Image Classification Heads, Data Augmentation, and Transfer Learning

A desk that classifies corporate-logo screenshots or satellite tiles for ESG monitoring rarely owns ImageNet-scale labelled data. Fitting a deep CNN from random initialisation on a few thousand images typically overfits long before it reaches useful validation accuracy. This note closes the vision pipeline: attach a lightweight classification head to convolutional features, regularise with data augmentation, and reuse transfer learning when pretraining on a large corpus is cheaper than collecting labels.