Image Processing and Enhancement

Introduction and Context

Computer vision systems rarely operate directly on ideal images. Images may contain sensor noise, motion blur, poor illumination, low contrast, compression artifacts, or irrelevant background detail. Image processing provides the mathematical and computational tools used to improve image quality, reveal useful structures, and prepare images for later tasks such as segmentation, feature extraction, recognition, and measurement.

This chapter focuses on enhancement and restoration at the image level. Enhancement aims to produce an image that is more useful for a particular task or more visually interpretable. Restoration attempts to recover an estimate of an original image from a degraded observation using a model of the degradation process. The distinction is important: enhancement is often application-dependent and may deliberately alter image appearance, whereas restoration is usually model-based and seeks fidelity to the original scene.

Image processing can be performed in the spatial domain, where operations act directly on pixel values, or in the frequency domain, where an image is represented according to its spatial frequencies. Spatial methods are often intuitive and local. Frequency-domain methods provide a useful way to analyze blur, periodic noise, and smoothing or sharpening behavior.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain the roles of convolution, correlation, kernels, and boundary handling.

2. Apply common spatial-domain filters for smoothing, sharpening, and denoising.

3. Interpret the frequency-domain meaning of low-pass, high-pass, and band-pass filtering.

4. Use histogram techniques and contrast transformations appropriately.

5. Distinguish global, adaptive, and multilevel thresholding.

6. Explain binary and grayscale morphological operations.

7. Compare image enhancement with image restoration.

8. Select suitable methods for practical computer-vision applications and evaluate their limitations.

1. Image Representation and Local Operators

A grayscale image can be represented as a function f(x, y), where x and y identify spatial position and f is the measured intensity. For an 8-bit image, intensity commonly lies in the range 0 to 255, although other bit depths and normalized ranges are possible. A color image may be represented by three channels, such as red, green, and blue, or by alternatives such as HSV, Lab, or YCbCr.

Many enhancement operations use a neighborhood around each pixel. A kernel, mask, or filter is a small matrix of coefficients. The kernel is moved across the image, and a weighted sum of neighboring pixels is calculated. For an image f and kernel h of dimensions (2a + 1) by (2b + 1), discrete convolution is:

(g = f * h)(x, y) = Σ from m = -a to a Σ from n = -b to b h(m, n) f(x - m, y - n).

Here, f is the input image, h is the kernel, and g is the output. Convolution reverses the kernel before applying it. In many computer-vision libraries, an operation called filtering is actually correlation:

g(x, y) = Σ h(m, n) f(x + m, y + n).

For symmetric kernels, such as the usual mean or Gaussian kernels, convolution and correlation produce the same result. For asymmetric kernels, such as directional edge detectors, the distinction matters.

Boundary Handling

A kernel extends beyond the image near its boundary. Common strategies include:

• Zero padding: missing pixels are treated as zero. This can create dark borders.
• Replication: the nearest boundary pixel is repeated.
• Reflection: image values are mirrored at the boundary.
• Circular wrapping: the image is treated as periodic.
• Valid convolution: only positions where the complete kernel fits are retained.

Boundary choice affects measurements, especially for small images or large kernels, and should be recorded in reproducible systems.

Linear and Nonlinear Filters

A linear filter computes a weighted sum of input values. Mean, Gaussian, and Laplacian filters are common examples. A nonlinear filter does not satisfy the linearity property and may be better suited to impulse noise or edge preservation. The median filter is a central example: it replaces each pixel with the median of values in its neighborhood.

2. Spatial-Domain Smoothing and Denoising

Smoothing reduces rapid intensity variation. It can suppress noise but may also blur edges and fine details. The correct choice depends on the noise model, the spatial scale of features, and the later vision task.

Mean Filtering

A normalized k by k mean filter assigns each coefficient the value 1/k². For a 3 by 3 filter:

h = (1/9) [[1, 1, 1], [1, 1, 1], [1, 1, 1]].

The output is the local arithmetic average. Mean filtering is simple and reduces zero-mean random noise, but it blurs edges and is sensitive to outliers.

Gaussian Filtering

A continuous two-dimensional Gaussian function is:

G(x, y) = 1/(2πσ²) exp(-(x² + y²)/(2σ²)),

where σ controls the spatial spread. In a discrete kernel, samples of this function are calculated over a finite window and normalized so that their coefficients sum to one. The normalization preserves constant regions: if every input pixel has value c, the output remains c.

