Glossary
What is computer vision
In short
Computer vision is the field of AI in which programs understand images and video: they find objects, read text, count and measure.
AI basics · computer vision
MMOOZZGG.com editorsChecked: 1 min read
For a child
You look out of the window and see at once: there is a cat on the fence, there is a yellow car. For a computer, a picture is at first just thousands of colored dots, like a mosaic. It is taught to make sense of the dots: where the cat’s ears are, where the car’s wheels are. Once it learns, it can tell by itself what is in a photo, and even count all the pigeons in the park.
In plain words
A camera gives a computer only a grid of pixels, each with its own color. To understand what is in a picture, a neural network is trained on a huge number of labeled photos. It finds edges in the pixels, then shapes, then whole objects. After that the tasks differ: say what is shown, draw a box around each person, separate the road from the sidewalk, read a car number or the text on a receipt. This is how a phone that unlocks by your face and a gate that lets a car in by its number work.
For business
In manufacturing, such systems look for defects on the line and check whether helmets are on. In retail they count visitors and empty shelves, in logistics they read barcodes and railcar numbers, and in accounting they recognize scans of delivery notes. The cost is made up of cameras and lighting, image labeling, computers on site or in the cloud, and tuning for your conditions. A model trained on other people’s photos often errs in your warehouse with different light, so you need your own examples, including rare defects. If cameras film people, the rules on personal data and biometrics apply. It is sensible to test the solution on one site before buying equipment for the whole shop floor.
For a specialist
Computer vision combines image classification, object detection, semantic and instance segmentation, optical character recognition (OCR), tracking and pose estimation. The leap in quality came with convolutional neural networks: AlexNet (Krizhevsky, Sutskever, Hinton, NIPS 2012) was trained on about 1.3 million ImageNet images with 1,000 classes, after which deep models pushed out hand-crafted features. YOLO (Redmon et al., 2015) reduced detection to a single regression of boxes and classes in one pass over the whole frame and ran at about 45 frames per second. Vision Transformer (Dosovitskiy et al., 2020) showed that a transformer on a sequence of image fragments (patches), with large-scale pre-training, matches convolutional networks. Today vision models are often built into multimodal systems together with text. Metrics: accuracy, mAP for detection, IoU for segmentation.
Example
A plant puts a camera over the packing line. The model spots crushed boxes and crooked labels and alerts the operator, and a log of defects by shift reaches the shop manager every morning.
Related terms
Questions and answers
Sources
- Krizhevsky, Sutskever, Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NIPS 2012papers.nips.cc
- Redmon et al., You Only Look Once: Unified, Real-Time Object Detection, 2015arxiv.org
- Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2020arxiv.org
Where companies use it
Not sure whether your company needs this?
AI solutions map