#> Explain the steps in digital image #> What is visual percep1on?

visual percep1on? Explain the #> Explain different color models (RGB, #> What is Discrete Cosine Transform
processing with a neat diagram. elements of human visual percep1on. CMY, HSI, YCbCr). (DCT)? Explain with its role in JPEG
Ans. The major steps in Digital Image Ans. Visual Perception is the ability of the Ans. RGB Color Model (100 words): compression.
human eye and brain to interpret and understand The RGB model is the most common color Ans. The Discrete Cosine Transform (DCT) is
Processing are:
the information contained in light patterns representation used in digital displays, cameras, a mathematical technique that represents an image
1. Image Acquisition – Capturing the image and scanners. It is based on three primary colors:
received from the environment. In image as a sum of cosine functions oscillating at different
through a sensor or camera. Red, Green, and Blue. Each pixel is represented
processing, it plays a vital role because most frequencies. Unlike the Fourier Transform, which
2. Image Enhancement – Improving image processing techniques are designed to match the as a combination of these components, with uses both sine and cosine, DCT relies only on
quality (contrast stretching, noise reduction). way humans perceive images. Our perception is intensity values ranging from 0 to 255 in 8-bit cosine functions, making it more efficient for
3. Image Restoration – Removing influenced by brightness, contrast, color, and systems. When combined at full intensity, they image compression because most of the signal
degradations like blurring and distortions motion, which guide how images are interpreted. produce white, while absence of all gives black. energy is concentrated in a few low-frequency
using mathematical models. Elements of Human Visual Perception: Different combinations generate a wide range of components. The formula for 2D DCT is:
4. Color Image Processing – Handling 1. Luminance and Brightness – Luminance is the colors. This is an additive model, meaning colors
RGB/HSV color models and color measurable intensity of light, while brightness is are created by adding light. It is widely used in 𝐹(𝑢, 𝑣)
transformations. the subjective perception of that intensity. monitors, mobile screens, and image processing 𝑀−1 𝑁−1
(2𝑥 + 1)𝑢𝜋 (2𝑦 + 1)𝑣𝜋
5. Wavelets & Multiresolution Processing – 2. Contrast – The difference in intensity between applications. = 𝛼(𝑢)𝛼(𝑣) 0 0 𝑓(𝑥, 𝑦) 𝑐𝑜𝑠 1 2 𝑐𝑜𝑠 1 2
2𝑀 2𝑁
Analyzing images at different scales for objects and their background that makes features 𝑥=0 𝑦=0
distinguishable. CMY Color Model (100 words): Role in JPEG Compression:
compression and recognition.
3. Color Perception – Human eyes detect colors The CMY model uses Cyan, Magenta, and > An image is divided into 8×8 pixel blocks.
6. Compression – Reducing image storage Yellow as primary colors. It is a subtractive
through cones sensitive to red, green, and blue > DCT is applied to each block, converting spatial
and transmission size using lossless/lossy model, meaning colors are formed by subtracting
wavelengths. data into frequency coefficients.
methods. 4. Visual Angle and Resolution – Defines how light from white. White light minus red gives > Low-frequency coefficients (which carry most
7. Morphological Processing – Extracting much detail the human eye can perceive at a given cyan, minus green gives magenta, and minus blue visual information) are retained, while many high-
shapes, edges, and structures from images. distance. gives yellow. When all three are combined at full frequency coefficients (less important to human
8. Segmentation – Dividing an image into 5. Adaptation – The eye’s ability to adjust to intensity, they produce black. In practice, the eye) are quantized or discarded.
meaningful regions or objects. varying levels of illumination. CMYK model (adding black, K) is used to > This reduces storage while maintaining
9. Representation & Description – improve quality and reduce ink usage. CMY is acceptable image quality.
Representing segmented regions and #> Explain image sampling and commonly applied in printing and publishing
extracting features like shape, texture. quan1za1on with diagrams. industries, where inks absorb certain wavelengths #> Explain the concept of Singular Value
10. Object Recognition – Identifying and Ans. Image Sampling and Quantization are the and reflect others. It is the exact complement of
the RGB model used in displays.
Decomposition (SVD) in image
classifying objects in the image. two fundamental steps in converting a continuous
processing.
image into a digital image.
> Image Sampling: Sampling refers to dividing HSI Color Model (100 words): Ans. Singular Value Decomposition (SVD) is a
the continuous image into small discrete points The HSI model represents colors in terms of Hue, mathematical technique used in image processing
(pixels). Each pixel represents the image intensity Saturation, and Intensity, making it closer to to factorize an image matrix into three separate
at a particular location. The number of samples in human visual perception. Hue specifies the matrices. If an image is represented as a matrix 𝐴
the x and y directions determines the spatial dominant color (such as red, blue, green). of size 𝑚 × 𝑛𝑚, its SVD is given by:
resolution. A higher sampling rate gives better Saturation defines the purity or vividness of the 𝐴 = 𝑈𝛴𝑉 !
quality but requires more storage, while lower color, with high saturation being pure and low Here, U and V are orthogonal matrices containing
sampling causes loss of detail and aliasing. saturation tending toward gray. Intensity eigenvectors, while Σ is a diagonal matrix
> Image Quantization: After sampling, each represents brightness, determining how light or containing singular values arranged in decreasing
pixel’s intensity value is approximated to the dark the color appears. This model is often order. The singular values indicate the importance
nearest level from a set of finite intensity levels. represented as a hexacone or cylinder. HSI is of each corresponding component in representing
#> Describe the components of an particularly useful in image analysis and the image.
