AI & Machine Learning Roadmap

Master data processing, modeling, and deep learning algorithms.

Overall Progress0%

Computer Vision

Explanation

Computer vision is the field of enabling machines to interpret visual information. It starts from a simple truth: to a computer, an image is just a grid of numbers. Everything — classification, detection, segmentation — is built on learning patterns in that grid. This lesson covers how images are represented numerically, the three core vision tasks, and the convolutional neural network (CNN), whose operation you will work through by hand. You will also learn about transfer learning, the challenges of real-world images, and the ethical considerations of vision systems. The convolution TRY_IT is a tiny, self-contained demonstration using nested Python lists — not a complete CNN.

Images as Numerical Data

An image is a grid of pixels.
Width and height give the grid dimensions.
Channels describe color: 3 for RGB, 1 for grayscale.
Each pixel value is an intensity (e.g., 0–255).

A color image is therefore a 3D array: height × width × 3. Conceptually this is a tensor of numbers.

RGB and Grayscale

RGB stores a red, green, and blue intensity per pixel.
Grayscale stores a single intensity per pixel.
Mixing the three channels reproduces a color.

The Three Core Vision Tasks

Task	Question	Output
Image classification	What is in this image?	A single label
Object detection	What objects are present and where?	Bounding boxes + labels
Image segmentation	Which pixels belong to which object?	Pixel-level labels

Traditional vs Deep-Learning Vision

Traditional vision used hand-crafted features (edges, corners, SIFT).
Deep learning learns features automatically from data.
CNNs dominate modern vision because they learn hierarchical features.

Convolution: The Core Idea

Convolution slides a small filter (kernel) over the image, multiplying and summing at each position to produce a feature map. It detects local patterns — edges, corners, textures — wherever they appear.

Kernel/filter: a small matrix of learned weights (e.g., 3×3).
Stride: how many pixels the kernel moves each step.
Padding: extra border pixels added so the output keeps a desired size.

Convolution Worked Example

Slide a 2×2 kernel over a 3×3 image with stride 1 and no padding. The output is 2×2.

text

1Image:      Kernel:      21 2 0       1  033 1 2       0 -140 1 35 6Top-left window (1,2,3,1):71*1 + 2*0 + 3*0 + 1*(-1) = 0

The kernel slides right and down, producing one output value per position. The result is a feature map that highlights where the pattern matched.

Stride and Padding

Stride 1: the kernel moves one pixel at a time (finer coverage).
Stride 2: it skips a pixel (smaller output, faster).
Padding adds zeros around the border so the output size can match the input.

Feature Maps and Pooling

Each filter produces one feature map showing where its pattern appears.
Pooling downsamples a feature map to reduce size and add robustness.
Max pooling takes the largest value in each region (keeps the strongest response).

CNN Architecture

Image

↓

Convolutional layers

↓

Pooling

↓

Feature maps

↓

Fully-connected layers

↓

Prediction

Early convolutional layers detect low-level patterns (edges); deeper layers compose them into objects. The final layers classify.

Hierarchical Feature Learning

Early layers: edges, colors, textures.
Middle layers: shapes, parts.
Later layers: whole objects and faces.
This hierarchy is why CNNs generalize so well.

Training a CNN (Concept)

The network learns both the filter weights and the classifier.
Backpropagation updates filters just like any other weights.
Data augmentation (flips, rotations) increases effective data and reduces overfitting.

Transfer Learning

Reuse a network pre-trained on a large dataset (e.g., ImageNet).
Freeze early layers and fine-tune the top for your task.
This is how most real projects get strong results with little data.

Vision Challenges

Lighting changes, viewpoint changes, and occlusion hide objects.
Scale and noise degrade performance.
Class imbalance and label errors corrupt training.

Privacy and Ethics

Facial recognition raises privacy and surveillance concerns.
Vision models can inherit bias from training data.
Responsible deployment requires consent, fairness checks, and limits on misuse.

Common Mistakes

Confusing classification, detection, and segmentation.
Misunderstanding convolution as a one-step operation.
Skipping image preprocessing (resizing, normalization).
Assuming a CNN automatically solves every vision problem.

Examples

A 3×3 image as numbers

A grayscale image is a grid of intensities.

text

10   255 1282255 128 03128 0   255

Output

(255 = bright, 0 = dark; the pattern forms a shape)

An edge-detection kernel

A simple kernel highlights vertical edges.

text

1Kernel:  [ -1  0  1 ]2         [ -1  0  1 ]3         [ -1  0  1 ]4Effect: strong response where brightness changes horizontally

Output

(a vertical-edge detector)

Key Takeaways

An image is a grid of numbers: height × width × channels.
Classification labels the image; detection adds boxes; segmentation labels pixels.
Convolution slides a kernel and multiplies-and-sums to build feature maps.
CNNs learn a hierarchy of features from edges to objects.
Transfer learning reuses pre-trained features for small datasets.
Vision systems carry privacy, bias, and fairness responsibilities.

← PreviousDeep Learning Basics Next →Natural Language Processing