Skip to main content

Computer Vision

Updated May 04, 2023 ·

Overview​

Computer vision aims to enable computers to see and interpret digital images, playing a crucial role in technologies like self-driving cars.

  • Help computers understand digital images
  • Essential for self-driving cars
  • Used by manufacturers like Tesla, Volvo, Audi, and BMW
  • Utilizes multiple cameras to detect objects, lane markings, and traffic signs

Image Data​

Understanding how image data is structured is key to grasping how computer vision works. Below is an example of a grayscale and a colored image data and the pixels that makes up the each image.

What an image data looks like:

  • Images are made up of pixels containing color and intensity information
  • Grayscale images have pixel intensities between 0 and 255

Similarly, we can have

  • Colored images use the RGB system (Red, Green, Blue)
  • Requires three rasters for each color channel
  • Digital images can be represented as numerical data for machine learning models

Face Recognition​

Building a face recognition system involves processing images to identify individuals.

  • Input: Pictures of people, such as instructors
  • Neural network processes pixel intensities
  • Early neurons detect edges
  • Later neurons identify parts of objects (eyes, noses)
  • Final neurons recognize face shapes
  • Network outputs the identity of the person in the image

Diagram:

Another example is classifying image of vehicles as cars or trucks. The process will look like this:

  • The car images are turned into numbers
  • The pixel intensities are fed into the neural network
  • Neurons will learn to detect edges
  • Neurons will learn to distinguish more complex objects like wheels, doors, and windows
  • Neurons will learn to detect shapes of vehicles
  • Finally, the image will be classified as either a car or a truck

Training the Neural Network​

Training involves feeding the network large amounts of labeled data to learn patterns.

  • Provide images of faces (features) and corresponding identities (labels)
  • Learning algorithm figures out neuron computations during training
  • Middle layer neurons' functions are determined by the algorithm, reducing the need for manual intervention

Applications​

Computer vision has a wide range of applications, from recognizing objects to generating realistic images.

  • Facial Recognition: Identifies people from images
  • Self-Driving Vehicles: Detects objects and navigates safely
  • Medical Imaging: Automatically detects tumors in CT scans
  • Image Generation: Creates realistic images, like deep fakes
    • Deep fakes can generate new faces by understanding the structure of human faces