This process is called quantization. The number of
image processing system. quantization levels defines the gray-level computer vision, as it separates chromatic In image processing, SVD is widely used for
Ans. An Image Processing System is made up of resolution. For example, 8-bit quantization components from intensity, making tasks like image compression, since only the largest
several essential components that work together to provides 256 gray levels. Higher levels give segmentation and feature detection more effective. singular values and corresponding vectors are
acquire, process, and output digital images. The smoother images, while fewer levels cause false needed to reconstruct a good approximation of the
first stage is the Image Sensor or Acquisition contours. YCbCr Color Model (100 words): image, reducing storage requirements. It is also
Device, such as a camera or scanner, which The YCbCr model is widely used in television applied in noise reduction, watermarking, and
captures the real-world scene and converts it into broadcasting, video compression, and pattern recognition, because SVD provides
digital form. This data is then transferred to the JPEG/MPEG formats. It separates an image into stable and compact representations. The strength
Computer System, which acts as the core and luminance (Y) and chrominance (Cb, Cr) of SVD lies in its ability to capture essential image
executes algorithms for enhancement, restoration, components. Y represents brightness or intensity, features while discarding redundancies.
segmentation, and recognition. To improve while Cb measures the difference between blue
efficiency, Specialized Hardware like GPUs or and luminance, and Cr measures the difference #> What is Principal Component Analysis
frame grabbers may be used. The system also between red and luminance. This separation is (PCA)? Explain its role in image
relies on Software, which provides tools, useful because the human eye is more sensitive to
brightness details than color details. As a result,
dimensionality reduc1on.
algorithms, and user interfaces. The results are
viewed on Display Devices such as monitors, or chrominance components can be compressed Ans. Principal Component Analysis (PCA) is a
printed through Hardcopy Devices. Finally, Mass more than luminance, reducing file size while statistical technique used to reduce the
Storage Units are used to save raw and processed maintaining visual quality. This model is a dimensionality of data while preserving most of its
images for future use. Together, these components transformed version of RGB and is crucial for important information. In image processing, an
form a complete image processing environment. multimedia applications image can be seen as a high-dimensional dataset
where each pixel contributes to the overall
Components in short: #> Explain the Fourier Transform in
dimension. PCA works by transforming the
> Image Sensor/Acquisition Device image processing and its applications. #> Describe the Discrete Fourier
> Computer System original correlated pixel data into a new set of
Ans. The Fourier Transform (FT) is a powerful Transform (DFT) and its properties. uncorrelated variables, called principal
> Specialized Hardware mathematical tool used in image processing to Ans. The Discrete Fourier Transform (DFT) is
> Software components, which are linear combinations of the
represent an image in the frequency domain a mathematical technique that converts a finite original variables.
> Display Devices rather than the spatial domain. In simple terms, it sequence of digital image values into its frequency
> Hardcopy Devices The first principal component captures the
decomposes an image into sinusoidal components components. For a 2D image of size 𝑀 × 𝑁𝑀, the maximum variance (most important information)
> Mass Storage of different frequencies. Low-frequency DFT is given by: in the data, the second captures the next highest
#> Write short notes on basic components represent smooth variations variance, and so on. By keeping only the top few
(background or illumination), while high- (!/ .!/
principal components, we can represent the image
relationships between pixels (neighbors, &' *+
𝐹(𝑢, 𝑣) = ( ( 𝑓(𝑥, 𝑦)𝑒 !"#$( ( ) , )
frequency components capture edges, fine details, with fewer dimensions while discarding less
connectivity, regions, distance). and sharp transitions. The 2D Fourier Transform '01 +01
significant details.
Ans. In digital image processing, the is commonly applied to digital images to analyze PCA is widely used for image compression,
relationships between pixels are fundamental for Here, 𝑓(𝑥, 𝑦) is the image in the spatial domain,
and manipulate their frequency content. The pattern recognition, and face recognition
analyzing shapes, regions, and structures. The first and 𝐹(𝑢, 𝑣) represents its frequency-domain
Inverse Fourier Transform allows (Eigenfaces method). Its role in dimensionality
important concept is neighborhood. A pixel has equivalent. The inverse DFT reconstructs the
reconstruction of the original image from its reduction is crucial, as it reduces storage space,
surrounding pixels that form its neighbors. The image from its frequency components.
frequency representation. computational cost, and noise, while maintaining
most common are 4-neighbors, which are the Properties of DFT:
Applications: visual quality and important features.
pixels directly above, below, left, and right, and 1. Linearity – DFT of a sum equals the sum of
1. Image Filtering – Noise removal using low-
diagonal neighbors, which are the four corner DFTs.
pass filters and edge enhancement using high-pass
pixels. Combining these gives 8-neighbors, which 2. Periodicity – Both input and output sequences
filters.
include both direct and diagonal neighbors. Based are periodic.
2. Image Compression – Reducing data size by
on these neighborhoods, we define connectivity, 3. Translation/Shifting – Shifting in spatial
discarding less significant frequency components.
which determines whether two pixels belong to domain introduces a phase shift in frequency
3. Pattern Recognition – Frequency features aid
the same component. In 4-connectivity, only domain.
in object identification.
direct neighbors are considered, while in 8- 4. Convolution Theorem – Convolution in spatial
4. Image Restoration – Correcting blurring or
connectivity all surrounding pixels are included. A domain corresponds to multiplication in frequency
motion effects in degraded images.
mixed form, m-connectivity, avoids ambiguity in domain.
certain cases. A region in an image is then formed 5. Separability – 2D DFT can be computed using
by a group of connected pixels with similar successive 1D DFTs.
intensity values, representing a meaningful object.
To measure separation, distance metrics like
Euclidean, city-block, and chessboard distances
are used. These relationships are vital for
segmentation and boundary detection
#> Explain intensity transforma1ons (log, #> Differentiate between smoothing and #> Write short notes on sharpening in #> Explain adaptive filters and their uses
power-law, contrast stretching). sharpening spatial filters. frequency domain. Write short notes on in image restoration.
Ans. Intensity transformations are fundamental Ans. Smoothing Spatial Filters: Smoothing sharpening in frequency domain. Ans. Adaptive filters are advanced noise-
techniques in image enhancement where pixel spatial filters are image enhancement techniques Ans. Sharpening in the frequency domain is reduction techniques where the filter parameters
values are modified using mathematical functions. designed to reduce noise and minor variations in performed using high-pass filters (HPF). Since change depending on local image characteristics.
