L1 - Introduction to Image Analysis L2 - Image Enhancement 1 L3 - Morphology L4 - Image Segmentation 1
Digital image = a matrix of discrete integer values where each pixel Image enhancement means preprocessing an image to improve Morphology is image processing based on shape and set theory. It is Edges → Discontinuities or sudden change in image intensity
stores intensity or color. A pixel is a picture element, and its position is technical quality so later analysis such as segmentation or recognition mainly applied to binary images, where pixels are 0 or 1 after Canny
written as f(x,y), where x and y are coordinates and f is the intensity works better. Main degradations are noise, low contrast, blur, and thresholding. The key tool is the structuring element, a small binary
value. The image origin is top-left (0,0). A neighborhood means aliasing. Aliasing happens when an image is sampled too coarsely mask with a chosen origin. Its size controls how strong the effect is, and
surrounding pixels used in later filtering operations. during resizing; repetitive high-frequency patterns can create false Moire its shape controls which directions and structures are affected. Larger
patterns. Anti-aliasing means smoothing before downsampling. structuring elements cause stronger changes; if the element is larger than
Binary image has only 0 and 1 (black/white), useful for object a thin object or hole, that structure may disappear.
separation. Grayscale image usually uses 8-bit intensity, so values range
from 0–255 (256 levels, starting at 0). RGB image has three channels: Dilation adds pixels to object boundaries, so white objects grow, gaps
red, green, blue; each pixel stores three values. Equal RGB values shrink, and nearby parts may connect. It is useful when objects are
produce grayscale. broken or too thin. Erosion removes pixels from boundaries, so white
objects shrink, thin white lines may disappear, and small white objects Canny: Smooth image with Gaussian → compute gradient (Sobel) →
Intensity histogram shows how pixel intensities are distributed: x-axis can be removed. It is useful for separating touching objects and non-maximum suppression keeps only thin strongest edges → double
= intensity, y-axis = number of pixels. It is used to understand brightness removing thin protrusions. Erosion on a binary image can remove white threshold separates strong/weak edges → edge tracking keeps weak
distribution and choose thresholds. If two peaks appear, foreground and Point operations change each pixel independently, so they do not use lines. Dilation is commutative and associative; erosion is not edges only if connected to strong ones. Result: clean 1-pixel edges.
background may be separable. neighbourhood information. Gamma transformation uses s = c r^γ on commutative, not associative, and erosion is not the inverse of dilation.
intensities scaled between 0 and 1. If γ > 1, the image becomes darker; if
Thresholding converts grayscale to binary by choosing one threshold T: γ < 1, it becomes brighter. Always normalize to 0-1 before applying this Opening means erosion followed by dilation with the same structuring
pixels above T become white, below T become black (or reversed). Otsu transformation. Log transformation brightens darker intensities and element. It removes small white objects and helps separate connected
thresholding automatically chooses T by maximizing separation between compresses brighter ones. Contrast stretching rescales intensities to components. It is useful when small bright noise should disappear while Watershed Algorithm
foreground and background classes. It works best when histogram has use a wider range, improving contrast while roughly keeping histogram larger objects stay. Closing means dilation followed by erosion. It fills
two clear clusters. Adaptive thresholding is used when illumination shape. It is sensitive to outliers such as salt-and-pepper pixels. small holes, bridges narrow gaps, and connects disconnected object
varies: the image is divided into small blocks, and each block gets its exposure.adjust_gamma(img,2,1) parts. In exam wording, binary closing is used to fill small holes and
own local threshold. This handles uneven backgrounds better. exposure.adjust_log(img, 1) connect broken parts. If the structuring element becomes larger, opening
removes more small bright structures and closing fills larger holes or
CMYK is a subtractive color model mainly for printing: cyan, magenta, Histogram equalization redistributes intensities so the histogram gaps.
