← AI Terminology

Image Classification

Image classification is the computer vision task of assigning a single label (or probability distribution over labels) to an entire image — answering "what is the main subject of this image?"

It is the most fundamental vision task and the benchmark that launched the deep learning revolution (ImageNet, 2012).
Why It Matters in AI
AlexNet winning the 2012 ImageNet challenge using deep convolutional networks triggered the modern AI era. Image classification is the "hello world" of computer vision: master it and the techniques extend to detection, segmentation, and medical imaging. Pre-trained classifiers (ResNet, EfficientNet, ViT) are transfer learning backbones for almost every vision application.
Key Points
Aspect Description
Output Single class label (hard) or class probability vector (soft, softmax) over N classes
Evaluation Top-1 accuracy (correct class ranked first); Top-5 (correct class in top 5)
Key models AlexNet, VGG, ResNet, EfficientNet, ViT — each generation improved accuracy at same compute
Current SOTA ViT-G/14 + Inception Head achieves 90.9% top-1 on ImageNet; human baseline ~95%
Key datasets ImageNet (1.2M images, 1000 classes), CIFAR-10/100, Food-101, iNaturalist
Transfer learning Pre-train on ImageNet → fine-tune on custom classes — works with as few as 100 examples
Simple Analogy
Showing a photo to a knowledgeable friend and asking "What's this a picture of?" — they return "golden retriever." Image classification is automating that question: given any image, predict the category it belongs to from a predefined list, as fast and accurately as possible.
Common Usage Examples
  • resnet50 = torchvision.models.resnet50(pretrained=True); pred = resnet50(image_tensor).argmax()
  • Transfer learning: model.fc = nn.Linear(2048, num_custom_classes) — replace ImageNet head
  • timm.create_model('vit_large_patch16_224', pretrained=True, num_classes=5) — ViT classifier
  • Medical imaging: classify chest X-rays into (normal, pneumonia, COVID) — CheXpert, NIH ChestX-ray14
  • Zero-shot: CLIP classifies images into novel categories via text description without any training
Summary
In short: Image classification — predicting what's in a photo — is the foundational vision task whose solved form (ImageNet-trained ResNet/ViT) now serves as the transfer learning backbone for virtually all other vision applications.