The three most common transformations are log, intensity. They work by replacing each pixel’s edges and fine structures correspond to high Unlike fixed filters (mean or median), adaptive
power-law, and contrast stretching. value with the average or median of its frequencies, these filters enhance such details by filters adjust dynamically to preserve edges and
> Log Transformation: Defined as 𝑠 = neighborhood, thus removing abrupt changes in allowing high-frequency components to pass details while removing noise.
𝑐𝑙𝑜𝑔 (1 + 𝑟), where 𝑟 is input intensity. It brightness. The most common linear smoothing while suppressing low-frequency background One common type is the adaptive median filter,
expands low-intensity values while compressing filter is the mean filter, where convolution with a information. which expands its neighborhood size if noise
high-intensity ones. This is useful in images with kernel produces a blurred effect. Another popular > Ideal HPF: Passes all frequencies above cutoff, persists, effectively handling high-density salt-
very dark regions, such as medical X-rays, where choice is the Gaussian filter, which applies but produces ringing artifacts. and-pepper noise. Another example is the
hidden details need to be highlighted. weighted averaging, giving more importance to > Butterworth HPF: Provides smoother adaptive Wiener filter, which calculates local
> Power-law (Gamma) Transformation: central pixels for natural blurring. Non-linear transitions, order controls sharpness. mean and variance to filter each pixel differently.
Expressed as 𝑠 = 𝑐𝑟 " . The value of gamma filters like the median filter are particularly > Gaussian HPF: Most effective; enhances edges Adaptive filters are widely used in image
(γ\gammaγ) determines the effect. If 𝛾 < 1, the effective in removing impulse noise (salt-and- naturally without introducing artifacts. restoration because they balance noise
image becomes brighter; if 𝛾 > 1, it becomes pepper noise). Applications: Sharpening is critical in medical suppression and detail preservation. For example,
darker. Power-law transformations are widely The main effect of smoothing is noise reduction, imaging (to highlight tumors), satellite image in satellite images, adaptive filters reduce
used in display systems, where gamma correction but it comes at the cost of blurring edges and fine processing (to detect terrain boundaries), and atmospheric noise while maintaining terrain
ensures proper brightness on monitors and TVs. details. Smoothing is widely used in document analysis (for character recognition). edges. In medical imaging, they suppress random
> Contrast Stretching: Enhances contrast by preprocessing before segmentation, compression, It improves interpretability by making boundaries noise while preserving fine structures critical for
expanding the range of gray levels. For example, and feature extraction. It prepares images for more visible. diagnosis.
in a low-contrast image, gray levels might be further processing by eliminating unwanted noise, The main advantage is flexibility, as they adapt to
compressed into a narrow range, making details though careful use is required to avoid losing #> Explain different noise models in non-uniform noise across different regions of an
unclear. By stretching them to span the full 0–255 important structures. images. image. However, they are computationally more
range, visibility improves drastically. Ans. Noise in digital images refers to unwanted expensive compared to fixed filters. Overall,
Sharpening Spatial Filters: Sharpening random variations in pixel values that degrade adaptive filters are powerful tools for high-quality
#> What is histogram equalization? spatial filters are designed to enhance fine details image restoration where preserving edges is
visual quality. It can arise due to sensors,
Explain with example. and emphasize edges in an image. Unlike essential.
transmission errors, or environmental factors.
smoothing, which suppresses high-frequency
Ans. Histogram equalization is a widely used content, sharpening highlights these components
Common noise models include:
image enhancement technique that improves > Gaussian Noise: Follows a normal distribution, #> Write short notes on band-reject and
by enhancing intensity changes. This is typically band-pass filters.
contrast by redistributing pixel intensity values. In appearing as random variations across the image.
achieved through derivative-based operators. Ans. Band-reject filters (BRF) and band-pass
many images, especially low-contrast ones, gray It is common in electronic circuits and sensors.
For example, the Sobel and Prewitt operators filters (BPF) are frequency domain filters used to
levels are concentrated in a narrow range, making > Salt-and-Pepper Noise: Appears as random
detect edges by approximating gradients, while selectively modify certain frequency components.
details difficult to observe. Histogram black and white pixels. It is caused by faulty
the Laplacian operator enhances regions of rapid > Band-reject filters attenuate frequencies within
equalization spreads these values across the entire pixels, transmission errors, or dust on sensors.
intensity change. a specified range (band) while allowing low and
intensity range (0–255 for 8-bit images), thereby > Poisson Noise: Also known as photon shot
Sharpening makes blurred or unclear features high frequencies outside the band to pass. They
enhancing visibility. noise, it occurs due to the statistical nature of
more visible, which is crucial in applications like are useful for removing periodic noise patterns
The method works by using the cumulative photon detection in low-light conditions.
medical imaging, satellite analysis, and character (e.g., stripes in scanned images). The notch filter
distribution function (CDF) of the histogram to > Speckle Noise: Multiplicative in nature,
recognition. However, sharpening also amplifies is a special case of BRF targeting very narrow
remap pixel intensities. Dark areas become commonly found in coherent imaging systems like
noise because noise exists in the high-frequency bands.
brighter, while bright areas adjust to reveal hidden radar and ultrasound.
domain, similar to edges. To counter this, it is > Band-pass filters, on the other hand, allow only
details. > Uniform Noise: Random values are uniformly
often combined with smoothing techniques for a specific band of frequencies to pass while
Example: Consider a dim image where most pixel distributed over a range.
balanced enhancement. blocking low and high frequencies. These are
values lie between 40 and 100. After histogram Understanding noise models is crucial for
equalization, these values are redistributed across selecting appropriate filtering and restoration useful for enhancing specific texture details in
the full 0–255 range. As a result, the image #> Explain frequency domain filters – images, as certain features may exist in mid-
techniques. For example, Gaussian noise is best
appears clearer with enhanced contrast. Ideal, BuHerworth, and Gaussian. removed by mean or Wiener filters, while salt- frequency ranges.