Gaussian filtering is useful when noise is approximately independent and Gaussian. Increasing σ produces stronger smoothing and removes progressively lower spatial frequencies. A kernel window should be large enough to capture the significant Gaussian mass; in practice, a radius of approximately 3σ is often used, although the exact choice depends on implementation and efficiency requirements.

A two-dimensional Gaussian is separable:

G(x, y) = G1(x)G1(y),

so a two-dimensional convolution can be implemented as two one-dimensional convolutions. This reduces computational cost from approximately O(Nk²) to O(Nk) for an N-pixel image and a k by k separable kernel.

Median Filtering

For each neighborhood, the median value is selected rather than the mean. Median filtering is particularly effective against salt-and-pepper noise because isolated extreme values have little effect on the median. It preserves step edges better than a mean filter, but it can remove thin lines or small features when the window is too large.

Bilateral Filtering

Bilateral filtering combines spatial proximity and intensity similarity. A simplified form is:

g(p) = [1/Wp] Σq in neighborhood S Gs(||p - q||) Gr(|f(p) - f(q)|) f(q),

where p is the target pixel, q is a neighboring pixel, Gs is a spatial Gaussian, Gr is an intensity-domain Gaussian, and Wp is the sum of the combined weights. Pixels that are nearby and have similar intensity receive high weights. Pixels across a strong edge receive lower weights, allowing smoothing within regions while preserving many edges.

Bilateral filtering is useful for photographic denoising and tone smoothing, but it is computationally more expensive than ordinary Gaussian filtering. Its parameters require careful selection: a large intensity scale may blur edges, while a small one may leave noise.

Choosing a Denoising Method

The noise type should guide the method:

• Gaussian-like sensor noise: Gaussian, Wiener, or other statistical filters.
• Salt-and-pepper noise: median or related rank filters.
• Speckle noise: often a multiplicative model; logarithmic transformation and specialized filters may help.
• Periodic interference: frequency-domain notch filtering.
• Motion blur: restoration methods based on a blur model rather than simple smoothing.

Denoising must be evaluated with the intended task in mind. A filter that improves visual appearance may remove small objects important for detection.

3. Sharpening and Edge Enhancement

Sharpening emphasizes rapid intensity changes. It can improve perceived detail, but it also amplifies noise and may create halos.

The discrete Laplacian is a second-derivative operator. One common four-neighbor kernel is:

L = [[0, 1, 0], [1, -4, 1], [0, 1, 0]].

Another convention uses the negative of this kernel. The sign must therefore be considered when constructing the sharpened image. If L is defined as above, a common sharpening expression is:

g = f - α(L * f),

where α is a positive sharpening strength. With the opposite Laplacian convention, addition may be required instead. The Laplacian responds strongly where intensity changes rapidly, including edges and noise.

The Sobel operators estimate first derivatives while including some smoothing. Typical horizontal and vertical kernels are:

Sx = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]],

Sy = [[-1, -2, -1], [0, 0, 0], [1, 2, 1]].

The gradient components are gx = Sx f and gy = Sy f. Gradient magnitude may be estimated as:

|∇f| = sqrt(gx² + gy²),

or, for lower computational cost, approximately as |gx| + |gy|. Gradient direction is computed using atan2(gy, gx), subject to the coordinate convention. Sobel responses are useful for edge detection, but their numeric ranges may exceed the display range and require scaling or normalization.

Unsharp masking first produces a smoothed image f_s, then forms a detail image d = f - f_s. Sharpening can be written as:

g = f + k d,

where k controls enhancement. This method is intuitive and allows control over the spatial scale of details through the smoothing filter.

4. Frequency-Domain Filtering

The Fourier transform represents an image as a combination of spatial frequencies. Low spatial frequencies correspond to slowly changing intensity regions, such as gradual illumination. High spatial frequencies correspond to edges, fine texture, and noise. For a two-dimensional discrete image, the discrete Fourier transform is:

F(u, v) = Σ from x = 0 to M-1 Σ from y = 0 to N-1 f(x, y) exp[-j2π(ux/M + vy/N)].

The inverse transform reconstructs the image. Here M and N are image dimensions, (x, y) are spatial coordinates, (u, v) are frequency indices, and j is the imaginary unit. In practice, the fast Fourier transform computes these values efficiently.

The convolution theorem states that spatial convolution corresponds to multiplication in the frequency domain:

f * h ↔ F(u, v)H(u, v).

