← Analysing an area of computer science

Computer vision

Offered 2 of the last 4 years (2022, 2025) — a strong second-choice area if paired with computer graphics, since both share the idea of an image as a grid of values.

The foundations to define precisely

  • A digital image is an array of pixel intensity values — computer vision operates on that array, not on the picture as a human sees it.
  • Noise is random variation in pixel values from sensor limitations, low light, or compression, and it matters because edge detection looks for sharp intensity change — which noise mimics.
  • Depth (distance from the camera) can be recovered from stereo vision (disparity between two cameras), structured light or time-of-flight sensors, LiDAR, or monocular depth estimation.
Worked example Walking through Canny edge detection Explain each step of the Canny edge-detection algorithm and why the order matters. GAUSSIAN BLUR first: smooths the image to reduce noise. This is counter-intuitive, since it removes detail, but noise looks exactly like a sharp intensity change and would otherwise be mistaken for edges. GRADIENT CALCULATION (usually via a Sobel operator): finds where intensity changes sharply, in both strength and direction. NON-MAXIMUM SUPPRESSION: thins the detected edges down to a single pixel wide, discarding gradient points that are not the local maximum. DOUBLE THRESHOLDING: classifies edge pixels as strong, weak, or not-an-edge based on two intensity thresholds. HYSTERESIS TRACKING: keeps a weak edge pixel only if it connects to a strong one, which stops a real edge from breaking into disconnected dashes. ✓ The blur-first ordering is the step worth explaining: removing detail before looking for edges is what prevents noise from being detected as false edges. ✕

Too much or too little blur is a genuine trade-off, not a fixed setting

A common weak answer treats blurring as simply "good" with no limit. In fact, too much blur removes real, fine edges along with the noise, while too little leaves noise that gets mistaken for edges. The kernel size is a genuine judgement call balancing these — naming this trade-off, rather than treating blur as unconditionally beneficial, is what separates analysis from description.

Practise just this