~/Computer Vision

Introduction — Play the Emoji Scavenger Hunt

🔹 A Quick Overview of Computer Vision

Computer Vision is a system that can process, analyze and make sense of visual data the same way humans do.

5.1 Computer Vision vs Image Processing

5.2 Applications of Computer Vision

Computer Vision was first introduced in the 1970s. Today it is widely used across industries:

5.3 Computer Vision Tasks

Every CV application is built on a set of core tasks performed on input images:

5.4 Basics of Images — Pixels

5.5 Resolution

🔹 Two Ways to Express Resolution

5.6 Pixel Value

🔹 Why the 0-255 Range?

• Computer data = binary system (0s and 1s).
• Each pixel uses 1 byte = 8 bits.
• Each bit has 2 possible values (0 or 1).
• 8 bits → 2⁸ = 256 possibilities (0 to 255).
• 0 = no colour / black. 255 = full colour / white.

5.7 Grayscale Images

• Darkest shade = Black (pixel value 0).
• Lightest shade = White (pixel value 255).
• Intermediate shades = equal brightness of the three primary colours.
• Each pixel = 1 byte (8 bits). Image = single plane / 2D array of pixels.
• Size of grayscale image = Height × Width (single-plane).

5.8 RGB Images

🔹 How RGB Images Are Stored

Every RGB image is stored in three separate channels:

Each pixel has 3 values (one per channel). All three channels combined form a colour image. Size of RGB image = Height × Width × 3.

5.9 No-Code AI Tools for CV

🎨 1. Lobe

• Lobe.ai is an AutoML tool — a no-code AI tool.
• Works with image classification.
• Provide a set of labelled images → Lobe automatically finds the most optimal model to classify them.

🎯 2. Teachable Machine

• AI, ML and DL tool developed by Google in 2017.
• Runs on tensorflow.js.
• Web-based tool to train a model using images, audio, or poses — input via webcam or pictures.

🟠 3. Orange Data Mining

Has dedicated widgets for image classification — covered in Unit 4.

🔹 Use-Case Walkthrough — Coral Bleaching Detection

5.10 Image Features

🔹 What Makes a Good Feature?

Imagine a security camera capturing an image. Three types of patches we might try to find:

5.11 Convolution

An image convolution = element-wise multiplication of image arrays and another array called the kernel, followed by sum.

The kernel is passed over the whole image (slid across) to get the resulting array after convolution.

5.12 What is a Kernel?

🔹 Why Use Convolution in CNN?

• To extract features from images — edges, corners, patterns.
• Convolution is used in the Convolutional Neural Network (CNN) for feature extraction (next section).
• The center of the kernel overlaps each pixel; output becomes smaller (because edges can't be fully convolved).
• To keep the same size, we extend edges with zero padding.

5.13 Convolutional Neural Network (CNN)

🔹 CNN Workflow

5.14 Layers of CNN

🔲 1. Convolution Layer

• The first layer of a CNN.
• Objective: extract high-level features such as edges from the input image.
• The first Convolution Layer captures low-level features (edges, colour, gradient orientation).
• Added layers capture high-level features — giving the network a wholesome understanding of images.
• Uses the convolution operation with several kernels to produce several features.
• Output = Feature Map (also called Activation Map).
• Benefits of Feature Map:
  • Reduces image size for efficient processing.
  • Focuses only on features important for further processing — e.g., eyes, nose, mouth are enough to recognise a person; you don't need the whole face.

📈 2. Rectified Linear Unit (ReLU) Layer

• After the feature map is generated, it is passed to the ReLU layer.
• What it does: simply gets rid of all negative numbers in the feature map; lets positive numbers stay as-is.
• This introduces non-linearity in the feature map.
• Why? It makes colour change more abrupt and obvious — sharper edges.
• Smooth grey gradient → more abrupt edges → better features for later layers → stronger CNN.

🎯 3. Pooling Layer

• Similar to convolution layer — responsible for reducing the spatial size of the convolved feature while retaining important features.
• Two types of pooling:

🔹 Why Pooling?

• Makes the image smaller and more manageable.
• Makes the image resistant to small transformations, distortions, translations — a small difference in input creates a very similar pooled image.

🔗 4. Fully Connected (FC) Layer

• The final layer of the CNN.
• Takes results of convolution + pooling and uses them to classify the image into a label.
• Output of conv/pooling is flattened into a single vector of values, each representing a probability that a feature belongs to a certain label.
• Example: for an image of a cat, features like whiskers or fur should have high probability for the label "cat".

5.15 CNN Summary — Putting It All Together

5.16 Python Libraries for Computer Vision

Three major Python libraries are used to build CV projects:

🔹 Applications of OpenCV

• Face detection in photos and videos.
• Object tracking in live video streams.
• Motion detection in security footage.
• Image editing — rescaling, rotating, cropping, colour changes.
• OCR (Optical Character Recognition) — read text from images.
• Augmented reality apps.
• Self-driving cars — detect lanes, traffic signs, vehicles.
• Medical imaging — analyse X-rays, MRIs, CT scans.

🔹 Sample Python CV Code (OpenCV)

import cv2
img = cv2.imread('photo.jpg')
print("Shape:", img.shape)
print("Data type:", img.dtype)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
cv2.imshow('Original', img)
cv2.imshow('Grayscale', gray)
cv2.waitKey(0)
cv2.destroyAllWindows()

5.17 Practical Programs for CV (Class X)

As per the CBSE Class X syllabus (Practical list), you should be able to write these programs:

• Read an image and display using Python (OpenCV imread + imshow).
• Read an image and identify its shape using Python (img.shape gives Height × Width × Channels).

import cv2
img = cv2.imread('photo.jpg')
cv2.imshow('My Image', img)
cv2.waitKey(0)
cv2.destroyAllWindows()

import cv2
img = cv2.imread('photo.jpg')
print("Shape of image:", img.shape)
print("Height:", img.shape[0])
print("Width:", img.shape[1])
print("Channels:", img.shape[2])