Applications include medical imaging, satellite Ans. I n frequency domain filtering, an image is and-pepper noise is effectively removed using Both filters can be implemented as Ideal,
imagery, and photography, where bringing out transformed using Fourier Transform, modified median filters. Thus, noise models form the Butterworth, or Gaussian depending on
hidden details is crucial. However, it may with a filter function, and then inverse transformed foundation of image restoration and application requirements. They play an important
sometimes amplify noise in smooth regions back. Filtering in frequency domain is powerful enhancement techniques. role in noise removal, feature extraction, and
because it allows selective modification of image medical/satellite image processing.
#> Explain the basics of spatial filtering frequency components. #> Describe mean filters and order-
in image enhancement. > Ideal Filters: These have a sharp cutoff
statistics filters (e.g., median filter) #> What are notch filters? Explain with
Ans. Spatial filtering is an important technique in frequency. For low-pass, frequencies below cutoff
are passed; above cutoff are completely blocked. Ans. Mean filters are linear smoothing filters example.
image processing where operations are performed used to reduce random noise in images. The most Ans. Notch filters are specialized frequency
directly on pixels within a neighborhood using a While simple, the sharp transitions cause ringing
artifacts (Gibbs phenomenon) in the spatial common is the arithmetic mean filter, which domain filters designed to remove or suppress
filter mask (kernel). The mask is a small matrix replaces a pixel with the average value of its
domain. very narrow frequency components from an
that slides over the image, and the output pixel is neighbors. While effective against Gaussian noise,
> Butterworth Filters: These provide a smoother image. They are particularly effective against
obtained by combining neighborhood pixel values it causes blurring of edges and fine details.
transition between passband and stopband. The periodic noise, which appears as repeating
according to the filter. Variants include geometric mean filter and
sharpness of the cutoff depends on the filter’s patterns or stripes in the spatial domain.
> Linear Filtering: Uses convolution where each harmonic mean filter, which preserve edges
order. They strike a balance between performance In the frequency spectrum, periodic noise
output pixel is the weighted sum of neighbors. better and are useful for specific noise types.
and artifact reduction. manifests as distinct spikes at certain frequency
Example: averaging filter (smoothing). Order-statistics filters, on the other hand, are
> Gaussian Filters: Based on the Gaussian points. A notch filter selectively attenuates these
> Non-linear Filtering: Uses operations like nonlinear and depend on sorting pixel values in a
function, they provide the smoothest frequency frequencies while leaving the rest of the image
median instead of convolution. Example: median neighborhood. The most popular example is the
transition. Gaussian low-pass removes noise unaffected. For example, if an image has
filter for removing salt-and-pepper noise. median filter, which replaces a pixel with the
effectively, while Gaussian high-pass enhances horizontal stripes due to sensor malfunction, the
Spatial filters are of two major types: median of surrounding values. It is highly
edges naturally. corresponding spikes in the Fourier transform are
> Smoothing (Low-pass): Removes high- effective against salt-and-pepper noise without
Applications: removed using notch filters, thereby eliminating
frequency components like noise and fine edges. blurring edges. Other types include max filter
> Low-pass filtering: Noise reduction, the stripes in the restored image.
Produces blurred images but useful for noise (removes pepper noise) and min filter (removes
smoothing. Types of notch filters include Ideal Notch Reject,
reduction. salt noise).
> High-pass filtering: Edge enhancement, detail Butterworth Notch Reject, and Gaussian Notch
> Sharpening (High-pass): Highlights high- Thus, mean filters are suitable for Gaussian noise
extraction. Reject. Gaussian notch filters are preferred as they
frequency components like edges, fine lines, and reduction, while order-statistics filters excel in minimize artifacts.
details. Uses derivative-based operators like impulse noise removal, making them
#> Write short notes on smoothing in Application: Used in restoring scanned
Sobel, Prewitt, and Laplacian. complementary tools in image restoration. documents, medical X-rays, and satellite
Applications: Spatial filtering is used in image frequency domain.
images where periodic interference often occurs.
denoising, edge detection, medical image Ans. Smoothing in the frequency domain is
enhancement, and object recognition. It forms performed using low-pass filters (LPF). The
the foundation of many higher-level image principle is that noise and fine details correspond
processing tasks. to high-frequency components, while important
image structures lie in low frequencies. By
allowing low frequencies to pass and attenuating
high frequencies, smoothing suppresses noise and
produces cleaner images.
> Ideal LPF: Sharp cutoff, may cause ringing
artifacts.
> Butterworth LPF: Smoother than ideal,
transition controlled by order.
> Gaussian LPF: Smoothest, avoids ringing
completely, widely preferred.
Applications include denoising medical images,
remote sensing, preprocessing before
segmentation, and reducing compression
artifacts. For example, in satellite images,
smoothing helps reduce random variations caused
by atmospheric effects.
#> Explain the Wiener filter and its #> Explain sub-band coding in wavelet #> Differentiate between lossless and lossy #> Explain lossy predictive coding.
applications. transform. compression techniques. Ans. Lossy predictive coding is a compression
Ans. The Wiener filter is an optimal image Ans. Sub-band coding is a key technique in Ans. Image compression techniques are broadly technique where the value of a pixel is predicted
restoration technique that minimizes the mean wavelet transform-based image processing. The classified into lossless and lossy based on whether from its neighbors, but unlike lossless methods,
square error (MSE) between the original and main idea is to divide an image into frequency sub- information is preserved or discarded. the prediction error is quantized before
restored image. Unlike inverse filtering, which bands using filters, where each band represents > Lossless Compression: In lossless methods, the transmission. Quantization introduces small
only considers degradation, Wiener filtering different levels of detail. A pair of filters—low- reconstructed image is identical to the original distortions, but it significantly reduces the number
accounts for both degradation function and noise pass and high-pass—are applied along rows and with no loss of information. These techniques of bits required, leading to much higher
statistics. columns of the image. This creates four sub- exploit coding redundancy and statistical compression ratios.