Thus, if H is a frequency response, filtering can be performed by transforming the image, multiplying by H, and applying the inverse transform. The image is often shifted before visualization so that the zero-frequency component appears at the center.

Frequency Filters

A low-pass filter retains low frequencies and suppresses high frequencies. It smooths images and reduces fine noise. An ideal low-pass filter has an abrupt cutoff and can create ringing near sharp transitions. Butterworth and Gaussian low-pass filters provide smoother transitions and generally reduce ringing.

A high-pass filter suppresses low frequencies and retains high frequencies. It emphasizes edges and fine detail but can amplify high-frequency noise. A high-pass result may be combined with the original image for sharpening.

A band-pass filter retains a selected range of frequencies. A notch-reject filter suppresses narrow frequency regions and is especially useful for periodic interference. Periodic noise often appears as symmetrical bright points in the Fourier magnitude spectrum because real-valued images produce conjugate-symmetric frequency components.

Frequency filtering has practical complications. Multiplication in the transform domain corresponds to circular convolution unless padding is used. Without suitable padding, edges of the image may interact unnaturally. Transform magnitudes usually have a large dynamic range, so log magnitude display is commonly used for visualization; this display transformation does not itself alter the underlying image filter.

5. Contrast Enhancement and Histogram Methods

The histogram of a grayscale image is a count or distribution of pixels at each intensity. If h(rk) is the number of pixels with intensity rk and the image contains MN pixels, the normalized histogram is:

p(rk) = h(rk)/(MN).

A histogram describes intensity distribution but not spatial arrangement. Two images can have identical histograms and completely different structures.

Point Transformations

A point transformation maps each input intensity r to an output intensity s = T(r). Examples include:

• Negative transformation: s = L - 1 - r for an L-level image.
• Logarithmic transformation: s = c log(1 + r), useful for expanding dark values and compressing bright values.
• Gamma transformation: s = c r^γ, after suitable intensity normalization. Values γ less than 1 brighten midtones; values greater than 1 darken them.
• Piecewise-linear contrast stretching: maps a selected input interval to a broader output interval, often with clipping outside that interval.

The transformation must respect the image representation. For normalized intensities, r is usually in [0, 1]; for integer images, output values must be clipped or rounded appropriately.

Histogram Equalization

Histogram equalization uses the cumulative distribution function (CDF). For a normalized input intensity r, the transformation is conceptually:

s = (L - 1) CDF(r),

where L is the number of intensity levels. In a discrete implementation, the CDF is calculated from the normalized histogram and the result is quantized. The transformation tends to distribute intensities across a wider range and can improve global contrast.

Equalization does not guarantee a perfectly uniform histogram, particularly for discrete images. It may over-amplify noise, produce unnatural appearance, or reduce local contrast when illumination varies spatially.

Adaptive and CLAHE Enhancement

Adaptive histogram equalization operates on local neighborhoods rather than the whole image. Contrast-limited adaptive histogram equalization (CLAHE) limits histogram amplification in each tile and interpolates neighboring tile mappings. This reduces excessive enhancement of noise and is useful in medical, industrial, and low-light imagery. However, tile size and clip limit influence the result and must be validated for the application.

6. Thresholding and Segmentation Preparation

Thresholding converts an intensity image into a binary image. A global threshold T produces:

b(x, y) = 1 if f(x, y) >= T, and 0 otherwise.

Global thresholding works when foreground and background have sufficiently distinct intensity distributions and illumination is relatively uniform.

Otsu’s method selects a threshold that maximizes between-class variance, equivalently minimizing within-class variance, for a two-class partition. It is effective when the histogram is approximately bimodal but may fail with uneven illumination, unequal class sizes, or overlapping intensity distributions.

Adaptive thresholding computes a threshold from a local neighborhood. For example, the threshold may be a local mean minus an offset. Such methods handle gradual illumination changes but may produce fragmented results if the neighborhood size is inappropriate.

Thresholding is sensitive to preprocessing. Smoothing can reduce noise before thresholding, while contrast normalization may improve class separation. Thresholding should not be judged only by visual appearance; precision, recall, intersection over union, boundary accuracy, or task-specific measurements may be necessary.

7. Mathematical Morphology

Mathematical morphology analyzes shapes using a structuring element. It is especially useful for binary images, though grayscale morphology is also important. The structuring element encodes the scale and shape of the feature of interest.

