Computer Vision
Pixels, filters, feature maps and the convolutional stack that turns an image into a prediction.
About this track
A convolutional network is usually drawn as a row of coloured boxes that explains nothing. This track takes the boxes apart: what a filter is, what it produces, why the output shrinks, what padding restores, and what a stride costs you.
It starts with how an image is stored at all - grayscale intensities, then RGB channels - and finishes on the practical layer: augmentation, data loaders and transfer learning.
The stack is taken apart rather than drawn as coloured boxes: what a filter contains, what a feature map records, what pooling discards, and how many parameters each of those choices actually costs.
What you will be able to do
- Explain what a convolutional filter is and why it spans every input channel.
- Compute a layer's output size from the input, kernel, padding and stride.
- Count a convolutional layer's parameters, and say why the image size never appears.
- Describe what pooling buys and what it destroys.
- Say why weight sharing gives both a parameter saving and translation equivariance.
How the track is ordered
It starts below the network, with how an image is stored at all - grayscale intensities, then RGB channels - because a filter's depth only makes sense once channels do. Then the convolution itself: edge detection by hand, feature maps, padding, stride, pooling and parameter counting. After that the components that make a real network work, including ReLU, dense layers and residual connections, and finally the applied layer: augmentation, data loaders, transfer learning, object detection and segmentation. The deep learning track is assumed.
Where this leads
Transfer learning is where most practical vision work actually starts, and the generative AI track covers the modern architectures that now compete with convolutional networks on the same tasks.
All 48 modules, in teaching order
- 01How Neural Networks Process ImagesUnroll a 2D pixel grid into a 1D vector and see what spatial information flattening destroys — and why CNNs exist to avoid it.
- 02Grayscale Image ProcessingSee a grayscale image for what it is: a grid of brightness numbers you can filter, threshold and edit directly.
- 03RGB Image ProcessingSplit a colour image into red, green and blue channels and see how three grids of numbers combine into every pixel.
- 04Colour Spaces: RGB and HSVConvert the same pixel between RGB and HSV, then see how a colour range that needs six RGB inequalities becomes a single HSV threshold.
- 05Histograms and EqualisationPlot an image histogram, apply equalisation, and watch a low-contrast photograph spread across the full intensity range in one pass.
- 06ThresholdingTurn a greyscale image into a binary mask with global and adaptive thresholds, and see where uniform lighting stops being enough.
- 07Erosion and DilationApply erosion and dilation to a binary mask, then combine them into opening, closing, gradient and top-hat to clean segmentation output.
- 08Blur: Gaussian, Median and BilateralApply Gaussian, median and bilateral blur to one noisy image side by side and see which filter removes noise while preserving edges.
- 09Convolution Kernels by HandSlide a 3x3 kernel across an image by hand, multiply and sum at each position, and watch edge detection and blur fall out of nine numbers.
- 10Real-time Edge DetectionSlide a sensitivity threshold across a real image and watch edges appear wherever brightness changes sharply enough.
- 11Resizing and InterpolationResize an image with nearest, bilinear and bicubic interpolation side by side and see exactly where each method invents or drops detail.
- 12Affine TransformsBuild a 2x3 affine matrix from rotation, scale, shear and translation, apply it to an image, and see every pixel land in its new position.
- 13Template MatchingSlide a template across an image, compute the match score at every position, and see why changes in scale and rotation break the method.
- 14Harris Corners and KeypointsCompute the Harris corner response at every pixel, threshold the result, and see why corners survive rotation and small shifts.
- 15Haar Cascade Detection ModelsDrag a Haar rectangle feature over a face, watch the integral image sum it in four lookups, and see a cascade discard 24,000 windows.
- 16Convolutional LayerDrag a convolution filter across an image and watch it build a feature map one value at a time, the core operation inside a CNN.
- 17Receptive FieldTrace one unit's receptive field back through conv and pooling layers and see how depth, stride and dilation decide what it can observe.
- 18Padding in CNNAdd zero-padding around an image and watch the output keep its spatial dimensions instead of shrinking with every convolution layer.
- 19Strides in CNNIncrease the stride from 1 to 2 and watch the feature map shrink to a quarter of its area in a single convolution step.