The Wiener filter in the frequency domain is given images: correlation between pixels. Common methods The process involves three steps:
by: > LL (approximation): low-frequency content include Run-Length Encoding (RLE), Huffman 1. Prediction – Estimate current pixel value from
with coarse structure. Coding, Arithmetic Coding, and LZW (used in past neighbors.
⎡ ⎤
𝑯∗ (𝒖, 𝒗) > LH (horizontal details): highlights vertical 2. Error Calculation – Difference between actual
"(𝒖, 𝒗) = ⎢
𝑭 ⎥ 𝑮(𝒖, 𝒗) PNG and GIF formats). Lossless compression is
⎢∣ 𝑯(𝒖, 𝒗) ∣𝟐 + 𝑺𝒏(𝒖, 𝒗)⎥ edges. preferred in applications where accuracy is and predicted pixel.
⎣ 𝑺𝒇 (𝒖, 𝒗)⎦ > HL (vertical details): highlights horizontal critical, such as medical imaging, satellite images, 3. Quantization & Encoding – The error is
Here, 𝐻(𝑢, 𝑣) is the degradation function, 𝑆# is the edges. quantized into fewer levels and encoded.
and technical drawings. However, the
noise power spectrum, and 𝑆$ is the original image > HH (diagonal details): fine high-frequency For example, if the actual pixel is 125, the
compression ratio is generally low (around 2:1 to
power spectrum. components. 3:1). predicted value is 124, error = +1. After
This formulation ensures that frequencies where This process can be repeated on the LL band to > Lossy Compression: Lossy techniques allow quantization, this may be approximated as 0,
noise dominates are suppressed, while useful achieve multi-level decomposition. Each sub- some loss of image information in exchange for which slightly changes the pixel upon
information is preserved. band can then be processed separately for much higher compression ratios. They exploit reconstruction. However, to the human eye, this
Applications: compression, enhancement, or noise reduction. both redundancy and psycho-visual irrelevance by difference is negligible.
> Medical Imaging: Enhancing CT or MRI scans In image compression, sub-band coding plays a discarding details less noticeable to the human Lossy predictive coding is widely used in video
corrupted by noise. major role. Since the human eye is more sensitive eye. Techniques like Transform Coding (DCT in compression (MPEG, H.264) and image
> Satellite Imaging: Removing atmospheric blur to low-frequency components, the LL band is JPEG), Wavelet Coding (JPEG2000), and compression systems where some quality can be
and noise. preserved with higher accuracy, while high- Quantization are commonly used. Lossy sacrificed for efficiency.
> Restoration of Old Photos: Improving quality frequency bands (LH, HL, HH) can be quantized compression achieves ratios as high as 20:1 or
of degraded images. or discarded, reducing storage requirements. more, making it ideal for multimedia storage, #> Write short notes on transform
Sub-band coding enables efficient image streaming, and web usage. However, repeated coding.
#> What is thresholding segmenta1on? representation, where the most important visual compression can degrade quality.
information is captured in fewer coefficients. This Ans. Transform coding is one of the most widely
Explain with examples used techniques in image compression. The main
principle is the foundation of JPEG2000
Ans. Thresholding is one of the simplest and most compression, making wavelet-based sub-band #> Explain variable length coding with an idea is to transform an image from the spatial
widely used image segmentation techniques. It example (e.g., Huffman coding). domain (pixel values) into a frequency domain,
coding superior to traditional block-based
separates objects from the background based on where redundancies can be better identified and
approaches. Ans. Variable Length Coding (VLC) is a popular
pixel intensity values. The idea is to choose a removed.
technique in image compression where frequently
threshold 𝑇: #>Write short notes on multi-resolution The most common transforms are the Discrete
occurring symbols are assigned shorter codes,
𝟏 𝒇(𝒙, 𝒚) ≥ 𝑻 expansions. Cosine Transform (DCT) and the Discrete
𝐠(𝐱, 𝐲) = J while less frequent symbols are assigned longer
𝟎 𝒇(𝒙, 𝒚) > 𝑻 Ans. Multi-resolution expansion is a fundamental codes. This ensures that the overall average code Wavelet Transform (DWT). These transforms
Here, pixels above threshold are assigned to the concept in image processing that allows an image length is reduced, leading to compression without compact most of the image’s energy into a few
object (foreground), and those below are assigned to be represented at different levels of detail or loss of information. Unlike fixed-length coding, low-frequency coefficients, while the high-
to the background. resolution. Instead of processing the entire image where every symbol uses the same number of bits, frequency coefficients (fine details and noise) are
> Global Thresholding: A single threshold value at full resolution, multi-resolution techniques VLC adapts according to symbol probability. small and can be quantized or discarded.
is used for the entire image. For example, decompose the image into a hierarchy of lower- Huffman coding is the most widely used VLC For example, in JPEG compression, the image is
separating text from a white paper in scanned resolution approximations along with the details method. It works by constructing a binary tree divided into 8×8 blocks, each block is transformed
documents. lost at each level. This is usually achieved through based on symbol frequencies. The symbols with using DCT, and then coefficients are quantized
> Local (Adaptive) Thresholding: Different wavelet transforms or pyramid representations. higher probability are placed closer to the root, and encoded. This achieves high compression
thresholds are used for different regions, useful In a pyramid structure, the original image is giving them shorter codes, while low-probability while maintaining acceptable visual quality.
when lighting is uneven. successively smoothed and subsampled, symbols get longer codes. Transform coding reduces correlation, exploits
> Otsu’s Method: An automatic thresholding producing smaller versions that capture the overall Example: Suppose pixel intensities occur with psycho-visual redundancy, and provides
technique that maximizes variance between object structure but with reduced detail. The difference probabilities: A=0.4, B=0.3, C=0.2, D=0.1. scalability. It is used in JPEG, JPEG2000,
and background. between successive levels (detail information) is Huffman coding assigns codes like: A=0, B=10, MPEG, and modern image/video codecs,
Examples: also stored. Similarly, in wavelet-based multi- C=110, D=111. Here, the most frequent symbol making it a cornerstone of lossy compression.
