This lecture introduces how neural networks are adapted to visual data. We begin with a simple question: What does a computer actually receive when it sees an image? We then explore why ordinary fully connected neural networks are not a natural fit for images, how convolution exploits spatial structure, and how CNNs build increasingly useful visual representations.
The lecture then connects CNNs to three increasingly informative computer-vision tasks: classification, object detection, and image segmentation. The emphasis is on intuition, visual examples, prediction, discussion, and short active-learning activities rather than mathematical derivation.
Learning Objectives
After this lecture, you should be able to:
Describe an image as numerical data that a neural network can process.
Explain why spatial structure matters when processing images.
Explain intuitively what a convolutional filter does.
Interpret a feature map as the response of a learned visual filter.
Describe the basic flow of a convolutional neural network (CNN).
Distinguish image classification, object detection, and image segmentation.
Explain the basic idea of an encoder–decoder and why U-Nets use skip connections.