Additive White Gaussian Noise (AWGN) is noise with Gaussian distribution, zero mean, and a uniform power spectrum across frequencies.
Key estimation sources:
- Neighboring pixels in a local mask
- Pixels from different realizations (multiple captures)
- Specific pixels elsewhere in the same image (non-local information)
- Whole-image priors or databases (learned priors)
Common simple estimator: averaging similar pixels to reduce variance.
2. Statistical basis: averaging and the Central Limit Theorem
Averaging several independent noisy measurements reduces variance. If you average $N$ independent samples affected by noise with variance $\sigma^2$, the result has variance $\sigma^2/N$.
Example: For repeated captures of the same scene (co-registered), the pixel-wise average is
$$\hat{f}{i,k} = \frac{1}{N} \sum{n=1}^{N} g_{i,k}^{(n)}$$
If each $g_{i,k}^{(n)} = f_{i,k} + \mathcal{N}(\mu,\sigma)$, then
$$\hat{f}{i,k} = f{i,k} + \mathcal{N}\left(\mu,\frac{\sigma^2}{N}\right)$$
Benefits and requirements:
- Reduces noise variance by factor $N$.
- Requires co-registration (images aligned) and identical scene.
💡 Věděli jste?Fun fact: Averaging $N$ independent noisy realizations reduces the noise variance by factor $N$, so variance goes as $\sigma^2/N$.
3. Spatial averaging (local filtering)
When multiple realizations are not available, we use spatial neighbors in the same image. A basic spatial estimator averages pixels in a neighborhood $\Omega_{i,k}$:
$$\hat{f}{i,k} = \frac{1}{N} \sum{n\in\Omega_{i,k}} g_n$$
A common implementation is a square uniform mask (box filter) of size $L\times L$ with $N=L^2$; its PSF (point spread function) has equal weights:
$$h_{m,n} = \frac{1}{N} \quad \text{for } m,n\text{ in mask}$$
Pros and cons:
- Simple and fast
- Good for reducing random noise in flat areas
- Blurs edges and fine details because it treats all neighbors equally
Gaussian (weighted) filter
A weighted alternative is the 2D Gaussian filter with weights decreasing with distance:
$$h(x,y) = \frac{1}{2\pi\sigma^2} e^{-\frac{x^2+y^2}{2\sigma^2}}$$
Properties:
- Smoothly reduces influence of distant pixels
- Better spectral properties than box filter
- Controls smoothing via $\sigma$
- Still blurs edges if used globally
Table: Box vs Gaussian filter
| Property | Box (Averaging) | Gaussian |
|---|---|---|
| Implementation complexity | Low | Medium |
| Edge preservation | Poor | Better |
| Parameter controlling smoothness | Mask size $L$ | $\sigma$ |
| Spectral behavior | Poor | Better |
4. Choosing neighborhood and similarity
Key questions when using spatial averaging:
- How to choose $\Omega_{i,k}$? (adjacent square, adaptive region, similar patches)
- Are noise samples independent? (often assumed locally)
- Are image pixels correlated? (usually yes, especially near edges and textures)
- How to find truly similar pixels? (non-local methods, patch matching)
Advanced approaches include:
- Growing masks based on similarity or brightness
- Rotating or anisotropic masks
- Searching the whole image for similar patches (non-local m
Zaregistruj se pro celé shrnutíFlashcardsKnowledge testSummaryPodcastMindmap
Start for free Already have an account? Sign in
Image Denoising Basics
Klíčová slova: Image Enhancement: Contrast & Histogram, Image Enhancement: Color & Sharpening, Edge Detection, Image Denoising, Biomedical Imaging
Klíčové pojmy: Denoising estimates original pixels $\hat{f}_{i,k}$ from noisy observations $g_{i,k}=f_{i,k}+\mathcal{N}(\mu,\sigma)$., Averaging $N$ independent realizations reduces noise variance to $\sigma^2/N$., Spatial averaging uses neighborhood $\Omega_{i,k}$: $\hat{f}_{i,k}=\frac{1}{N}\sum_{n\in\Omega} g_n$., Box filter: uniform weights, simple, causes edge blurring; Gaussian: weighted by distance via $h(x,y)=\frac{1}{2\pi\sigma^2}e^{-\frac{x^2+y^2}{2\sigma^2}}$., Adaptive filtering controls smoothing by local features (e.g., gradients) to preserve edges., Non-local means average similar patches across the image using similarity-based weights., Impulse noise is handled by detection + replacement; median filter is robust for salt-and-pepper noise., Choosing neighborhood and similarity metric is critical for trade-off between denoising and detail preservation., Co-registered multi-frame averaging requires alignment but yields strong noise reduction., Advanced methods: transform-domain shrinkage, statistical priors, and deep learning-based denoisers.
