Image classification, deciding what a picture shows, is the task that drove most of modern computer vision. The network that does the heavy lifting is called a backbone: it turns raw pixels into features, and later modules for detection and segmentation reuse the same backbones.
Backbones have evolved quickly. Early CNN architectures such as AlexNet and VGG showed that depth helps, and ResNet made very deep networks trainable with skip connections. The Vision Transformer then showed that an image cut into patches can be handled by a transformer instead of convolutions.
In practice you rarely train a backbone from scratch. Transfer learning starts from a model pretrained on a large dataset and fine-tunes it on your own images. Watch the lessons in order, then continue on to Object Detection.
How CNN backbones developed over time, from LeNet and AlexNet through VGG, Inception, ResNet, DenseNet, and EfficientNet.
Why adding layers can hurt performance, and how residual connections and identity mappings fix the problem.
The Vision Transformer: how an image is split into patches and processed as a sequence, and how it compares with CNNs.
Building an image classifier by fine-tuning a pretrained ResNet50 in PyTorch, with data augmentation and preprocessing.