For binary sets A and B, erosion is:

A ⊖ B = {z | B translated by z is contained in A}.

Dilation is:

A ⊕ B = {z | the translated B intersects A}.

In intuitive terms, erosion shrinks foreground objects and removes structures that cannot contain the structuring element. Dilation expands objects and can bridge small gaps.

Opening is erosion followed by dilation:

A ○ B = (A ⊖ B) ⊕ B.

It removes small objects and narrow protrusions while generally preserving larger shapes. Closing is dilation followed by erosion:

A ● B = (A ⊕ B) ⊖ B.

It fills small holes and gaps and can connect nearby regions.

The structuring element matters greatly. A disk is suitable for approximately round structures, a line element can detect or preserve directional features, and a square may produce block-like effects. Morphological gradients, top-hat operations, and reconstruction-based methods provide additional tools for boundary extraction, background correction, and object isolation.

Morphology is not merely a cosmetic operation. Repeated erosion can remove legitimate small objects; repeated dilation can merge objects that should remain separate. Parameters should be selected according to object scale and validated against labeled data.

8. Image Restoration

Let the observed image be modeled as:

g(x, y) = h(x, y) * f(x, y) + η(x, y),

where f is the unknown original image, h is the point-spread or blur function, η is additive noise, and g is the observed image. Restoration seeks an estimate of f using knowledge or assumptions about h and η.

Inverse Filtering

In the frequency domain, if G = HF, an ideal inverse estimate is:

F̂(u, v) = G(u, v)/H(u, v),

where H is nonzero. In practice, inverse filtering is unstable when H is small because noise is strongly amplified. Accurate knowledge of the blur model is also required.

Wiener Filtering

Wiener filtering balances deblurring and noise suppression. A common frequency-domain form is:

F̂(u, v) = [H*(u, v)/( |H(u, v)|² + K(u, v) )] G(u, v),

where H* is the complex conjugate of H. In the simplest constant-noise formulation, K represents a ratio related to noise and signal power spectra. The exact parameterization depends on the statistical model. Wiener filtering assumes a linear degradation model and relies on estimates of blur and signal/noise statistics.

Restoration cannot recover information that has been completely destroyed. Severe blur, clipping, saturation, and undersampling make exact recovery impossible. Modern learning-based restoration can produce visually plausible detail, but generated detail may not be faithful to the original evidence. This distinction is critical in scientific, forensic, medical, and legal settings.

9. Worked Example: A 3 × 3 Mean Filter

Consider the neighborhood:

[10 20 30]

[20 30 40]

[30 40 50]

Using a 3 × 3 normalized mean filter, the center output is the average of all nine values:

(10 + 20 + 30 + 20 + 30 + 40 + 30 + 40 + 50)/9 = 270/9 = 30.

The center pixel is therefore replaced by 30. In this particular neighborhood, the original center is already 30, so the output does not change. If the center had been an isolated value of 255 while the other eight values remained unchanged, the mean would be (270 - 30 + 255)/9 = 495/9 = 55, substantially reducing the impulse. A median filter would instead sort the values and select the fifth value; for the original neighborhood, the median is 30. This illustrates why the mean is influenced by extreme values while the median is more robust to isolated impulses.

10. Applied Case: Enhancing Industrial Surface Inspection

Background

Consider an illustrative automated inspection system for detecting scratches and small pits on painted metal panels. Images are acquired by a fixed camera under illumination that varies slightly across the panel.

Problem

Raw images contain sensor noise, gradual brightness variation, and reflections. A single global threshold produces false defects in darker regions and misses defects in brighter regions. Excessive smoothing removes narrow scratches.

Analysis

The defects are high-contrast, spatially small structures. The illumination variation is low-frequency, while scratches contain higher-frequency components. Salt-and-pepper artifacts are occasionally introduced by transmission errors.

Approach

First, defective or corrupted frames are rejected through acquisition-quality checks. A small median filter removes isolated impulse artifacts without strongly blurring line-like defects. A slowly varying background is estimated with a large Gaussian filter or morphological opening, and the image is normalized relative to that background. Local contrast enhancement may then be applied, with contrast limits chosen to avoid amplifying sensor noise.

Adaptive thresholding is used because illumination is not uniform. Morphological opening removes isolated foreground pixels, while closing reconnects fragmented scratch segments. Connected-component measurements are computed, including area, elongation, orientation, and mean contrast. Components are classified using validated rules or a machine-learning model trained on representative labeled panels.

