Splitting an image into meaningful regions without a neural network: Otsu thresholding, connected components, the watershed, mean shift, SLIC superpixels, graph cuts, and the Dice and Jaccard scores that judge them.
Comp9517
Why linear classifiers and MLPs struggle on images, how a convolutional layer fixes that, the size and parameter arithmetic, pooling, receptive fields, residual connections and the architecture line from LeNet to ViT.
Mathematical morphology as a repair kit for binary masks, from erosion and dilation to reconstruction, the distance transform and the top-hat, ending in a pipeline that counts nuts and bolts.
What moved, and where it will be next: change detection, brightness constancy and the aperture problem, Harris corners and Lucas--Kanade flow, then tracking as Bayesian filtering with the Kalman and particle filters.
Finding what is in an image and where: IoU and non-maximum suppression, region proposals, the R-CNN line from R-CNN to Faster R-CNN, one-stage detectors, focal loss with feature pyramids, and how detectors are scored.
Images as sums of waves: the convolution theorem, what a real spectrum looks like, deleting a periodic artefact with a notch, the difference of Gaussians as a band-pass filter, and Gaussian and Laplacian pyramids.
Labelling every pixel: fully convolutional networks, upsampling layers and their output-size trap, U-Net's skip connections, and Mask R-CNN with RoI Align.
Finding the same points in two photographs and gluing the photographs together: scale-space extrema, the 128-dimensional SIFT descriptor, the ratio test, and a homography fitted by RANSAC.