presentsUNAIVERSE
The UN AI timeline

    Learn AI /Computer Vision and Language Technology /Image classification

    Image classification

    Also called: image labelling.

    Detail

    In five seconds

    What it is

    A system assigns one or more predefined labels to an image based on patterns in its pixels.

    In plain words

    It is a way for a computer system to look at a picture and choose a label from a fixed list.

    An everyday example

    A programme officer opens a shared inbox after a storm response. Each message is quickly placed into one of three trays: “urgent action”, “information only” or “send to logistics”. The officer uses clues such as words in the subject line, who sent it and whether it mentions blocked roads or missing supplies. At the end, each message gets a category, even though the officer does not rewrite the message or mark every sentence separately.

    Where the story stops working

    A human sorter can use outside context, ask questions and change the categories on the spot. An image classifier usually works only with the labels it was trained for and only from the image data it receives. Also, an email can be read sentence by sentence, while image classification usually treats the whole picture as one item rather than locating each object inside it.

    Remember thisImage classification assigns predefined labels to an image as a whole. It does not describe the image freely or identify where each object appears.

    Understand it

    Understand it

    Why it matters

    Many real systems do not chat. They sort. Image classification is one of the simplest and most widely used AI tasks for that kind of sorting. If you understand this term, you can ask practical questions: What are the allowed labels? Who chose them? What images was the model trained on? Does it produce one label or several? What happens when the picture does not fit any label well? Those questions matter because a classifier can save time, but it can also mislabel images when lighting, location, camera quality or local context differ from the training examples. For non-specialists, the value is not in building the model yourself. It is in recognising what the tool can and cannot tell you before you rely on it for a workflow or decision.

    How image classification turns a picture into a label

    After viewing this, a reader should be able to say that image classification takes an image, analyses pixel patterns with a trained model, scores known categories and outputs the best-fitting label.

    1. 01

      1. Input image

      A single image is provided to the system, such as a photo of a road, document or crop field.

      Draw a thumbnail image in a rounded rectangle labelled “Image”. Use a simple neutral scene icon, not a person’s face.

    2. 02

      2. Pixels become numbers

      The picture is represented as pixel values so a model can process it mathematically.

      Show a small grid overlay on the image and a side panel with a few example number triples to suggest pixel values.

    3. 03

      3. Trained model compares patterns

      The model checks the numeric pattern against patterns learned from labelled training examples.

      Draw a larger centre box labelled “Trained model” with simple stacked layers or abstract nodes. Add a note below: “learned from many labelled examples”.

    4. 04

      4. Scores for known labels

      The system produces scores for the categories it knows, such as “damaged road 0.82” and “undamaged road 0.18”.

      Show a short ranked list with horizontal bars for three labels and their scores.

    5. 05

      5. Output label

      The highest-scoring label is returned as the prediction, sometimes with the scores shown as well.

      Highlight the top label in a final output box. Include a small note: “score is not certainty”.

    A left-to-right flow diagram with four boxes connected by arrows. It starts with one example image, shows that the image becomes pixel values, then passes through a trained model, and ends with ranked labels. Read from left to right as a process from input to output.

    On hover or tap, each step highlights and reveals one sentence of explanation; a toggle switches between single-label and multi-label examples.

    Describe this without the diagram

    The diagram explains image classification as a five-step flow. First, one image enters the system. Second, the image is represented by pixel numbers. Third, a trained model analyses the numeric pattern using what it learned from many labelled examples. Fourth, it produces scores for the categories it knows. Fifth, the top category is returned as the label, with a reminder that a high score does not guarantee the answer is correct. The main idea is that the system maps a whole image to one or more predefined categories.

    With reduced motion: A static left-to-right diagram with the same five labelled boxes and arrows. The selected example label is highlighted without animation, and the single-label or multi-label view changes only when the reader presses the toggle.

    Image classification takes a picture as input and returns a label such as “crop field”, “damaged road”, “invoice” or “not safe helmet use”. A label is a category chosen in advance. The system does not produce any category it likes; it selects from the set it was built to use.