Outcome

The processing pipeline improves separation between defects and background while preserving narrow scratches better than a large mean filter. Performance should be reported using defect-level recall, false alarms per panel, localization accuracy, and processing time—not simply by visual inspection.

Learning Points

The case illustrates that enhancement is task-dependent. No single filter solves all problems. Acquisition quality, noise modeling, illumination correction, thresholding, morphology, and quantitative evaluation must be designed together. A visually attractive image is not necessarily the image that produces the most reliable inspection decisions.

11. Practical Design Principles and Limitations

A robust workflow begins with the imaging conditions. Record sensor type, bit depth, exposure, illumination, compression, and color space. Avoid converting to a low-bit-depth format prematurely. Preserve the original image and document every transformation.

Filter parameters should be selected according to feature scale. A kernel larger than the target object may remove it. Smoothing before edge detection can improve stability, but too much smoothing displaces or weakens boundaries. For color images, processing each RGB channel independently can create color artifacts; luminance-preserving spaces are often preferable for intensity enhancement.

Quantitative evaluation should use a task-specific reference. Possible measures include mean squared error and peak signal-to-noise ratio when a clean reference exists, structural similarity for perceptual comparison, and segmentation or recognition metrics when enhancement is a preprocessing step. No metric is universally sufficient.

Image processing can also create ethical and governance concerns. Enhancement of medical, biometric, surveillance, or forensic images may change interpretation. Systems should distinguish enhancement from evidence recovery, retain provenance, log parameter settings, and avoid presenting generated detail as observed fact. Privacy protection is necessary when images contain faces, license plates, medical information, or workplace activity. Bias can arise when an enhancement pipeline performs differently across skin tones, devices, environments, or object classes. Validation must therefore include the populations and conditions in which the system will be deployed.

12. Contemporary Developments

Modern vision systems increasingly use learned image restoration and enhancement models, including convolutional neural networks, transformer architectures, and diffusion-based methods. These systems can learn complex mappings for denoising, super-resolution, low-light enhancement, and deblurring. Their advantages include strong performance on representative data and the ability to model nontrivial image statistics.

Their limitations include dependence on training distributions, sensitivity to domain shift, high computational cost, and the possibility of hallucinating plausible but unsupported structures. Classical filters remain valuable because they are interpretable, computationally efficient, and easy to validate. Hybrid systems often combine physically informed preprocessing with learned models. In safety-critical applications, uncertainty estimation, out-of-distribution testing, calibration, and human review are important safeguards.

Chapter Summary

Image processing improves or analyzes images before higher-level computer-vision operations. Spatial filtering uses local neighborhoods and kernels; convolution, correlation, boundary handling, and normalization determine the result. Mean and Gaussian filters smooth noise but blur edges, whereas median and bilateral filters can preserve important structures under suitable conditions. Sharpening uses derivatives, the Laplacian, gradients, or unsharp masking and must be controlled because it can amplify noise.

Frequency-domain filtering interprets images as spatial frequencies. Low-pass filters smooth, high-pass filters emphasize detail, and notch filters can suppress periodic interference. Histogram transformations, equalization, and CLAHE modify contrast, while thresholding converts intensity information into regions for segmentation. Morphological erosion, dilation, opening, and closing operate according to a structuring element and are useful for shape-based cleanup.

Restoration differs from enhancement because it models degradation and attempts to estimate an original image. Inverse filtering is unstable when the blur transfer function is small; Wiener filtering provides a statistical compromise between deblurring and noise suppression. All methods require parameter selection, quantitative evaluation, awareness of task requirements, and careful documentation. Modern learned methods extend these capabilities but introduce risks involving generalization, interpretability, and hallucinated detail.

Key Terms

Convolution: Weighted operation in which a kernel is reversed before being applied to an image.

Correlation: Sliding weighted operation that does not reverse the kernel.

Kernel: A finite matrix of coefficients used by a local image operator.

Boundary handling: The strategy used when a filter neighborhood extends beyond an image.

Gaussian filter: A smoothing filter whose weights follow a Gaussian distribution.

Median filter: A nonlinear filter that replaces a pixel with the median of its neighborhood.

Bilateral filter: An edge-aware filter combining spatial and intensity similarity.

Laplacian: A second-derivative operator used for edge enhancement and sharpening.

Gradient: A vector of first spatial derivatives indicating intensity change and direction.

