The clearest classroom contrast is this: image processing takes an image in and gives an image out, while computer vision takes an image in and gives out a description, a label or a decision about what the image shows. Sharpening a photo is image processing. Your phone deciding that the face in front of it is yours is computer vision. For a slide: “Image processing changes pixels to make an image better or different; computer vision analyses pixels to work out what is actually in the image.”
TechnoLynx builds computer vision systems through its Computer Vision R&D work.
Is “image processing changes pixels, computer vision understands the image” accurate?
Yes, as an introductory characterisation it is accurate. Image processing covers operations such as contrast enhancement, noise removal, edge extraction and rotation, and it neither requires assumptions about the content nor produces interpretations of it. Computer vision aims at “understanding”, which Wikipedia’s article describes as turning visual images into descriptions of the world that can elicit appropriate action.
The one refinement worth teaching is that the difference lies in the goal, not the method. The same arithmetic on pixels can appear in both fields; what separates them is whether the output is a changed image or knowledge about the image. A 2021 arXiv paper on a document-scanning app built with OpenCV puts it plainly: computer vision may or may not use image-processing algorithms, and “the main difference in Image Processing and Computer Vision lies in the goal”.
If they learn “image processing uses filters, computer vision uses AI”, they will be confused the first time they meet a vision system full of filters.
Image processing and computer vision side by side
| Image processing | Computer vision | |
|---|---|---|
| Input | An image | An image or video |
| Output | A new image | A label, a bounding box, an analysis or a system’s behaviour |
| Typical operations | Brightness and contrast, noise reduction, sharpening, rescaling, rotation | Detecting, recognising and classifying objects, then deciding (pass/fail, match/no match) |
| Assumes anything about content? | No | Often relies on assumptions about the scene |
| Classroom example | A photo filter that brightens a selfie | Face unlock checking whether the face is the owner’s |
Does image processing come under computer vision, or are they separate fields?
Neither answer is quite right: the two sit on a continuum with no agreed boundary. Gonzalez, Woods and Eddins, in the opening chapter of Digital Image Processing Using MATLAB, state that there is no general agreement on where image processing stops and computer vision starts, and they call the strict “image in, image out” definition “limiting and somewhat artificial” (by that rule, even computing an image’s average brightness would not count as image processing). Wikipedia’s article goes further and notes that the techniques overlap so much that the fields can be read as one field with different names, kept apart partly because research groups, journals and companies need to present themselves as belonging to one.
The textbook’s alternative is a three-level picture that works well on a whiteboard:
- Low level: noise reduction, contrast enhancement, sharpening; image in, image out.
- Mid level: splitting an image into regions, describing them and recognising individual objects; image in, attributes out (edges, contours, the identity of an object).
- High level: “making sense” of a collection of recognised objects, up to the cognitive work normally associated with human vision.
Low level is what students would call image processing and high level is plainly computer vision; the middle is where the argument lives. That is also why image processing is common as a preprocessing stage for computer vision: brightening a photo so text is clearer improves the detector that then finds and reads the words.
Wikipedia’s article notes that the best current algorithms for such recognition tasks are based on convolutional neural networks. Students who go on to study the subject will meet this, so it is worth one sentence in the lesson: the pixel/meaning split describes what a system is for, not how its insides are drawn.
Computer vision examples students already know from daily life
Three everyday examples cover the main computer vision tasks without any maths, and each one pairs naturally with an image-processing counterpart on the same device.
Face unlock (recognition). Each time an iPhone unlocks with Face ID, the TrueDepth camera captures depth data and an infrared image, and that information is matched against a stored mathematical representation of the owner’s face, according to Apple’s support documentation. Apple also says Face ID uses machine learning and needs depth information, which a printed photo does not have. The pairing: a filter on the same phone that brightens or sharpens a photo changes pixels without recognising anyone.
Photo app face grouping (detection, then recognition). Google’s help pages for Google Photos describe face grouping in steps: first the app detects whether a photo contains a face, then it builds face models that numerically represent faces and estimates whether different images show the same person. This is a useful example because it separates “is there a face?” from “whose face is it?”, which are two different questions to a computer.
A car braking for a pedestrian (detection and tracking). Frame this as driver-assist braking rather than full self-driving, so students do not take away the idea that autonomous cars are an everyday product. Waymo’s blog on its perception system describes software that turns video, lidar and radar data into a form the Waymo Driver can interpret, and fuses sensors to improve confidence in its detections. The tracking part is knowing that the pedestrian seen now is the one seen half a second ago.
