Computer Vision and Image Processing
What this domain tests
Computer Vision and Image Processing tests how digital images are represented, how features are extracted, how images are classified, and where vision is applied. Small convolution or pixel questions are likely.
Core rules and concepts
- A digital image is a grid of pixels. Grayscale stores intensity, often 0-255; colour uses red, green and blue channels. Resolution is width x height.
- Feature extraction finds useful patterns: edges, colour histograms, texture, HOG or SIFT. Convolution applies a kernel that slides over the image; edge kernels respond to sharp intensity changes. CNNs learn features instead of hand-coding them.
- A classification pipeline resizes, normalises, augments training images, trains a model, and evaluates on a test set.
- Applications include medical imaging, OCR, face recognition, quality inspection and self-driving perception.
Worked example
A 3x3 grayscale patch has left column 10 and right column 80. Apply the vertical-edge kernel [[-1,0,1],[-1,0,1],[-1,0,1]]. Each row gives -10 + 0 + 80 = 70, so the sum is 210. The large response signals a strong vertical edge. Reasoning: the kernel subtracts the left side from the right side, so a flat region would give 0.
Traps that cost marks
Normalise pixel values consistently across training and test data. Augment the training set, not the test set. High accuracy can mislead on imbalanced image classes. Convolution and pooling are different: convolution learns filters, pooling reduces spatial size. Do not confuse image classification with object detection, which also localises objects.