Fourier transform: Representation of an image in terms of spatial-frequency components.

Low-pass filter: A filter that retains low spatial frequencies and suppresses high frequencies.

Notch filter: A frequency filter that rejects selected narrow frequency bands.

Histogram: Distribution of image pixels over intensity levels.

Histogram equalization: Contrast transformation based on the cumulative distribution function.

Thresholding: Conversion of an intensity image into classes, often foreground and background.

Structuring element: Shape used to probe and transform image regions in morphology.

Opening: Morphological erosion followed by dilation.

Closing: Morphological dilation followed by erosion.

Image restoration: Model-based estimation of an image degraded by blur and noise.

Wiener filtering: Statistical restoration method balancing inverse filtering and noise suppression.

Review Questions

Short Answer Questions

1. What is the difference between convolution and correlation?

2. Why is a median filter effective for salt-and-pepper noise?

3. What types of image structures are represented by low and high spatial frequencies?

4. What is the purpose of a structuring element in mathematical morphology?

5. Why can inverse filtering amplify noise?

Long Answer Questions

1. Explain Gaussian smoothing, including the role of σ, normalization, separability, and edge blurring.

2. Compare global thresholding, Otsu’s method, and adaptive thresholding, including situations in which each may fail.

3. Explain histogram equalization and CLAHE and discuss their potential effects on noise and visual appearance.

4. Describe spatial-domain and frequency-domain filtering and explain the convolution theorem.

5. Compare enhancement and restoration, using inverse filtering and Wiener filtering to support your explanation.

Analytical / Application Questions

1. Design a preprocessing pipeline for document images photographed under uneven illumination. Justify the choice of denoising, contrast correction, thresholding, and morphology operations.

2. An inspection image contains periodic horizontal interference and small isolated impulse noise. Propose a frequency- and spatial-domain solution, including how you would evaluate it.

3. A super-resolution model creates visually convincing details in a forensic image. Analyze the evidential, ethical, and governance risks and recommend safeguards before the image is used in a decision.

Multiple Choice Questions with Answers

Question 1

Which operation is most resistant to isolated salt-and-pepper noise?

A. Median filtering

B. Mean filtering

C. Ideal high-pass filtering

D. Histogram stretching

Correct Answer: A. Median filtering

Question 2

What does a low-pass filter generally do to an image?

A. Enhances only isolated points

B. Suppresses gradual illumination

C. Smooths the image by reducing high frequencies

D. Converts the image directly to a binary image

Correct Answer: C. Smooths the image by reducing high frequencies

Question 3

For a normalized 3 × 3 mean filter, what is the output at the center of the worked neighborhood in the chapter?

A. 20

B. 30

C. 40

D. 270

Correct Answer: B. 30

Question 4

Which operation is defined as erosion followed by dilation?

A. Closing

B. Opening

C. Gradient thresholding

D. Unsharp masking

Correct Answer: B. Opening

Question 5

What is a major weakness of ideal inverse filtering?

A. It cannot represent low frequencies

B. It always changes grayscale images to color

C. It amplifies noise where the blur response is small

D. It requires no degradation model

Correct Answer: C. It amplifies noise where the blur response is small

Question 6

Which method uses the cumulative distribution function of the intensity histogram?

A. Histogram equalization

B. Median filtering

C. Bilateral filtering

D. Morphological closing

Correct Answer: A. Histogram equalization

Question 7

Why is adaptive thresholding useful under uneven illumination?

A. It uses thresholds derived from local neighborhoods

B. It removes all high-frequency information

C. It requires a perfectly bimodal global histogram

D. It guarantees that every object is detected

Correct Answer: A. It uses thresholds derived from local neighborhoods

Question 8

Which filter combines spatial distance with intensity similarity to preserve many edges?

A. Mean filter

B. Bilateral filter

C. Ideal low-pass filter

D. Laplacian-only filter

Correct Answer: B. Bilateral filter

Question 9

In the degradation model g = h * f + η, what does η represent?

A. The original image

B. The blur kernel

C. The observed image

D. Additive noise

Correct Answer: D. Additive noise

Question 10

What is a principal risk of learned image restoration in forensic or scientific use?

A. It can never process color images

B. It may generate plausible details unsupported by the original evidence

C. It always produces lower contrast than the input

D. It cannot be evaluated quantitatively

Correct Answer: B. It may generate plausible details unsupported by the original evidence