If the class wants a fourth, ask students to name one themselves and decide which task it is. Code.org’s machine-learning lesson video lists image recognition among its AI examples, alongside email filters and video recommendations.
Check each year which of these phone features and apps your students actually recognise.
How do I explain “the computer learns from examples” without neural networks or maths?
Have students do the labelling themselves. Code.org’s machine-learning lesson has students train a real machine learning model and explore how training data lets it classify new data, and defines machine learning as “how computers recognize patterns and make decisions without being explicitly programmed”. The lesson also suggests training the model badly on purpose, or not at all, so students see what goes wrong. That one move teaches more about training data than any diagram.
Google’s Teachable Machine does the same with a webcam: students capture example images for each category, click once to train, and the model sorts new images into those categories, with no coding. Google states the training happens in the browser and the examples stay on the computer unless the user saves the project to Google Drive.
The spoken analogy can be as simple as: show a child many pictures labelled “cat” and “dog”, and eventually they sort new pictures correctly without being told the rule. Say out loud that this is a teaching picture; a computer finding statistical patterns is not the same as a person understanding what a cat is. This is also where the computer vision vs machine learning question comes up: many vision applications are pre-programmed for a particular task, but learning-based methods are increasingly common.
Free, no-install browser demos that show object detection live
Browser pages run object detection on a webcam feed or an uploaded photo, and Pixlane’s needs no sign-up for core use. Pick one, test it on the school network before the lesson, and keep a second as a fallback.
- Pixlane Object Detection: labels 80 common object classes in images, video files or a live camera feed with boxes and confidence scores; the site says processing runs on the device, media is not uploaded, and no account is needed for core use.
- Skycally Object Detection: the COCO-SSD model, recognising 80 everyday categories, running entirely in the browser through TensorFlow.js with no frames sent to a server. It draws a separate box for each object in the frame, whether one or twenty.
- TensorFlow.js demos: Google’s gallery includes Emoji Scavenger Hunt (find real objects matching emojis with a phone camera) and a webcam Pac-Man controller trained in the browser.
- Teachable Machine: the learning-from-examples demo described above.
For the image-processing half of the pairing, the AI4K12 guidelines suggest showing the live output of vertical and horizontal edge detectors on webcam images. Put an edge-detected webcam feed next to an object-detection feed and the contrast teaches itself: one changes the picture, the other names what is in it.
Skycally’s page also makes a point worth borrowing: detection differs from classification because it locates each object with a bounding box rather than labelling the whole photo.
Limits to plan for
- School networks and device policies may block webcam access or the demo site, so test on a student account, not a staff one.
- These are third-party services, so recheck their features and sign-up rules before each lesson.
- The object-detection demos recognise a set list of 80 common categories, so they show the idea, not what a custom detector would achieve on a specific task. A demo that mislabels a pencil case is a good discussion prompt, not a verdict on the technology.
Embedding a demo in slides or a classroom page
A demo page can usually be placed inside another web page with an iframe, the HTML element that shows one site inside another. Hugging Face’s documentation says an embedded public Space lets an audience interact with a demo “without requiring any setup on their side”, using an iframe pointed at the Space’s URL. Classroomscreen offers a free Embed widget that takes embed code from tools such as Google Slides, PowerPoint or Canva and displays it for students during a lesson. For coding classes, CodePen pens can be embedded in other sites and suit workshops where nobody should spend time installing editors.
For a slide deck or a Google Classroom post, a plain link to the demo is the simpler route.
A lesson order that holds the contrast together
- Put the one-line definition on the board.
- Run the edge-detection webcam feed, then the object-detection feed, side by side.
- Walk through face unlock, photo grouping and driver-assist braking, naming the task in each.
- Let students train Teachable Machine or AI for Oceans, including one deliberately bad model.
- Finish with the low, mid and high levels, and ask where each demo sits.
Starting from things students already use and ending on the continuum means they leave with a definition that will survive their next course, rather than one they will have to unlearn. The question left open is the one the textbook raises: where does recognising an object stop being processing and start being understanding? There is no settled answer, and letting a class argue about it is a fair test of whether they grasped the goal-based distinction.
FAQ
Is computer vision also AI?
Yes, in both sources cited here. The Gonzalez, Woods and Eddins textbook describes computer vision as a branch of artificial intelligence, and the 2021 arXiv OpenCV paper calls it a field within AI.
Is computer vision the same as image processing?
No. Their techniques overlap heavily and there is no agreed boundary, but image processing aims to change an image, while computer vision aims to interpret it.