> In medical imaging, thresholding separates resolution expansion, an image is split into sub- (A) uses only 1 bit, while the least frequent (D)
tumors from healthy tissue. bands (low and high frequency), giving both uses 3 bits. #> Explain wavelet coding with
> In biometric systems, it isolates fingerprints or coarse approximation and fine details. applica1ons in image compression.
face regions. This approach is important in applications like #> Write short notes on bit-plane coding. Ans. Wavelet coding is a modern image
image compression, progressive transmission, and
Ans. Bit-plane coding is a compression technique compression technique that uses the Discrete
#> What are wavelets? Explain their feature extraction, since different levels of
where an image is decomposed into its binary Wavelet Transform (DWT) to represent an
resolution can be used depending on the task. For
importance in image processing. example, a lower-resolution version is sufficient representation across different bit planes. Each image in multiple frequency sub-bands. Unlike
Ans. Wavelets are mathematical functions that pixel intensity (e.g., 8-bit grayscale image) can be block-based transforms such as DCT, wavelets
for quick preview, while higher resolutions are
represent signals at different resolutions. Unlike expressed in binary form, ranging from the Most provide both spatial and frequency localization,
used for accurate analysis or reconstruction.
the Fourier Transform, which only analyzes Significant Bit (MSB) plane to the Least making them highly effective for compressing
Thus, multi-resolution expansions provide
signals in the frequency domain, wavelets provide Significant Bit (LSB) plane. The MSB planes natural images.
flexibility, efficient storage, and computational
both time (or space) and frequency information contain important structural details of the image, In wavelet coding, an image is decomposed into
savings in image processing, making them a key
simultaneously. A wavelet transform while the LSB planes mostly capture finer noise- four sub-bands at each level: LL
tool in modern image analysis and compression
decomposes an image into a set of sub-images like details. (approximation), LH (horizontal detail), HL
techniques.
with different frequency bands and resolutions. In bit-plane coding, higher-order planes are often (vertical detail), and HH (diagonal detail). The
The basic idea is to use a small “mother wavelet” #> Explain the fundamentals of image compressed with more priority, while lower-order LL sub-band contains most of the image energy
function, which is shifted and scaled to capture compression. planes may be discarded in lossy compression to and is further decomposed, creating a multi-
local details. achieve higher ratios. For example, an 8-bit pixel resolution structure. The detail coefficients (LH,
Ans. Image compression is the process of
In image processing, wavelets are highly value 201 (11001001) will be split across 8 planes. HL, HH) are typically small and can be quantized
reducing the amount of data required to represent
important because they can represent edges, This method is useful in progressive transmission, or thresholded for compression.
an image without losing essential visual quality.
textures, and fine details efficiently. Applications where the image is first transmitted using the MSB Applications include JPEG2000, which uses
Since digital images consist of millions of pixels,
include image compression (JPEG2000), planes for a rough view and later refined with wavelet coding instead of DCT. It avoids blocking
storing and transmitting them requires large
denoising, edge detection, and watermarking. additional planes. It forms the basis of many artifacts, supports progressive transmission, and
memory and bandwidth. Compression solves this
Wavelets preserve both global structure and local image compression algorithms and improves achieves better quality at high compression ratios.
by removing redundancy in the data.
variations, which makes them superior to simple efficiency. Wavelet coding is also used in medical imaging,
There are two main types of redundancies
Fourier methods for non-stationary signals. satellite images, and multimedia storage where
exploited in compression:
For example, in image compression, high- #> Explain lossless predictive coding with clarity is important.
1. Coding redundancy – Using efficient codes
frequency coefficients (fine details) can be (e.g., Huffman coding) to reduce the average example.
discarded while retaining low-frequency number of bits per pixel. Ans. Lossless predictive coding is a compression
components (smooth areas), drastically reducing 2. Spatial and temporal redundancy – method where the value of a pixel is predicted
file size without noticeable loss of quality. Thus, Neighboring pixels or frames often carry similar from its neighboring pixels, and only the
wavelets provide an excellent balance between information. prediction error (difference between actual and
compact representation and visual fidelity. 3. Psycho-visual redundancy – Human eyes are predicted value) is encoded. Since neighboring
less sensitive to certain details (like high- pixels are usually correlated, the prediction error
frequency information), so these can be discarded tends to be small and requires fewer bits to
in lossy compression. represent, leading to compression without any loss
Compression techniques are classified as lossless of information.
and lossy. Lossless retains exact data (e.g., PNG, For example, if we use a simple predictor like
GIF), while lossy sacrifices minor details for x̂(i,j) = x(i,j–1) (predict current pixel from its left
higher compression (e.g., JPEG, MPEG). neighbor), then the prediction error is:
The process of compression generally involves e(i,j) = x(i,j) – x̂(i,j).
transform coding (DCT or wavelets), If the actual pixel value is 125 and its predicted
quantization, and entropy coding. Applications value is 124, the error is only +1, which can be
include storage (cloud, archives), transmission represented with fewer bits. During decoding, the
(video calls, streaming), and efficient image original pixel is reconstructed using the predicted
databases. value and the error, ensuring perfect recovery.
-------------------------PYQ---------------------------- #> What is image filtering? Compare and #> What is image transform? How an #> Explain the concept of inverse
contrast between filtering in the spatial image can be enhanced by the filtering and its applications in image
#> What is image histogram? Discuss and frequency domain.
