1.

Object Detection using Sliding Window and Region Proposal

Sliding Window Technique:

 A fixed-size window slides across the image (horizontally and vertically).

 At each location, the sub-region is passed to a classifier (e.g., SVM, CNN) to determine
whether it contains the object.

 Drawbacks: High computational cost, slow, and doesn’t handle scale changes well.

Region Proposal Methods:

 Instead of exhaustive sliding, these methods propose regions likely to contain objects.

 Selective Search: Groups similar regions based on color, texture, size.

 Edge Boxes: Generates boxes based on edge information.

 These proposals are then passed to CNNs for classification and bounding box regression.

Viola-Jones for Face Detection:

 Uses Haar-like features computed with an integral image.

 A cascade of classifiers quickly eliminates non-face regions.

 Adaboost is used to select the best features.

Deep Learning Models for Face Detection:

 MTCNN, RetinaFace, and others use CNNs to detect faces at different scales.

 Provide higher accuracy and robustness to pose, illumination, and occlusion variations.

2. Comparison of YOLO, SSD, and Faster R-CNN; Harris and Shi-Tomasi Corner Detection

Feature YOLO SSD Faster R-CNN

Type Single-shot Single-shot Two-stage

Speed Very fast Fast Slower

Accuracy Moderate-high Moderate-high High

Pipeline Unified CNN Multi-scale CNN RPN + Detection

Harris Corner Detector:

 Measures intensity change in all directions.

 Uses the second-moment matrix.

 Corner response: R=det(M)−k(trace(M))2R = det(M) - k(trace(M))^2R=det(M)−k(trace(M))2

Shi-Tomasi:

 Improves Harris by using minimum eigenvalue of the matrix MMM.

3. Hough Transform and Morphological Operations

Hough Transform:

 Used to detect lines and shapes (e.g., circles).

 Transforms each point to parameter space (e.g., lines in polar form: ρ=xcos⁡θ+ysin⁡θ\rho = x\
cos\theta + y\sin\thetaρ=xcosθ+ysinθ).

 Peaks in accumulator space indicate lines.

Morphological Operations:

 Applied to binary images.

 Based on structuring elements.

Erosion:

 Shrinks white regions.

 Removes small noise and separates objects.

Dilation:

 Expands white regions.

 Fills small holes.

Sobel Edge Detection:

 Uses gradient filters in X and Y directions.

 Highlights edges based on intensity change.

Canny Edge Detection:

 Steps: Gaussian blur → Gradient → Non-maximum suppression → Hysteresis thresholding.

 Produces clean and continuous edges.

4. Image Segmentation Techniques

Thresholding:

 Converts grayscale to binary using a threshold value.

Global Thresholding:

 Single threshold for the whole image.

Adaptive Thresholding:

 Threshold computed locally for different regions.

Region-Based Segmentation:

 Groups pixels with similar properties.

 Region growing starts from a seed and includes neighboring pixels.

Opening:

 Erosion followed by dilation.

 Removes small objects.

Closing:

 Dilation followed by erosion.

 Fills small holes.

5. SIFT, SURF, Viola-Jones, and Panorama Matching

SIFT (Scale-Invariant Feature Transform):

 Detects keypoints invariant to scale, rotation.

 Steps: Scale-space extrema → Keypoint localization → Orientation → Descriptor.

SURF (Speeded-Up Robust Features):

 Faster than SIFT using integral images and box filters.

 Less accurate but computationally efficient.

Viola-Jones:

 As explained earlier, uses Haar features, Adaboost, and cascade classifiers.

Keypoint Matching in Panorama Creation:

 Detect keypoints using SIFT/SURF.

 Match descriptors between images.

 Estimate transformation (homography).

 Warp and blend images to create panorama.

6. CNN Architectures and Harris Corner Detection

YOLO:

 Predicts bounding boxes and class probabilities directly from image in one pass.

 Very fast; used in real-time applications.

Faster R-CNN:
  Region Proposal Network (RPN) suggests regions.

 These are classified and refined by the detector head.

Harris Corner Detector (Revisited):

 Computes gradient matrix for each pixel.

 Uses eigenvalues of matrix to detect corners.

7. Fundamentals of Computer Vision

Computer Vision:

 Field enabling machines to interpret and understand visual information.

 Applications: object detection, facial recognition, autonomous vehicles, etc.

Pixels:

 Smallest element of an image.

 Each pixel stores color/intensity values.

Resolution:

 Number of pixels in width × height.

 Higher resolution = more detail.

Image Representation:

 Grayscale: one value per pixel.

 RGB: three values (Red, Green, Blue).

