Hello Model
← Model library

Image Classification

Assign a label to a whole image: defect/ok, species, product type, diagnosis.

Build a plan for this →

Typical projects

Detect defective products on a production lineIdentify plant diseases from leaf photosClassify X-ray images

Three ways to build it

Starter

No-code / AutoML vision (Teachable Machine, Roboflow, Vertex AutoML)

Upload labeled images and get a model in minutes, no code required.

Best for: No or little data, or new to ML

Roboflow · Teachable Machine

Standard

Fine-tuned pretrained CNN/ViT (ConvNeXt, EfficientNet) via timm

Transfer learning gets high accuracy from a few hundred images per class.

Best for: Some labeled data and Python experience

PyTorch · timm · torchvision

Advanced

Vision foundation model (DINOv2 / CLIP) fine-tuning + distillation to a small edge model

Best accuracy, then compress for fast/cheap inference.

Best for: Lots of data and an experienced team

PyTorch · timm · ONNX · TensorRT

How success is measured

Accuracy and per-class recall (for defect detection, missed defects matter most)

The data you'll need

  • Collect images in the same conditions as production (lighting, angle, camera).
  • Start with 100–300 images per class for transfer learning.
  • Organise as one folder per class, e.g. data/train/cat/*.jpg and data/train/dog/*.jpg

Labeling

Folder-per-class is enough. For larger sets use Label Studio, CVAT or Roboflow.

Preparing the data

  • Resize to model input size (e.g. 224×224)
  • Data augmentation: flips, rotations, color jitter
  • Remove near-duplicate images across train/test splits

Start with a baseline

A pretrained model used as a frozen feature extractor + logistic regression.

Evaluating the model

  • Confusion matrix
  • Look at misclassified images in a grid
  • Test on images from a different day/camera

Monitoring in production

  • Image brightness/size drift
  • Prediction confidence distribution
  • Sample images for periodic human audit

Common pitfalls

  • Model learns the background, not the object
  • Same item photographed twice ends up in train and test

Example code

CodeQuick start
python
# Zero-shot with CLIP: no training needed
# pip install transformers pillow torch
from transformers import pipeline
clf = pipeline("zero-shot-image-classification", model="openai/clip-vit-base-patch32")
print(clf("leaf.jpg", candidate_labels=["healthy leaf", "leaf with rust disease"]))
CodeTrain your own model
python
# pip install torch torchvision timm
import timm, torch
from torch.utils.data import DataLoader
from torchvision import datasets, transforms

tf = transforms.Compose([transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(),
                         transforms.ToTensor(), transforms.Normalize([0.485,0.456,0.406],[0.229,0.224,0.225])])
train = datasets.ImageFolder("data/train", tf)
loader = DataLoader(train, batch_size=32, shuffle=True)

model = timm.create_model("convnext_tiny", pretrained=True, num_classes=len(train.classes))
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
opt = torch.optim.AdamW(model.parameters(), lr=1e-4)

for epoch in range(5):
    for x, y in loader:
        x, y = x.to(device), y.to(device)
        loss = torch.nn.functional.cross_entropy(model(x), y)
        opt.zero_grad(); loss.backward(); opt.step()
    print(epoch, loss.item())
torch.save(model.state_dict(), "model.pt")

Ready to build one? Get a personalised plan →

Or read about Object Detection next.