yellow, black. HSV color model separates color meaning: Hue = true becomes more spread/flat; it is less sensitive to outliers. Adaptive
color, Saturation = purity of color, Value = brightness. This is useful histogram equalization does this locally, so it is useful when brightness Boundary extraction uses erosion: boundary = A - (A erosion B). You
because hue stays stable even if lighting changes, so segmentation by varies across the image. Intensity histogram plots intensity on the x- erode the object, then subtract the eroded version from the original,
color becomes easier than in RGB. HSI separates color into hue, axis and number of pixels on the y-axis. A histogram concentrated on the leaving only the inner boundary. This is useful when you want outlines
saturation, and average intensity of RGB ((R+G+B)/3), while HSL uses high side means a bright image. A narrow histogram means low contrast. instead of full objects. Connected components labels each separate
hue, saturation, and lightness, where lightness reflects the midpoint Histograms help decide whether contrast enhancement is possible, but binary object with a unique label. It is useful for counting objects such
between the darkest and brightest channel rather than the average they do not tell you where pixels are located in rows or columns. as cells or blobs after segmentation.
brightness (max + min/2). HSV are 0-1 where RGB values are 0-255.
Hit-and-miss transform detects specific binary patterns, often using
structuring elements with required 0s and 1s, for example corners.
Thresholding converts grayscale to binary using a threshold T: pixels
Image formation: a detector grid measures incoming energy and above T become one class and the rest the other. Increasing T usually
converts it into pixel values. Bright regions produce larger values. makes fewer pixels stay bright/foreground. Otsu’s method chooses a Thinning repeatedly removes boundary pixels while preserving
Problems during acquisition: noise = random unwanted variation (grainy global threshold that maximizes separation between foreground and connectivity, producing a thinner version of the object. Skeletonization
texture), saturation = too many pixels at maximum brightness, losing background by minimizing intra-class variance and maximizing inter- reduces an object to its centerline structure and is linked to the distance
detail. class variance. Adaptive thresholding computes a local threshold per transform, where each foreground pixel stores distance to the nearest
neighbourhood, so it works better when background intensity is uneven. background pixel.
Intensity resolution = number of gray levels determined by bit depth: Increasing block size makes the threshold more global; smaller blocks
2-bit, 4-bit, 8-bit, 16-bit. Higher bit depth gives smoother intensity make it more local and sensitive to local variation.
differences. Spatial resolution = number of pixels per area; more pixels binary_global = image > global_thresh
means sharper detail. 8-bit is 256 values because 28 = 256. Spatial
resolution is density of pixels. Greater spatial resolution = more pixels Convolution requires 3 steps: 1. position the mask over the current
are used to display images. pixel. 2. form all products of filter elements with the corresponding
elements of the neighborhood. 3. add up all the products. Stride is how
Lookup table (LUT) / colormap changes display only, not pixel values. many steps it takes. Local maxima
in the distance Noise in the gradient magnitude image can result in oversegmentation
Same grayscale image can appear in gray, summer, or pseudocolor. How to handle oversegmentation
Useful for visualization, especially in medical images. Contrast Spatial filtering uses a kernel, mask, or filter, meaning a neighbourhood transform define
stretching also changes display range to improve visibility without around each pixel. Average filtering smooths Gaussian noise but blurs the skeleton, and
changing structure. edges. A larger kernel causes stronger smoothing and can erase thin local width can
structures. Median filtering replaces a pixel by the median of its be estimated
neighbourhood, so it removes salt-and-pepper noise better and preserves from this
edges better than averaging. Laplacian filtering highlights rapid distance.
intensity changes, so it is used for sharpening. To sharpen, convolve
with a Laplacian filter and subtract the result from the original image. For grayscale morphology, you do not threshold to binary first.
Grayscale dilation replaces each pixel by the maximum in its
Uint8, values are limited to 0–255, so arithmetic can overflow and wrap neighbourhood, so bright regions expand and the image becomes
around, for example 250 + 10 becomes 4. Always check image datatype brighter overall. Grayscale erosion replaces each pixel by the minimum
int8 has range -127 to 127. uint8 has range 0 to 255 (u = unsigned). before applying transformations. in its neighbourhood, so dark regions expand and the image becomes
250 + 10 in uint8 = 4 because it wraps around and also counts 0. Distance transform: Computes a distance map or a distance field. For
darker overall. Grayscale opening removes light spots smaller than the every non-zero pixel it calculates the distance to the nearest zero pixel
structuring element.
Fundamental workflow: preprocessing (contrast, denoising, sharpening) Watershed: Treat gradient image as a topographic surface → local
→ segmentation → recognition → quantification. Preprocessing minima become basins → flood from minima upward → when basins
improves later analysis; segmentation isolates objects; quantification meet, build dams → dams become object boundaries. To avoid over-
measures size, shape, or intensity. segmentation, first limit minima (markers). Good for separating
touching objects.