## Introduction
Image denoising is the process of estimating a clean image from a noisy observation. The goal is to remove or reduce unwanted random variations (noise) while preserving meaningful structures such as edges and textures. Denoising is essential in photography, remote sensing, surveillance, and scientific imaging where noise corrupts measurements.
> Definition: Noise is any unwanted random variation in pixel values that does not belong to the original scene.
## 1. Noise models and estimation goals
A noise model describes how noise is added to an original image. The common additive noise model is:
$$g_{i,k} = f_{i,k} + \mathcal{N}(\mu,\sigma)$$
where $g_{i,k}$ is the observed pixel, $f_{i,k}$ is the original pixel, and $\mathcal{N}(\mu,\sigma)$ is a noise random variable (often Gaussian). Denoising aims to estimate $f_{i,k}$ (denoted $\hat{f}_{i,k}$) from $g_{i,k}$.
> Definition: Additive White Gaussian Noise (AWGN) is noise with Gaussian distribution, zero mean, and a uniform power spectrum across frequencies.
Key estimation sources:
- Neighboring pixels in a local mask
- Pixels from different realizations (multiple captures)
- Specific pixels elsewhere in the same image (non-local information)
- Whole-image priors or databases (learned priors)
Common simple estimator: averaging similar pixels to reduce variance.
## 2. Statistical basis: averaging and the Central Limit Theorem
Averaging several independent noisy measurements reduces variance. If you average $N$ independent samples affected by noise with variance $\sigma^2$, the result has variance $\sigma^2/N$.
Example: For repeated captures of the same scene (co-registered), the pixel-wise average is
$$\hat{f}_{i,k} = \frac{1}{N} \sum_{n=1}^{N} g_{i,k}^{(n)}$$
If each $g_{i,k}^{(n)} = f_{i,k} + \mathcal{N}(\mu,\sigma)$, then
$$\hat{f}_{i,k} = f_{i,k} + \mathcal{N}\left(\mu,\frac{\sigma^2}{N}\right)$$
Benefits and requirements:
- Reduces noise variance by factor $N$.
- Requires co-registration (images aligned) and identical scene.
Fun fact: Averaging $N$ independent noisy realizations reduces the noise variance by factor $N$, so variance goes as $\sigma^2/N$.
## 3. Spatial averaging (local filtering)
When multiple realizations are not available, we use spatial neighbors in the same image. A basic spatial estimator averages pixels in a neighborhood $\Omega_{i,k}$:
$$\hat{f}_{i,k} = \frac{1}{N} \sum_{n\in\Omega_{i,k}} g_n$$
A common implementation is a square uniform mask (box filter) of size $L\times L$ with $N=L^2$; its PSF (point spread function) has equal weights:
$$h_{m,n} = \frac{1}{N} \quad \text{for } m,n\text{ in mask}$$
Pros and cons:
- Simple and fast
- Good for reducing random noise in flat areas
- Blurs edges and fine details because it treats all neighbors equally
### Gaussian (weighted) filter
A weighted alternative is the 2D Gaussian filter with weights decreasing with distance:
$$h(x,y) = \frac{1}{2\pi\sigma^2} e^{-\frac{x^2+y^2}{2\sigma^2}}$$
Properties:
- Smoothly reduces influence of distant pixels
- Better spectral properties than box filter
- Controls smoothing via $\sigma$
- Still blurs edges if used globally
Table: Box vs Gaussian filter
| Property | Box (Averaging) | Gaussian |
|---|---:|---:|
| Implementation complexity | Low | Medium |
| Edge preservation | Poor | Better |
| Parameter controlling smoothness | Mask size $L$ | $\sigma$ |
| Spectral behavior | Poor | Better |
## 4. Choosing neighborhood and similarity
Key questions when using spatial averaging:
- How to choose $\Omega_{i,k}$? (adjacent square, adaptive region, similar patches)
- Are noise samples independent? (often assumed locally)
- Are image pixels correlated? (usually yes, especially near edges and textures)
- How to find truly similar pixels? (non-local methods, patch matching)
Advanced approaches include:
- Growing masks based on similarity or brightness
- Rotating or anisotropic masks
- Searching the whole image for similar patches (non-local m