Contrast Stretching intensity restoration.
histogram equalization method for Ans. Image filtering is a fundamental operation in
image processing where an image is modified to transformation? Ans. Inverse filtering is one of the earliest and
image enhancement. Ans. An image transform is a mathematical simplest methods of image restoration. In digital
enhance features, remove noise, or extract useful
Ans. An image histogram is a graphical information. Filtering involves applying a operation applied to convert an image from one imaging, when an image gets degraded by blur
representation of the frequency distribution of mathematical operation to pixels using a filter domain to another (e.g., spatial domain to (e.g., motion blur, defocus blur) or noise, it can be
pixel intensities in an image. On the x-axis, it mask (kernel) in the spatial domain or modifying frequency domain) or to modify pixel intensity modeled as the original image passed through a
shows possible gray levels (0–255 for an 8-bit frequency components in the frequency domain. values. Transforms are used in image processing degradation function (system) plus noise. The goal
image), and on the y-axis, it shows the number of Spatial Domain Filtering: for enhancement, compression, filtering, and of inverse filtering is to reverse the effect of this
pixels corresponding to each intensity. The > Performed directly on pixels. feature extraction. Common transforms include degradation by applying the inverse of the
histogram gives useful information about image > Uses masks (e.g., 3×3 or 5×5) applied through Fourier Transform, Discrete Cosine Transform degradation function in the frequency domain.
contrast, brightness, and dynamic range. For convolution. (DCT), Wavelet Transform, and intensity-based Mathematically, if the degraded image is
example, a histogram concentrated on the left > Examples: transforms. represented as 𝐺(𝑢, 𝑣) and the degradation
indicates a dark image, while one spread across >> Smoothing filters (Mean, Median, Gaussian) Contrast Stretching for Enhancement: function as 𝐻(𝑢, 𝑣), then the restored image can
the full range shows good contrast. for noise reduction. Contrast stretching is an intensity be estimated by dividing:
Histogram Equalization is a popular technique >> Sharpening filters (Laplacian, Sobel) for edge transformation technique used to improve the 𝐹(𝑢, 𝑣) = 𝐺(𝑢, 𝑣) / 𝐻(𝑢, 𝑣).
for image enhancement that improves contrast by detection. visibility of details in an image. Many images Applications in Image Restoration:
redistributing intensity values more uniformly > Advantages: simple, intuitive, low have narrow contrast (i.e., pixel values 1. Deblurring Images: Useful in correcting
across the available range. The method uses the computational cost for small kernels. concentrated in a limited range), making them motion blur or lens defocus in photographs.
cumulative distribution function (CDF) of the > Limitation: less efficient for large masks and look dull or washed out. 2. Satellite Imaging: Helps restore blurred
histogram to map original pixel values to new may blur details. Contrast stretching spreads out the pixel intensity images captured from space due to atmospheric
ones. Frequency Domain Filtering: values to occupy the full available range (0–255 disturbances.
For example, a low-contrast image may have pixel > Image is transformed using Fourier Transform for 8-bit images). For example: 3. Medical Imaging: Enhances MRI, CT, or X-
values clustered between 50–150. After histogram into frequency components. > A dark image with values mostly between 50 ray images for better diagnosis.
equalization, the values are stretched to cover the > Filters modify low or high frequencies (e.g., and 150 is stretched to cover 0–255. 4. Astronomical Imaging: Restores telescope
full 0–255 range, making dark areas darker and Low-Pass for smoothing, High-Pass for > This increases brightness differences, making images distorted by atmospheric turbulence.
bright areas brighter. sharpening). hidden details clearer. Limitation: It is very sensitive to noise—when
This method is widely used in medical imaging, > Examples: Ideal, Butterworth, and Gaussian 𝐻(𝑢, 𝑣) has very small values, division can
remote sensing, and photography to enhance filters. #> What is Image Compression and amplify noise significantly.
details. However, it may sometimes over-enhance > Advantages: efficient for large kernels, precise
noise or produce unnatural effects in smooth why is it important in digital
control over frequency components. #> Describe the concept of Thresholding
regions. imaging? in image segmentation and its role in
> Limitation: requires transformation back to
spatial domain; more complex to implement. Ans. Image compression is the process of separating objects from the background.
#> What is image smoothing? Explain Comparison: Spatial filtering is simple and local, reducing the size of a digital image file without Ans. Thresholding is one of the most fundamental
different linear and non-linear while frequency domain filtering is powerful for losing essential visual quality. It achieves this by techniques in image segmentation, used to
smoothing spatial filters with suitable global modifications. In practice, the choice removing redundancies in data (spatial or spectral) separate objects of interest from the background.
depends on the application and efficiency needs. and representing the image with fewer bits. It works by converting a grayscale image into a
examples. Compression can be of two types: binary image using a chosen threshold value (T).
Ans. Image smoothing is a technique used to #> Define image segmentation. Discuss > Lossless Compression: No information is lost, All pixels with intensity greater than T are
reduce noise, blur small details, and make an watershed algorithm of image and the original image can be perfectly assigned to the object (foreground), while those
image appear softer. It works by averaging or segmentation. reconstructed (e.g., PNG, GIF). with intensity less than T are assigned to the
modifying pixel values using their neighbors. Ans. Image segmentation is the process of > Lossy Compression: Some less important background. This binary representation simplifies
Smoothing is often used as a preprocessing step in partitioning an image into meaningful regions that information is discarded, achieving higher the image, making further analysis easier.
image enhancement, segmentation, and correspond to objects or areas of interest. The goal compression at the cost of slight quality loss (e.g., Role in Separating Objects from Background:
recognition. is to simplify or change the representation of an JPEG, WebP). Thresholding is highly effective when there is a
> Linear Smoothing Filters: These filters replace image into something more useful for analysis. Importance in Digital Imaging: distinct intensity difference between objects and
each pixel with a weighted sum of its neighbors. Segmentation is a crucial step in applications like 1. Storage Efficiency: Compressed images background. For instance:
The most common is the Mean Filter, where each medical imaging, object detection, and pattern require less memory, making it possible to store > In document processing, black text is separated
pixel is replaced by the average of values in its recognition. large image databases or archives. from a white background.