 Stored as 2D or 3D arrays.

Image Formation:

 Through lenses projecting scene onto a sensor.

 Pinhole model, perspective projection, and lens distortions affect the image.

Brightness: Intensity level (e.g., dark vs. bright image).

Contrast: Difference between darkest and brightest regions.

Hue: Type of color (e.g., red, blue).

Saturation: Intensity or purity of color (gray = low saturation).

✅ 1. Filtering in Image Processing

➤ Definition:
Filtering is the process of modifying or enhancing an image by emphasizing or removing certain
features like noise, edges, or textures.

➤ Types of Filters:

A. Linear Filters:

Apply a linear transformation to pixel values.

 Mean Filter (Averaging): Reduces noise by replacing each pixel with the average of its
neighbors.

 Gaussian Filter: Applies a weighted average using a Gaussian kernel. Smoothens image while
preserving edges better than mean filtering.

B. Non-linear Filters:

 Median Filter: Replaces pixel value with the median of its neighborhood. Very effective in
removing salt-and-pepper noise.

➤ Example:

An image with salt-and-pepper noise can be cleaned using a median filter, which removes outliers
(black or white dots) while preserving edges.

✅ 2. Convolution in Image Processing

➤ Definition:

Convolution is a mathematical operation used to apply filters to images.

➤ How It Works:

 A kernel (filter matrix) is slid across the image.

 At each location, the sum of the element-wise product of the kernel and the overlapping
image region is computed.

 This value replaces the central pixel.

➤ Mathematical Expression:

G(x,y)=∑i=−kk∑j=−kkI(x+i,y+j)⋅K(i,j)G(x, y) = \sum_{i=-k}^{k} \sum_{j=-k}^{k} I(x+i, y+j) \cdot K(i,

Where:

 III is the input image

 KKK is the kernel

 GGG is the output image

➤ Example:

Using a 3×3 sharpening kernel:

CopyEdit

[ 0 -1 0

-1 5 -1

0 -1 0 ]

applied via convolution enhances edges in an image.

✅ 3. Edge Detection

Edge detection helps identify object boundaries by detecting changes in intensity.

➤ A. Sobel Edge Detection

➤ How it works:

 Applies two 3×3 kernels: one for horizontal (GxG_xGx), one for vertical (GyG_yGy) gradients.

 Combined gradient magnitude:

G=Gx2+Gy2G = \sqrt{G_x^2 + G_y^2}G=Gx2+Gy2

➤ Kernels:

text

CopyEdit

Gx = [ -1 0 1 Gy = [ -1 -2 -1

-2 0 2 0 0 0

-1 0 1 ] 1 2 1]

➤ Example:

Apply Sobel to detect roads in satellite images by emphasizing edges in horizontal and vertical
directions.

➤ B. Canny Edge Detection

Canny is a multi-stage edge detection algorithm:

1. Noise Reduction: Gaussian blur

2. Gradient Calculation: Sobel-like operation

3. Non-Maximum Suppression: Thins edges

4. Double Thresholding: Strong and weak edges

5. Edge Tracking by Hysteresis: Connects weak edges to strong ones

Used in medical imaging (e.g., MRI, X-rays) to detect boundaries of tissues or bones accurately.

✅ 4. Image Transformations

➤ A. Fourier Transform (FT)

➤ Definition:

Transforms an image from spatial domain to frequency domain. Useful to analyze frequency
content.

➤ How it works:

F(u,v)=∑x∑yf(x,y)⋅e−j2π(ux/M+vy/N)F(u, v) = \sum_{x} \sum_{y} f(x, y) \cdot e^{-j2\pi(ux/M +

➤ Use Cases:

 Filtering (e.g., low-pass to remove high-frequency noise)

 Image compression (JPEG uses DCT, a related concept)

 Pattern recognition

➤ Example:

A fingerprint image with periodic noise can be denoised by applying Fourier Transform, masking
high-frequency components, and applying Inverse Fourier Transform.

➤ B. Hough Transform

➤ Definition:

Used to detect geometric shapes (lines, circles) in images.

➤ For Lines:

A line can be expressed as:

ρ=xcos⁡θ+ysin⁡θ\rho = x\cos\theta + y\sin\thetaρ=xcosθ+ysinθ

Each edge point votes in the accumulator space for possible lines passing through it.

➤ For Circles:

Circle equation: (x−a)2+(y−b)2=r2(x - a)^2 + (y - b)^2 = r^2(x−a)2+(y−b)2=r2

➤ Example:

Used in license plate detection or lane detection in autonomous driving by detecting straight lines
on the road.