Digital image = a matrix of discrete integer values where each pixel Image enhancement means preprocessing an image to improve Morphology is image processing based on shape and set theory. It is Edges → Discontinuities or sudden change in image intensity
stores intensity or color. A pixel is a picture element, and its position is technical quality so later analysis such as segmentation or recognition mainly applied to binary images, where pixels are 0 or 1 after Canny
written as f(x,y), where x and y are coordinates and f is the intensity works better. Main degradations are noise, low contrast, blur, and thresholding. The key tool is the structuring element, a small binary
value. The image origin is top-left (0,0). A neighborhood means aliasing. Aliasing happens when an image is sampled too coarsely mask with a chosen origin. Its size controls how strong the effect is, and
surrounding pixels used in later filtering operations. during resizing; repetitive high-frequency patterns can create false Moire its shape controls which directions and structures are affected. Larger
patterns. Anti-aliasing means smoothing before downsampling. structuring elements cause stronger changes; if the element is larger than
Binary image has only 0 and 1 (black/white), useful for object a thin object or hole, that structure may disappear.
separation. Grayscale image usually uses 8-bit intensity, so values range
from 0–255 (256 levels, starting at 0). RGB image has three channels: Dilation adds pixels to object boundaries, so white objects grow, gaps
red, green, blue; each pixel stores three values. Equal RGB values shrink, and nearby parts may connect. It is useful when objects are
produce grayscale. broken or too thin. Erosion removes pixels from boundaries, so white
objects shrink, thin white lines may disappear, and small white objects Canny: Smooth image with Gaussian → compute gradient (Sobel) →
Intensity histogram shows how pixel intensities are distributed: x-axis can be removed. It is useful for separating touching objects and non-maximum suppression keeps only thin strongest edges → double
= intensity, y-axis = number of pixels. It is used to understand brightness removing thin protrusions. Erosion on a binary image can remove white threshold separates strong/weak edges → edge tracking keeps weak
distribution and choose thresholds. If two peaks appear, foreground and Point operations change each pixel independently, so they do not use lines. Dilation is commutative and associative; erosion is not edges only if connected to strong ones. Result: clean 1-pixel edges.
background may be separable. neighbourhood information. Gamma transformation uses s = c r^γ on commutative, not associative, and erosion is not the inverse of dilation.
intensities scaled between 0 and 1. If γ > 1, the image becomes darker; if
Thresholding converts grayscale to binary by choosing one threshold T: γ < 1, it becomes brighter. Always normalize to 0-1 before applying this Opening means erosion followed by dilation with the same structuring
pixels above T become white, below T become black (or reversed). Otsu transformation. Log transformation brightens darker intensities and element. It removes small white objects and helps separate connected
thresholding automatically chooses T by maximizing separation between compresses brighter ones. Contrast stretching rescales intensities to components. It is useful when small bright noise should disappear while Watershed Algorithm
foreground and background classes. It works best when histogram has use a wider range, improving contrast while roughly keeping histogram larger objects stay. Closing means dilation followed by erosion. It fills
two clear clusters. Adaptive thresholding is used when illumination shape. It is sensitive to outliers such as salt-and-pepper pixels. small holes, bridges narrow gaps, and connects disconnected object
varies: the image is divided into small blocks, and each block gets its exposure.adjust_gamma(img,2,1) parts. In exam wording, binary closing is used to fill small holes and
own local threshold. This handles uneven backgrounds better. exposure.adjust_log(img, 1) connect broken parts. If the structuring element becomes larger, opening
removes more small bright structures and closing fills larger holes or
CMYK is a subtractive color model mainly for printing: cyan, magenta, Histogram equalization redistributes intensities so the histogram gaps.
yellow, black. HSV color model separates color meaning: Hue = true becomes more spread/flat; it is less sensitive to outliers. Adaptive
color, Saturation = purity of color, Value = brightness. This is useful histogram equalization does this locally, so it is useful when brightness Boundary extraction uses erosion: boundary = A - (A erosion B). You
because hue stays stable even if lighting changes, so segmentation by varies across the image. Intensity histogram plots intensity on the x- erode the object, then subtract the eroded version from the original,
color becomes easier than in RGB. HSI separates color into hue, axis and number of pixels on the y-axis. A histogram concentrated on the leaving only the inner boundary. This is useful when you want outlines
saturation, and average intensity of RGB ((R+G+B)/3), while HSL uses high side means a bright image. A narrow histogram means low contrast. instead of full objects. Connected components labels each separate
hue, saturation, and lightness, where lightness reflects the midpoint Histograms help decide whether contrast enhancement is possible, but binary object with a unique label. It is useful for counting objects such
between the darkest and brightest channel rather than the average they do not tell you where pixels are located in rows or columns. as cells or blobs after segmentation.
brightness (max + min/2). HSV are 0-1 where RGB values are 0-255.