    Inside the system, the image is converted into numbers from its pixels, the tiny picture elements. A model then compares the pattern in those numbers with patterns it learned from many labelled examples during training. It outputs scores for the possible labels, and the highest score often becomes the predicted label.

    The output may be one label, such as “flooded”, or several labels if the system allows more than one category at once, such as “road”, “vehicle” and “smoke”. Some systems also show a confidence score, but that score is not a guarantee of correctness.

    This is often confused with object detection and image segmentation. Image classification answers “What category best fits this whole image?” Detection asks “Where in the image is each object?” Segmentation goes further and marks which pixels belong to which object or region.

    How “Sorting urgent and non-urgent messages” maps onto the real thing

    In the storyIn the AI system
    Each incoming messageOne input image
    The trays with fixed namesThe predefined labels or classes
    Clues used for sortingPatterns in pixel values learned during training
    Placing a message into one trayAssigning the most likely class to the image

    A worked example

    Before the morning repair schedule is issued, a road inspection supervisor reviews dashboard photographs using a classifier with two labels: “pothole present” and “no pothole”.

    Starting from: Photo ID 1842, taken from a vehicle dashboard camera on a dry road in daylight.

    1. The image is resized to the dimensions expected by the model and its pixels are converted into numerical values.
    2. The trained model applies patterns learned from labelled photographs, including road texture, edges and shadows.
    3. It produces normalised probabilities: “pothole present” 0.91 and “no pothole” 0.09.
    4. Because this is single-label classification, the higher-probability class becomes the prediction.

    Result: Photo ID 1842 is labelled “pothole present” and placed in the supervisor’s review queue. The prediction still requires checking, particularly if the camera angle, lighting or road surface differs from the model’s training examples.

    In a United Nations workplace

    Following a scheduled bilateral with a bidder, an evaluation officer in a procurement unit drafts a note for the file on a demonstrated image-classification module. Using pre-cleared test photographs, the module assigns one or more predefined labels, such as “intact packaging”, “water damage” and “crushed packaging”, to each image.

    The officer can record whether the module consistently maps whole images to the expected labels, including how it handles photographs with more than one type of damage. This tests classification performance, not whether the system can locate each damaged area within an image.

    Before you try thisDo not enter bid submissions, commercially confidential images, photographs containing identifiable people or other procurement-sensitive material into a tool that has not been approved for those data. Demo outputs are not independent proof that a requirement has been met. Results need an agreed test set, documented criteria and human verification.

    Where you may hear it

    • During a vendor demonstration in a procurement meeting
    • At an ICT governance board review of a proposed visual inspection tool
    • In an information management briefing on satellite-image triage
    • During a model-testing session before an image-processing tool is cleared

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • If the system labels an image correctly once, it has understood what is in the picture like a person does.

      A classifier detects statistical patterns linked to labels in training data. It can perform well on familiar cases without having human-like understanding or broader common sense. That is why it may fail on unusual angles, new contexts or misleading backgrounds.

    • Image classification tells you where every object is in the image.

      Usually it does not. Standard image classification labels the whole image or assigns image-level categories. If you need locations, you usually need object detection with bounding boxes or image segmentation that marks pixels.

    • A 95 percent confidence score means there is a 95 percent chance the answer is true in any situation.

      Not necessarily. The score reflects the model’s internal scoring behaviour, which may not match real-world reliability. Scores can look high even when the image is very different from the training data or when classes are poorly defined.

    Quick check

    Which task best matches image classification?

    Go deeper

    Go deeper

    The simple version

    It is a way for a computer system to look at a picture and choose a label from a fixed list.

    The technical version

    In technical terms, image classification maps an input image to one class or a set of classes. During training, a model learns parameters from labelled examples so that patterns in pixel values are associated with target labels. At inference time, the image is preprocessed, passed through the model, and converted into class scores, often called logits or probabilities after calibration or normalisation. In single-label classification, the system usually selects the highest-scoring class. In multi-label classification, it may assign several labels if each passes a threshold. This differs from object detection, which predicts classes plus locations, and from image segmentation, which predicts a class for each pixel or region.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Image classification assigns predefined labels to an image as a whole. It does not describe the image freely or identify where each object appears.