An image is a grid of thousands of pixels, and feeding every pixel into an ordinary network ignores the fact that nearby pixels belong together. A convolutional neural network (CNN) is built around that fact. It slides small filters across the image, and the same filter is reused at every position, so a pattern such as an edge is detected wherever it appears.
Each filter produces a feature map showing where its pattern was found. Pooling then shrinks the maps and keeps the strongest responses, and stacking several layers lets early layers find edges while later ones find shapes and whole objects. A final fully connected layer turns those features into a class prediction. The filter values are learned during training, just like any other weights.
Watch the lessons in order, then continue to Computer Vision & Transfer Learning.
3Blue1Brown builds up convolution as a sliding window of pairwise products, then shows it at work on images for blurring and edge detection. The closing sections on polynomials and FFTs go beyond what CNNs need, so the image-processing part is the key section.
Video 23minLuis Serrano explains how a CNN recognizes images, starting from a tiny example: small filters slide over the picture, and gradient descent finds the filter values that work best.
Video 32minStatQuest on one of the best-known uses of neural networks: classifying images with a CNN, and why this beats feeding raw pixels into an ordinary network.
Video 15min