Image Classification
The fundamental computer vision task of assigning a predefined category label to an entire input image, which drove the deep learning revolution and underpins many downstream vision tasks.
Image classification assigns one or more predefined category labels to an input image. It is the most fundamental computer vision problem, and AlexNet's decisive victory in the 2012 ImageNet challenge ignited the modern deep learning era.
Today, architectures like ResNet, EfficientNet, and Vision Transformer (ViT) achieve superhuman accuracy on ImageNet. Feature representations learned through classification serve as powerful backbones for downstream tasks including detection and segmentation.
- Single-label classification: Assigns exactly one class per image. A softmax function produces a probability distribution, and the highest-probability class becomes the prediction
- Multi-label classification: Assigns multiple labels simultaneously. Independent sigmoid activations per class with a confidence threshold determine predictions
- Zero-shot classification: Models like CLIP classify images into categories never seen during training by leveraging a shared text-image embedding space, enabling open-vocabulary recognition
Standard metrics include top-1 and top-5 accuracy. In production, confusion matrices, precision, recall, and F1 scores provide per-class analysis. Class imbalance is addressed through weighted loss functions or focal loss. Knowledge distillation compresses large models for efficient edge deployment.