Summary of Image Enhancement and Noise Suppression

Image Enhancement & Noise Suppression for Students

SummaryKnowledge testFlashcardsPodcastMindmap

Introduction

Image enhancement improves the visual quality and usefulness of images by changing contrast, brightness, color balance, or dynamic range. The goal is to make images easier to interpret by humans or subsequent automated analysis—without inventing new information or performing restoration tasks like deblurring or denoising.

Definition: Image enhancement = selective improvement of some image features at the expense of others to increase subjective impression or information usability.

Overview of Enhancement Goals and Limits

  • Goals:
    • Improve subjective visual impression
    • Facilitate image analysis and interpretation
    • Make more efficient use of available image information
    • Increase diagnostic yield in imaging contexts
  • Limits: Enhancement does not create new information, it does not restore missing data, and it is not a substitute for denoising or edge sharpening.

Basic Concepts and Metrics

Intensity, Range and Bit Depth

  • Intensity range: difference between minimum and maximum pixel values.
  • Bit depth: number of quantization levels; determines contrast resolution.

Measures of Contrast

  • RMS contrast: root-mean-square of intensity deviations.
  • Michelson contrast: for periodic patterns or paired extremes, $\mathrm{Michelson}=\dfrac{I_{\max}-I_{\min}}{I_{\max}+I_{\min}}$.

Definition: Contrast resolution = minimum distinguishable difference between adjacent quantization levels; it depends on intensity range and bit depth.

Point-wise Contrast and Brightness Transformations

Point-wise transforms map each input pixel value $f(i,k)$ to an output value $g(i,k)$ independently.

Linear Transformations

  • Brightness change: $g(i,k)=f(i,k)+b$.
  • Gain (contrast scaling): $g(i,k)=k,f(i,k)$.
  • Combined linear: $g(i,k)=k,f(i,k)+b$.
  • Implemented via a Look-Up-Table (LUT) in discrete images.

Practical example: increase contrast by $k=2$ and decrease bias by $b=20$: $g(i,k)=2,f(i,k)-20$.

💡 Did you know?Did you know that applying a pure linear scaling to all RGB channels equally preserves hue but can clip highlights or shadows if values exceed the allowable range?

Piece-wise Linear (Windowing)

  • Divide intensity domain into intervals $R_1,R_2,\dots,R_N$ and apply different linear parameters on each.
  • Useful for emphasizing a specific dynamic range while compressing others.

Radiological Window (CT/MRI example)

  • Window defined by Width $W$ (range of emphasis) and Level $L$ (center of window).
  • Mapping formula (normalized to output range $q_{\max}$): $$g(i,k)=\dfrac{f(i,k)-L+\dfrac{W}{2}}{W},q_{\max}$$
  • Examples for CT: Bone window $W=3000\ \mathrm{HU},\ L=1524\ \mathrm{HU}$; Brain window $W=80\ \mathrm{HU},\ L=1064\ \mathrm{HU}$.
💡 Did you know?Fun fact: Using windowing in CT allows radiologists to see either soft tissue detail or bone detail from the same underlying dataset by changing $W$ and $L$.

Full Grayscale Range Stretching

  • Map image minimum $\min f$ to $0$ and maximum $\max f$ to $q_{\max}$ to use full dynamic range: $$g(i,k)=\dfrac{f(i,k)-\min f}{\max f-\min f},q_{\max}$$
  • Watch out for outliers that may be artifacts; robust methods ignore small fractions of extreme values.

Nonlinear Contrast: Gamma and Logarithmic Transforms

Gamma Transform

  • Applies power-law mapping after normalizing intensities to $[0,1]$: $$g(i,k)=\mathcal{N}\bigl(f(i,k)\bigr)^{\gamma}$$
  • If $\gamma<1$ image brightens (shadows lifted); if $\gamma>1$ image darkens.
  • Used both for aesthetic adjustments and to correct nonlinearities in capture/display (gamma correction).

Gamma correction (inverse transform) recovers linear brightness when device gamma is known.

💡 Did you know?Did you know that display systems historically used gamma because CRTs had a nonlinear luminance response, and many image pipelines still compensate for that today?

