A two-stage pipeline: Haar Cascade classifiers handle the human-face / dog detection step (the same Viola-Jones approach OpenCV ships out of the box), and any dog images then go through a Convolutional Neural Network trained via transfer learning to identify the breed.
Transfer learning is the practical trick here: rather than train a CNN from scratch on a dataset that’s far too small for it, we take a network that was already pre-trained on ImageNet (millions of generic images), freeze its early feature extractors, and only fine-tune the top classifier layer on dog breeds. Fast to train, surprisingly accurate, and the dominant pattern for image classification before vision transformers showed up.