neighborhood (e.g., a 3×3 mask). This reduces The Watershed Algorithm is a region-based 2. Faster Transmission: Smaller file sizes reduce > In medical imaging, tumors or bones can be
random noise but also blurs edges. Another linear segmentation method inspired by topography. bandwidth requirements, enabling quicker sharing isolated from surrounding tissues.
method is the Gaussian Filter, which uses a Think of an image as a landscape where gray-level over the internet and in real-time applications like > In industrial inspection, defects can be detected
weighted average giving more importance to values represent elevations: video conferencing. against uniform surfaces.
nearer pixels, producing smoother and more > Bright regions = high peaks 3. Cost Reduction: Saves storage and Types of Thresholding:
natural blurring. > Dark regions = valleys transmission costs, especially important in cloud 1. Global Thresholding: A single threshold value
> Non-Linear Smoothing Filters: Instead of The idea is to “flood” the landscape with water services and large-scale imaging systems. is used for the entire image.
averaging, non-linear filters use order-based starting from local minima. As water fills valleys, 4. Application Support: Essential in medical 2. Local Thresholding: Different thresholds are
statistics. The most common is the Median Filter, dams (watershed lines) are built where water from imaging, satellite imaging, social media applied to different regions of the image.
where the central pixel is replaced by the median different catchment basins would merge. These platforms, and multimedia applications where 3. Adaptive Thresholding: Threshold values
of its neighborhood values. This is especially watershed lines form the boundaries of objects in massive amounts of image data are processed. change dynamically depending on local image
effective for removing salt-and-pepper noise the image. characteristics, making it suitable for non-uniform
while preserving edges better than mean filtering. The algorithm works best on gradient images, #> What are the basic components of illumination.
Examples: where object edges correspond to high gradients
> Mean filter for Gaussian noise reduction. an image compression model?
(ridges). However, it is sensitive to noise and may
> Median filter for impulse noise removal. Ans. An image compression model is designed #> Explain the steps involved in
lead to over-segmentation. To address this,
Thus, smoothing filters play a key role in noise preprocessing with smoothing or using marker- to reduce the number of bits required to represent digital image processing from
reduction, though with trade-offs in edge clarity. based watershed (where markers guide flooding) an image, while maintaining acceptable quality. acquisition to the final output.
improves results. The basic components are: Ans. Digital image processing involves a series
Applications: medical imaging (cell or organ 1. Source Encoder (Mapper):
#> What is image sharpening? of steps that transform a raw image into a
boundary detection), document analysis, and > Transforms the input image into a more compact
Explain in detail. meaningful output.
industrial inspection. form by removing redundancies.
Ans. Image sharpening is the process of 1. Image Acquisition: The process begins with
> Techniques used: transform coding (like DCT, capturing the image using sensors, scanners, or
enhancing fine details and edges in an image. It DWT) or predictive coding.
#> Discuss the advantages and limitations cameras, followed by digitization through
makes boundaries between objects clearer and 2. Quantizer:
improves visibility of important structures. Unlike of Weiner filtering in image restoration. sampling and quantization.
> Approximates the transformed values by
smoothing (which blurs details), sharpening Ans. Wiener filtering is a powerful technique for 2. Preprocessing: Enhances image quality by
reducing precision.
highlights abrupt changes in intensity, making the image restoration that minimizes the mean square removing noise, correcting illumination, or
> This is the step where lossy compression
image look crisper and more informative. error between the restored image and the original. resizing.
occurs, as less significant details are discarded.
Techniques in Spatial Domain: Unlike inverse filtering, it not only considers the 3. Image Enhancement: Improves the visual
3. Symbol Encoder (Entropy Encoder): appearance for human interpretation using
1. Gradient-based methods: These use edge- degradation function but also accounts for noise
> Converts quantized values into binary codes techniques like histogram equalization, contrast
detection operators (like Sobel or Prewitt) to characteristics, making it more robust.
efficiently. stretching, or spatial filtering.
emphasize transitions in brightness. The edge map Advantages:
> Methods like Huffman coding or Arithmetic
is then combined with the original image to > Effectively reduces both blur and noise 4. Image Restoration: Removes degradations
coding assign shorter codes to frequent symbols
highlight boundaries. simultaneously. such as blur and noise with filters like Wiener or
and longer codes to rare ones.
2. Laplacian-based methods: These detect > Provides an optimal balance between detail inverse filtering.
4. Decoder:
regions of rapid intensity change and are widely recovery and noise suppression. 5. Segmentation: Divides the image into
> Performs the inverse operations: entropy meaningful regions or objects using thresholding,
used for edge enhancement. > Widely used in medical, astronomical, and
decoding, inverse quantization, and inverse edge detection, or watershed algorithms.
3. Unsharp Masking and High-Boost Filtering: satellite imaging for high-quality restoration.
transform. 6. Representation and Description: Extracts
Here, a blurred version of the image is subtracted > Works efficiently in the frequency domain for
> Reconstructs the image (exact in lossless,
from the original to extract details. Adding these large images. features like shape, edges, or texture for further
approximate in lossy).
details back makes the image sharper. Increasing Limitations: analysis.
the detail contribution gives stronger sharpening. > Requires prior knowledge of noise power and 7. Recognition and Interpretation: Assigns
Frequency Domain Methods: Sharpening can original image statistics, which may not always be labels and derives useful information from the
also be done by applying high-pass filters in the available. image.
frequency domain. These filters suppress low- > More complex to implement compared to simple
frequency background information and emphasize filters.
high-frequency details such as edges and textures. > Performance decreases under very high noise or
severe blur.
> Slightly higher computational requirements.