Histogram Equalization (HE)

Purpose

  • Redistribute gray levels so that the image histogram is closer to uniform, often producing a visu
Sign up for the full summary
FlashcardsKnowledge testSummaryPodcastMindmap
Start for free

Already have an account? Sign in

Image Enhancement Techniques

Klíčová slova: Image Enhancement Techniques, Edge and Sharpening Enhancement, Image Denoising, Biomedical Image Analysis

Klíčové pojmy: Image enhancement improves visual interpretability but does not create new information, Contrast measured by RMS and Michelson helps quantify changes, Linear transforms: $g=k\,f+b$ for brightness/contrast adjustments, Piece-wise linear windowing targets a specific intensity range using Width $W$ and Level $L$, Gamma transform: $g=\mathcal{N}(f)^{\gamma}$; $\gamma<1$ brightens shadows, Histogram Equalization maps via CDF: $q_g=q_{\max}D_f(q_f)$, Use AHE/CLAHE for local contrast; CLAHE clips histogram to limit noise amplification, Enhance color images by modifying the lightness channel in HSI/ Lab to preserve hue, White balance scales RGB channels so neutral reference becomes gray, Use 3×3 color correction matrix $P$ estimated from known color chart pairs, Pseudo-coloring maps scalar intensities to RGB via LUTs to increase perceptual contrast, When stretching dynamic range, ignore outlier extremes or use robust thresholds

## Introduction Image enhancement improves the visual quality and usefulness of images by changing contrast, brightness, color balance, or dynamic range. The goal is to make images easier to interpret by humans or subsequent automated analysis—without inventing new information or performing restoration tasks like deblurring or denoising. > **Definition:** Image enhancement = selective improvement of some image features at the expense of others to increase subjective impression or information usability. ## Overview of Enhancement Goals and Limits - Goals: - Improve subjective visual impression - Facilitate image analysis and interpretation - Make more efficient use of available image information - Increase diagnostic yield in imaging contexts - Limits: Enhancement does not create new information, it does not restore missing data, and it is not a substitute for denoising or edge sharpening. ## Basic Concepts and Metrics ### Intensity, Range and Bit Depth - **Intensity range**: difference between minimum and maximum pixel values. - **Bit depth**: number of quantization levels; determines contrast resolution. ### Measures of Contrast - **RMS contrast**: root-mean-square of intensity deviations. - **Michelson contrast**: for periodic patterns or paired extremes, $\mathrm{Michelson}=\dfrac{I_{\max}-I_{\min}}{I_{\max}+I_{\min}}$. > **Definition:** Contrast resolution = minimum distinguishable difference between adjacent quantization levels; it depends on intensity range and bit depth. ## Point-wise Contrast and Brightness Transformations Point-wise transforms map each input pixel value $f(i,k)$ to an output value $g(i,k)$ independently. ### Linear Transformations - Brightness change: $g(i,k)=f(i,k)+b$. - Gain (contrast scaling): $g(i,k)=k\,f(i,k)$. - Combined linear: $g(i,k)=k\,f(i,k)+b$. - Implemented via a Look-Up-Table (LUT) in discrete images. Practical example: increase contrast by $k=2$ and decrease bias by $b=20$: $g(i,k)=2\,f(i,k)-20$. Did you know that applying a pure linear scaling to all RGB channels equally preserves hue but can clip highlights or shadows if values exceed the allowable range? ### Piece-wise Linear (Windowing) - Divide intensity domain into intervals $R_1,R_2,\dots,R_N$ and apply different linear parameters on each. - Useful for emphasizing a specific dynamic range while compressing others. #### Radiological Window (CT/MRI example) - Window defined by **Width** $W$ (range of emphasis) and **Level** $L$ (center of window). - Mapping formula (normalized to output range $q_{\max}$): $$g(i,k)=\dfrac{f(i,k)-L+\dfrac{W}{2}}{W}\,q_{\max}$$ - Examples for CT: Bone window $W=3000\ \mathrm{HU},\ L=1524\ \mathrm{HU}$; Brain window $W=80\ \mathrm{HU},\ L=1064\ \mathrm{HU}$. Fun fact: Using windowing in CT allows radiologists to see either soft tissue detail or bone detail from the same underlying dataset by changing $W$ and $L$. ### Full Grayscale Range Stretching - Map image minimum $\min f$ to $0$ and maximum $\max f$ to $q_{\max}$ to use full dynamic range: $$g(i,k)=\dfrac{f(i,k)-\min f}{\max f-\min f}\,q_{\max}$$ - Watch out for outliers that may be artifacts; robust methods ignore small fractions of extreme values. ## Nonlinear Contrast: Gamma and Logarithmic Transforms ### Gamma Transform - Applies power-law mapping after normalizing intensities to $[0,1]$: $$g(i,k)=\mathcal{N}\bigl(f(i,k)\bigr)^{\gamma}$$ - If $\gamma<1$ image brightens (shadows lifted); if $\gamma>1$ image darkens. - Used both for aesthetic adjustments and to correct nonlinearities in capture/display (gamma correction). Gamma correction (inverse transform) recovers linear brightness when device gamma is known. Did you know that display systems historically used gamma because CRTs had a nonlinear luminance response, and many image pipelines still compensate for that today? ## Histogram Equalization (HE) ### Purpose - Redistribute gray levels so that the image histogram is closer to uniform, often producing a visu