Hit-and-miss transform detects specific binary patterns, often using
structuring elements with required 0s and 1s, for example corners.
Thresholding converts grayscale to binary using a threshold T: pixels
Image formation: a detector grid measures incoming energy and above T become one class and the rest the other. Increasing T usually
converts it into pixel values. Bright regions produce larger values. makes fewer pixels stay bright/foreground. Otsu’s method chooses a Thinning repeatedly removes boundary pixels while preserving
Problems during acquisition: noise = random unwanted variation (grainy global threshold that maximizes separation between foreground and connectivity, producing a thinner version of the object. Skeletonization
texture), saturation = too many pixels at maximum brightness, losing background by minimizing intra-class variance and maximizing inter- reduces an object to its centerline structure and is linked to the distance
detail. class variance. Adaptive thresholding computes a local threshold per transform, where each foreground pixel stores distance to the nearest
neighbourhood, so it works better when background intensity is uneven. background pixel.
Intensity resolution = number of gray levels determined by bit depth: Increasing block size makes the threshold more global; smaller blocks
2-bit, 4-bit, 8-bit, 16-bit. Higher bit depth gives smoother intensity make it more local and sensitive to local variation.
differences. Spatial resolution = number of pixels per area; more pixels binary_global = image > global_thresh
means sharper detail. 8-bit is 256 values because 28 = 256. Spatial
resolution is density of pixels. Greater spatial resolution = more pixels Convolution requires 3 steps: 1. position the mask over the current
are used to display images. pixel. 2. form all products of filter elements with the corresponding
elements of the neighborhood. 3. add up all the products. Stride is how
Lookup table (LUT) / colormap changes display only, not pixel values. many steps it takes. Local maxima
in the distance Noise in the gradient magnitude image can result in oversegmentation
Same grayscale image can appear in gray, summer, or pseudocolor. How to handle oversegmentation
Useful for visualization, especially in medical images. Contrast Spatial filtering uses a kernel, mask, or filter, meaning a neighbourhood transform define
stretching also changes display range to improve visibility without around each pixel. Average filtering smooths Gaussian noise but blurs the skeleton, and
changing structure. edges. A larger kernel causes stronger smoothing and can erase thin local width can
structures. Median filtering replaces a pixel by the median of its be estimated
neighbourhood, so it removes salt-and-pepper noise better and preserves from this
edges better than averaging. Laplacian filtering highlights rapid distance.
intensity changes, so it is used for sharpening. To sharpen, convolve
with a Laplacian filter and subtract the result from the original image. For grayscale morphology, you do not threshold to binary first.
Grayscale dilation replaces each pixel by the maximum in its
Uint8, values are limited to 0–255, so arithmetic can overflow and wrap neighbourhood, so bright regions expand and the image becomes
around, for example 250 + 10 becomes 4. Always check image datatype brighter overall. Grayscale erosion replaces each pixel by the minimum
int8 has range -127 to 127. uint8 has range 0 to 255 (u = unsigned). before applying transformations. in its neighbourhood, so dark regions expand and the image becomes
250 + 10 in uint8 = 4 because it wraps around and also counts 0. Distance transform: Computes a distance map or a distance field. For
darker overall. Grayscale opening removes light spots smaller than the every non-zero pixel it calculates the distance to the nearest zero pixel
structuring element.
Fundamental workflow: preprocessing (contrast, denoising, sharpening) Watershed: Treat gradient image as a topographic surface → local
→ segmentation → recognition → quantification. Preprocessing minima become basins → flood from minima upward → when basins
improves later analysis; segmentation isolates objects; quantification meet, build dams → dams become object boundaries. To avoid over-
measures size, shape, or intensity. segmentation, first limit minima (markers). Good for separating
touching objects.
