Does object recognition always require deep learning?

No. Feature-based, template-based and other methods can work for suitable tasks. Deep learning is one important approach.

Can it recognize an object it has never seen before?

That depends on the system and the intended label. Generalizing to a new instance of a known category differs from recognizing an entirely unsupported category.

Can it count objects in a photograph?

Counting normally requires identifying separate instances or a dedicated counting method. An image-level category label alone is insufficient.

AI Glossary / Computer vision and multimodal AI / Beginner

Object recognition

Definition

Object recognition is the process of identifying what an object is from visual information, such as an image or video. Depending on the task, the result may be a category, such as “chair”, or the identity of a particular object, such as a specific product model.

What it means

Recognition connects visible patterns with meaningful labels. A photograph may contain a chair even when it is partly hidden, seen from an unusual angle or shown under different lighting. A recognition system must cope with the variations relevant to its intended setting.

The term is sometimes used broadly to include several visual tasks. Image classification assigns labels to an image or crop. Object detection also identifies where object instances appear, often using boxes. Segmentation can assign labels or instance membership to pixels. A project should specify which output it needs instead of relying on “recognition” alone.

Recognizing a category is also different from establishing every property of an item. An image labeled “bottle” does not establish its contents, capacity or suitability for a particular use.

How it works

A system receives an image and transforms it into a representation useful for comparison or prediction. Traditional approaches may use designed features such as edges, shapes or local descriptors. Learned models can develop features from training examples. Template-based approaches may be sufficient for tightly controlled patterns.

The representation is used to produce labels, matches or scores. Training data and the label definitions strongly influence what distinctions the system can make. An intended distinction between two product versions may require details that are absent from the available view.

Evaluation uses images separate from those used to build the system. Relevant variation includes viewpoint, lighting, background, wear and partial visibility. Deployment also needs a response for uncertain or unsupported inputs. Selecting the largest score among known labels is not evidence that the object actually belongs to one of them.

Three practical examples

A catalog workflow proposes product categories for newly uploaded photographs, with uncertain cases sent for review.
A sorting station distinguishes several known package types under controlled lighting before a separate mechanism routes them.
A visual search feature retrieves similar-looking furniture from a collection, helping a user explore candidates without claiming an exact product identity.

Think of it this way

Think of identifying tools from a set of photographs. Shape, handle and visible markings can help distinguish a screwdriver from a wrench. Recognizing an exact model may require a label or a detail that the photograph does not show.

The analogy does not imply that software reasons like an experienced technician. It may rely on correlations in the images, including background cues that do not reliably indicate the object itself.

Worked example

Reading recognition performance carefully

Consider an invented test with 100 single-object photographs: 50 mugs and 50 bottles. A classifier correctly labels 45 mugs and 40 bottles. It mislabels the other 5 mugs as bottles and the other 10 bottles as mugs.

Overall accuracy is (45 + 40)/100 = 85%. For the label “mug”, there are 45 correct mug predictions out of 55 total mug predictions, giving precision of about 81.8%. Mug recall is 45/50 = 90% because 45 of the actual mugs were found.

These figures answer different questions. A team automatically assigning a “mug” catalog label may care about incorrect assignments, while a search tool trying to find all mugs may care about missed ones. The same 85% overall accuracy hides that distinction.

The test contains no unknown object categories and no images with multiple objects. It therefore provides no evidence that the system rejects a vase, locates two bottles separately or identifies their exact manufacturers. Those capabilities would need their own task definitions and tests.

Illustrative case study

Illustrative case study - fictional

A fictional homeware retailer builds a recognition trial using photographs on a white studio background. Initial results look encouraging. When customer-uploaded images are introduced, the system confuses some objects with visually similar surroundings.

The team reviews the mistakes and adds representative evaluation images with clutter, different viewing angles and partially covered products. It also separates broad category suggestions from exact product matching, which requires stronger evidence.

The revised workflow presents a suggestion for review instead of silently changing inventory records. The case illustrates task definition and testing decisions, not an independently measured improvement or a claim about a real retailer.

Common misconceptions and mistakes

Recognition is not automatically detection, counting or tracking. Those tasks require additional outputs and evaluation.
A model can recognize a broad category while failing to distinguish nearly identical product variants.
A high test score may not transfer to new cameras, backgrounds or operating conditions.
The highest class score is not necessarily a calibrated probability, and it does not guarantee that the input belongs to a supported class.
Labeling an object does not establish its hidden properties, condition or authenticity.

Business use

Businesses can use object recognition to assist search, catalog organization and visual inspection. The useful starting point is a precise decision: which objects need to be distinguished, from which views, and what happens when evidence is insufficient?

Measure errors in terms of that decision. A wrong search suggestion and a wrong automatic stock update can have different consequences. Keep a representative evaluation set and revisit it when products, cameras or image sources change. Where exact identifiers are available, combining visual evidence with readable labels or other records may be more useful than forcing an image-only answer.

Check your understanding

A classifier labels a photograph “bottle”. Does that prove it can locate and count every bottle in the photograph?

Answer: No. Classification provides a label. Localization and instance counting need additional capabilities and separate validation.

Related terms

Open set recognition
image classification
object detection
semantic segmentation
instance recognition
visual search

Learn next

Write down the exact output needed for a small recognition task: category, location, count or specific identity. Collect examples where those outputs differ. Build a confusion table before deciding whether an overall accuracy figure is adequate.

Sources

Sources: MathWorks, Object Recognition: https://www.mathworks.com/solutions/image-processing-computer-vision/object-recognition.html ; IBM, What is Image Recognition?: https://www.ibm.com/think/topics/image-recognition . Relevant provider explanations of recognition workflows and task distinctions were reviewed. The performance calculations and business scenario are illustrative.

Previous: Object orientation Next: Object - role modeling