Computer vision basics: image classification with CNNs

11 steps · about 11 h 8 min · Beginner

  1. Neural networks (guide, 3Blue1Brown, 1 h)

    Start with what neurons, layers, weights and learning mean, using digit recognition as the running example.

  2. But what is a convolution? (video, 3Blue1Brown (YouTube), 23 min)

    The sliding-window idea behind every CNN filter, shown with blurring and edge detection on images.

  3. CNN Explainer (interactive, Georgia Tech Polo Club, 30 min)

    Click through convolution, ReLU, pooling and softmax in a live network that classifies images.

  4. Creating Deep Convolutional Neural Networks for Image Classification (tutorial, The Programming Historian, 45 min)

    Train a real image classifier with no code in Teachable Machine to get a feel for labelled data and predictions.

  5. Convolutional Neural Networks (CNNs / ConvNets) (article, Stanford CS231n, 1 h)

    A rigorous but readable reference for conv layers, pooling, layer sizing and classic architectures.

  6. Convolutional Neural Network (CNN) (tutorial, TensorFlow Core, 40 min)

    Your first code: a small Keras CNN on CIFAR images in a few lines.

  7. Training a Classifier (tutorial, PyTorch Tutorials, 45 min)

    Repeat the task in PyTorch to see the full loop of data, model, loss, training and testing.

  8. Image Classification with Convolutional Neural Networks: All in One View (tutorial, The Carpentries Incubator, 2 h)

    A structured lesson on compiling, training and evaluating CNNs, with exercises that build on the basic models.

  9. Transfer Learning and Fine-tuning Convolutional Neural Networks (article, Stanford CS231n, 20 min)

    Learn why almost nobody trains from scratch and how pretrained networks are reused.

  10. Transfer Learning for Computer Vision Tutorial (tutorial, PyTorch Tutorials, 45 min)

    Apply the idea by fine-tuning a pretrained network on your own small image dataset.

  11. CS231n Deep Learning for Computer Vision (reference, Stanford CS231n, 3 h)

    Go deeper with the full course notes and assignments on visualizing networks, training tricks and beyond classification.