Computer Vision: Color Spaces, Image Kernels, and Filtering

Color Representation: RGB and HSV Color Spaces (Checkpoint 1 Task 1)

• HSV vs. RGB in Computer Vision Applications:

  • RGB Color Space: Primarily designed and utilized for hardware-oriented display systems and digital image storage.

  • HSV Color Space: Optimized for color manipulation, computer vision processing, and image analysis tasks such as object detection.

  • Decoupling of Chromatic and Achromatic Information: HSV separates pure chromatic content (Hue) from intensity and luminance attributes (Saturation and Value). This separation makes HSV significantly more intuitive than RGB when segmenting or tracking objects under varying lighting conditions.

• Components of the HSV Color Model:

  • Hue ($H$): Identifies the fundamental color type or wavelength. It is modeled geometrically as an angle ranging along a circular color wheel from $0^\bot$ to $360^\bot$.

  • Saturation ($S$): Dictates the purity or intensity of the color, ranging continuously from $0$ to $1$. A saturation value of $0$ represents a completely desaturated grayscale tone, whereas a value of $1$ represents fully saturated, pure color.

  • Value ($V$): Defines the brightness or luminance level of the color, scaling from $0$ to $1$. A value of $0$ corresponds to absolute darkness (black), whereas $1$ indicates full, maximum brightness.

• RGB Value Normalization:

  • Prior to running the rgb_to_hsv conversion algorithm, input RGB values must be normalized to the range of $[0, 1]$.

  • Normalization scales the raw channel values (typically spanning integer ranges like $[0, 255]$) into the expected mathematical domain required by HSV analytical formulas. Without normalization, the mathematical relations for computing hue, saturation, and brightness fail to produce valid results.

• Numerical Precision and Data Types:

  • The rgb_to_hsv function must return high-precision floating-point values rather than integer types such as uint8.

  • The higher-level convert_to_hsv function handles casting and scaling floats to integers when producing final uint8 arrays.

  • Casting or rounding to uint8 prematurely inside rgb_to_hsv would truncate decimal fractional components, leading to an irreversible loss of mathematical precision during intermediate calculations.

• Role of Extremum Channels and Delta:

  • Maximum ($\max$): The maximum normalized value among the red, green, and blue color channels: $\max = \max(R, G, B)$.

  • Minimum ($\min$): The minimum normalized value among the color channels: $\min = \min(R, G, B)$.

  • Delta ($\Delta$): The absolute dynamic range between the extremum channel intensities: $\Delta = \max - \min$.

  • Conditional Branching for Hue:

  • The formula used to calculate Hue depends strictly on which specific color channel matches $\max$. If the red channel is maximal, a specific mathematical equation is evaluated; distinct alternative equations are applied if the green or blue channel is maximal.

  • If $\Delta = 0$, all RGB components are identical (representing an achromatic shade of gray), requiring a separate conditional branch where Hue is defined as $0$.

• Pixel-by-Pixel Application:

  • Image-wide conversion is orchestrated by the convert_to_hsv procedure.

  • This procedure iterates across every coordinate and pixel in the image array, systematically invoking the core rgb_to_hsv helper function on each individual pixel.

• Behavior on Grayscale Image Inputs:

  • When a grayscale image is passed to the convert_to_hsv function, the red, green, and blue components are identical for every pixel ($R = G = B$).

  • Consequently, both Hue ($H$) and Saturation ($S$) evaluate strictly to $0$.

  • Only the Value ($V$) channel contains active, non-zero data corresponding to the grayscale intensity. The function still executes normally and returns complete three-channel HSV structures.

Geometric Preprocessing: Resizing, Aspect Ratio, and Padding (Checkpoint 1 Task 1)

• Need for Standardized Input Dimensions:

  • Standardizing input images to a uniform resolution ensures structural consistency across data pipelines and downstream computer vision algorithms.

• Aspect Ratio Preservation via Padding:

  • To achieve a fixed target size of $100 \times 100$ pixels without distorting or warping the original subject matter, padding is combined with resizing.

  • Padding introduces extra boundary pixels around the outer edges of the image canvas. This allows the primary image content to preserve its original aspect ratio while expanding the overall matrix dimensions to meet the uniform $100 \times 100$ standard.

Convolution and Image Kernels: Filtering Mechanics (Checkpoint 1 Task 2)

• Definition of an Image Kernel:

  • An image kernel is a compact, two-dimensional square matrix containing numerical coefficients or weights.

  • As it traverses across an image, these weights define the relative mathematical influence that neighboring pixels exert when computing the updated intensity of the central target pixel.

• Sliding-Window Operation:

  • A sliding window corresponding to the dimensions of the kernel (for example, a $3 \times 3$ subgrid) is superimposed over a local patch of the image.

  • Filtering computations are executed for that localized neighborhood.

  • The window subsequently shifts across the image grid to the next pixel coordinate by an interval determined by the stride, repeating the operation until the entire spatial domain is processed.

• Kernel Filtering Mechanism:

  • Local Window Extraction: At each pixel coordinate, the kernel extracts the raw values of the surrounding pixel neighborhood (e.g., a $3 \times 3$ window surrounding the current center pixel).