- 20Pooling Layers in CNNsSlide a 2x2 or 3x3 window across a feature map, take the max or average at each position, and watch the output shrink.
- 21Parameter Sharing in CNNSlide one small filter across every spatial position and compare its 9 parameters to the millions a fully connected layer would need.
- 22ReLU Activation in CNNApply ReLU to a feature map and watch every negative value drop to zero, the single nonlinearity that makes deep networks trainable.
- 23Pooling LayerShrink a feature map with max, average or min pooling and see how downsampling keeps the signal but discards the detail.
- 24Fully Connected Layer in CNNFlatten a CNN's feature maps into a 1D vector and feed it through a dense layer that mixes spatial features into a class decision.
- 251x1 ConvolutionsSee how a 1x1 convolution mixes channels without touching spatial dimensions, and why it appears in Inception, ResNet and every bottleneck.
- 26Depthwise Separable ConvolutionSplit a standard convolution into depthwise and pointwise steps, count the multiply-adds saved, and see why mobile models rely on them.
- 27Dilated ConvolutionsExpand a kernel's field of view without adding parameters or pooling, and spot the gridding artefact that appears when rates are wrong.
- 28Global Average Pooling against FlattenReplace a flatten layer with global average pooling, count the parameters removed, and see why modern CNNs skip the large dense head.
- 29Calculating Parameters in CNNCount a CNN's trainable parameters layer by layer and see why image size never appears in the formula — only the filter shape does.
- 30CNN ArchitectureWalk through a convolutional neural network layer by layer in 3D: filters, feature maps, pooling and the classification head.
- 31Data Loaders in CNNSee how a DataLoader batches, shuffles and augments images on the fly so the GPU never sits idle waiting for the next batch.
- 32Image Data AugmentationFlip, rotate, zoom and add noise to an image, and see how augmentation multiplies a small training set into a larger one.
- 33Transfer Learning with CNNSwap the classification head of a pretrained CNN, freeze the base, and fine-tune on your dataset with a fraction of the data.
- 34VGG-16Walk VGG-16 layer by layer and see why 89% of its 138 M parameters sit in three dense layers that do 1% of the work.
- 35InceptionNetRun the four inception branches in parallel, then switch off the 1x1 bottleneck and watch one module jump from 128 M to 303 M MACs.
- 36Anchor BoxesSee how anchor boxes turn object detection into a small regression, and follow the IoU-based rule that assigns each ground truth to a box.
- 37IoU and Non-Max SuppressionA detector proposes four boxes for two objects. IoU measures the overlap, and Non-Max Suppression uses that number to throw the duplicates away.
- 38Mean Average PrecisionSort detections by confidence, compute precision and recall at each step, and watch the area under that curve collapse into one mAP number.
- 39Object Detection with Bounding BoxesDrag a predicted box toward the ground truth and watch IoU climb, the loss fall, and the detection flip from a miss to a match at the 0.5 threshold.
- 40YOLO v8See all 8,400 raw YOLO predictions, then move the confidence and NMS thresholds and watch objects appear, duplicate and disappear.
- 41ResNet and Identity ShortcutsPush a signal through twenty contractive layers with and without an identity shortcut. One collapses toward zero; the other never can.
- 42ResNetBuild ResNet-18 through -152 from one description and watch 25.56 M parameters and 4.1 GMACs fall out of the block counts.
- 43Semantic, Instance and Panoptic SegmentationCompare semantic, instance and panoptic segmentation on the same scene and see the two questions that separate each task from the others.
- 44Semantic Segmentation and U-NetDownsample a mask to see more context, then upsample it back, and watch the boundary blur unless a skip connection carries the resolution pooling just destroyed.
- 45U-NetWalk U-Net level by level: 572 in, 388 out, every skip crop computed, and the case for skip connections made by turning them off.
- 46Mask R-CNNDrag a proposal on a feature map and watch RoIPool's double rounding move it eight input pixels while RoIAlign samples it exactly.
- 47Vision Transformer PatchesSplit an image into fixed-size patches, project each into a token, and follow the sequence through a transformer to see its quadratic cost.
- 48Grad-CAMCompute a Grad-CAM heatmap from gradients flowing into the last convolutional layer, then overlay it to see which regions drove the prediction.