Skip to content
Français

Vision AI: See, Count and Inspect

Turn photos into decisions you can defend. Sort product photos, count boxes on a pallet, find damage on returned items and match a customer's photo to the catalogue. Compare simple rules, pretrained models, embeddings, detectors and vision-language models on the same images, choose by the cost of a mistake, and run the chosen model as a small image service.

Read every lesson as a guest, or sign in to enroll and track your progress.

Is this the right course for you?What this course covers and does not cover, and 4 courses to take for other needs.

This course teaches vision for things in the world as a business tool: framing the decision and the cost of a mistake, simple baselines, pretrained classifiers and transfer learning, image embeddings and CLIP, detection and counting, promptable segmentation, inspection with and without damage labels, image matching, vision-language models and how to check them, and a small image service with a review queue, CPU latency and privacy by default. It does not teach how neural networks learn, documents and OCR, or the platform that deploys and monitors models.

Take another course if you want to:

What you'll learn

By the end of this course, you will be able to:

  • Frame a vision problem: the decision, the output type (label, box, mask, match, text), the camera and lighting conditions, and the cost of a miss and of a false alarm.
  • Build simple baselines (colour and size rules, colour histograms, HOG features with a linear model) and say when they are enough.
  • Apply and compare pretrained classifiers with transfer learning, CLIP zero-shot and linear probes, and few-shot methods, on the same images, and explain why each wins or loses.
  • Detect, count and segment objects with pretrained detectors and promptable segmentation, and measure them with IoU and mAP by class and by size.
  • Choose between supervised defect detection and anomaly-based inspection, find matching and duplicate images with embeddings, and use vision-language models where they help and check them where they fail.
  • Run a vision model in production: labelling and label quality, domain shift, CPU latency (ONNX, quantisation), review queues, privacy of people in images, and what not to build.

About this course

Larkfield’s returns warehouse checks about 600 returned items a day by eye. When a damaged candle goes back on sale as new, a customer complains and Larkfield pays twice. When a good candle is marked damaged, it is sold at the outlet and loses its value. “Use AI on the photos” sounds like the answer, but it is not yet a plan.

In this course, you follow Priya, Larkfield’s data scientist, and Hamid, the returns warehouse lead, as they turn photos into decisions. You start with the simplest methods, such as a colour histogram, and you test every method the same way, on the same images, in one comparison table. You see a simple rule beat a large model on one task and lose badly on another, and you learn why.

Then you go further: pretrained models and transfer learning, image embeddings and CLIP, detectors that count boxes on a pallet, masks that measure damage, inspection that learns only from good items, and vision-language models that describe a product, sometimes with details that are not in the photo. You choose the thresholds with Hamid, from the cost of each mistake and the number of checks his team can do in a day.

You finish with a small image service that answers on a laptop CPU, sends unsure images to a person, blurs faces by default and notices when a new camera changes the images. The Larkfield images are open photos and images drawn by our own code, each with its licence, and every notebook runs on a laptop CPU or on free Colab in a few minutes.

Before you start

You completed Machine Learning, From Problem to Reliable Model, or you have the same skills. You can read and change Python code in a notebook, and you know train, validation and test sets, precision and recall, and thresholds. The deep learning course helps but is not required. You need a free Google or Kaggle account for the notebooks, or Python 3.12 on your computer. You need no key, no graphics card and no money.

Courses to take first: Machine Learning: From Problem to Reliable Model.

Who this is for

ML engineers, data scientists and backend developers who must turn photos and camera images into decisions, such as sorting product photos, counting items, finding damage, matching a photo to a catalogue item or describing an image.

How we made this course

Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses

Inside the course

Final module · choose at least oneCase studies

3 case studies

Three end-to-end case studies on real open images. Inspection with few defects, fine-grained classification, and counting in crowded scenes. Complete at least one for the certificate; each one is a portfolio piece.

Complete at least one case study for the certificate. The others are optional. What each one covers

  1. Visual inspection with few defects120 min
  2. Classify products and pets from photos150 min
  3. Count and detect in crowded scenes150 min

After the case studiesFinal assignment

Use every module in one practice project. You collect evidence, and check it against the course outcomes.

  1. Final assignment180 min

Case studies

Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.

  1. Visual inspection with few defects2 hr

    Inspect industrial parts from an open inspection set: compare a supervised classifier, PatchCore and a CLIP prompt, and choose thresholds from an inspection budget.

  2. Classify products and pets from photos2 hr 30 min

    Fine-grained classification on open photos: compare a histogram baseline, zero-shot CLIP, a linear probe and fine-tuning against labelled images per class, with the worst class.

  3. Count and detect in crowded scenes2 hr 30 min

    Detect and count objects in crowded open photos: compare two detectors, a density method and a vision-language model on count error by crowding.

Reviews

Loading the reviews…

    Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.

    Share this course