  • Element-Wise Multiplication: Every extracted pixel value is multiplied by its corresponding numerical weight located at that relative position within the kernel.

  • Accumulation: The products of all element-wise multiplications across the window are summed together into a single scalar value.

  • Value Reassignment: The accumulated scalar sum is assigned as the new intensity value for the central pixel coordinate in the output image.

• Constructing a Normalized Mean Kernel:

  • Step 1 — Count Total Matrix Elements: Determine the total number of entries in the square kernel array. For a square kernel of size $K \times K$, total elements $N = K \times K$. For example, a $3 \times 3$ kernel contains $3 \times 3 = 9$ entries.

  • Step 2 — Calculate Uniform Value: Divide $1$ by the total entry count ($\frac{1}{N}$). For a $3 \times 3$ kernel, each entry evaluates to $\frac{1}{9}$.

  • Step 3 — Matrix Population: Assign the calculated fraction to every coordinate within the kernel matrix.

  • Step 4 — Normalization Check: Verify that the sum of all individual coefficients across the entire kernel sums exactly to $1$. Normalization preserves the overall energy and brightness of the image.

Spatial Constraints: Padding, Kernel Symmetry, and Stride (Checkpoint 1 Task 2)

• Impact of Padding on Output Image Dimensions:

  • Without padding, filtering cannot be computed at the boundary margins of the image because parts of the kernel window hang over empty space beyond the image border.

  • Omitting edge handling forces the output image to shrink, as border pixels cannot be evaluated.

  • Zero-padding introduces extra border rows and columns filled with zeros ($0$), enabling kernel placement directly over peripheral pixels while maintaining the original dimensional size throughout the filtering operation.

• Calculating Required Zero-Padding:

  • The amount of zero-padding border width required to maintain dimensional consistency is determined by the kernel size $K$ via the formula:   $\text{Padding} = \frac{K - 1}{2}$

  • For instance, a $3 \times 3$ kernel ($K = 3$) requires:   $\text{Padding} = \frac{3 - 1}{2} = 1\text{ pixel of zero-padding on every side}$

• Kernel Dimension and Center Symmetry:

  • Kernel dimensions must strictly be odd integers ($3 \times 3$, $5 \times 5$, $7 \times 7$, etc.).

  • An odd dimension guarantees the presence of an unambiguous, physical center pixel.

  • This central anchor ensures an equal, symmetrical distribution of neighboring pixels on all sides (top, bottom, left, and right) during window calculation.

• Definition and Function of Stride:

  • Stride specifies the step size or coordinate displacement of the sliding window between consecutive operations.

  • A setting of $\text{stride} = 1$ means the kernel shifts by exactly $1$ pixel coordinate horizontally across columns, and by $1$ pixel vertically across rows upon reaching the line boundary.

Image Smoothing Techniques: Mean Blur vs. Gaussian Blur (Checkpoint 1 Task 3)

• Weighting Differences Between Mean and Gaussian Filters:

  • Mean Blur: Employs uniform weighting. Every single pixel in the receptive window contributes equally ($\frac{1}{N}$) to the final output average, regardless of its spatial distance from the center.

  • Gaussian Blur: Employs non-uniform, distance-weighted coefficients derived from a Gaussian bell curve. Pixels nearest to the center receive the largest weights, while pixels farther toward the periphery receive progressively smaller weights.

• Evaluation of Filtering Characteristics:

  • Mean Filter Drawbacks: Produces harsh, blocky, and unnatural visual artifacts. Because it weights outliers and edge pixels equally to center pixels, it acts aggressively, excessively degrading and washing out major structural boundaries.

  • Gaussian Filter Advantages: Provides natural, continuous smoothing without blocky distortion. It removes localized granular variations while preserving dominant structural edges.

Feature Extraction and the Role of Kernel Dispersion (Checkpoint 1 Task 3)

• Pre-Processing for Feature Extraction and Edge Detection:

  • High-frequency noise and pixel-level variance interfere with gradient calculations in feature extraction pipelines.

  • Blurring acts as a low-pass filter that suppresses minor pixel-level variations and noise before edge detection algorithms are applied.

  • Eliminating high-frequency noise allows edge detection filters to reliably capture prominent macroscopic boundaries, silhouettes, and large structural contours.

• Workflow Preference for Gaussian Smoothing:

  • Gaussian blur is preferred over mean blur in computer vision workflows because it smooths out image noise without degrading major geometric structures.

  • Mean blur is overly destructive to clean structural lines, whereas Gaussian blur provides controlled noise suppression necessary for sharp edge localization.

• Impact of the Dispersion Parameter Sigma ($\sigma$):

  • The parameter $\sigma$ (sigma) governs the standard deviation and spatial spread of the Gaussian distribution within the kernel matrix.

  • Small $\sigma$: Results in a narrow, tightly concentrated distribution peak, producing a subtle, localized, and tight blur.

  • Large $\sigma$: Flattens and widens the spatial distribution, assigning larger relative weights to distant surrounding pixels, resulting in an expansive, broader blur effect across the image.