Image Recognition

Image recognition is the ability of AI systems to identify and categorise the contents of images — objects, people, scenes, text, activities, and more. It is one of the most mature and widely deployed applications of deep learning.

How image recognition works

Modern image recognition uses convolutional neural networks (CNNs) or vision transformers (ViTs) trained on millions of labelled images. During training, the model learns to identify visual features — edges, textures, shapes, patterns — that are associated with different categories.

Types of image recognition tasks

• Image classification — assigns a label to an entire image ("this is a cat")
• Object detection — identifies and locates multiple objects within an image with bounding boxes ("there is a cat at coordinates X,Y and a dog at coordinates A,B")
• Segmentation — classifies every pixel ("these pixels are cat, these are background")
• Face recognition — identifies specific individuals from facial features
• Optical character recognition (OCR) — reads text from images
• Scene understanding — identifies the context ("this is a kitchen," "this is a highway")

Business applications

• Manufacturing — visual quality inspection on production lines
• Healthcare — detecting tumours in medical scans, analysing pathology slides
• Retail — visual search ("find products that look like this photo"), inventory management
• Security — surveillance analysis, access control
• Agriculture — crop disease detection, yield estimation from aerial images
• Insurance — damage assessment from photos of vehicles or property

Accuracy and limitations

Modern image recognition exceeds human accuracy on many benchmarks but can fail in unexpected ways:

• Adversarial examples — subtle image modifications invisible to humans can cause misclassification
• Domain shift — a model trained on professional photos may fail on smartphone photos taken in poor lighting
• Bias — models may perform worse on underrepresented demographics or unusual contexts
• Context dependence — a model might correctly identify a stop sign in daylight but miss it when partially obscured by snow