AI-901
Deep Dive
Computer vision lets machines "see" and interpret images and video. The AI-901 tests the main tasks, the differences between them, and when to use each Azure service.
Contents
Computer vision is a field of AI that allows systems to extract meaningful information from digital images, videos, and other visual inputs. It uses deep learning (convolutional neural networks) to learn to recognize visual patterns.
Why is it hard for a machine?
How does deep learning solve it?
Convolutional neural networks (CNNs) automatically learn hierarchies of features: edges → shapes → textures → parts → objects. With millions of labeled images, they learn to "see" in a way similar to humans.
Image classification
What IS in this image?Output
A label (and its confidence): "dog (94%)"
Example
Classifying product photos in a catalog
Object detection
WHAT is in this image and WHERE?Output
Labels + bounding boxes (location rectangles)
Example
Quality control system: detecting defects in products on a conveyor belt
Semantic segmentation
Which class does EVERY PIXEL belong to?Output
A color mask where each color = a class (car, road, sky, pedestrian)
Example
Autonomous vehicles that need to understand the exact boundaries of each object
OCR (optical character recognition)
What TEXT is in this image?Output
Extracted text with position and confidence
Example
Digitizing scanned invoices, reading vehicle license plates
Facial analysis
Are there faces? Who are they? How do they look?Output
Face coordinates, identity (if a database exists), emotions, attributes
Example
Biometric access control, monitoring attention in classrooms
Multiclass classification
Each image belongs to exactly ONE class. The model picks the most probable category.
"This image is a cat (87%), dog (9%), rabbit (4%)"
Multi-label classification
An image can have MULTIPLE labels at once.
"This image contains: beach (95%), palm trees (89%), sunset (78%)"
Azure: image classification
Azure AI Vision — Image Analysis: for general categories (animals, vehicles, landscapes).
Custom Vision (classification): when you need business-specific categories (type of manufacturing defect, specific plant variety). Its retirement is announced for 2028; today Content Understanding or a multimodal model in Foundry is recommended.
Extract a description, tags, and objects from a photo in under a second, with no training required.
+1 more...
Object detection goes beyond classifying — it locates each object with a bounding box (a bounding rectangle with x, y, width, height coordinates) and assigns it a class.
🏭
Quality control
Detecting and locating defects in products on a production line. The system marks exactly where the defect is.
🛒
Retail analytics
Counting products on shelves, detecting when stock is low, analyzing inventory layout in-store.
📷
Security and surveillance
Detecting people, vehicles, or specific objects in security cameras. Alerting when something is detected in a restricted zone.
Classification vs detection — common trap
Classification: "There is a dog in the image." Detection: "There is a dog at coordinates [120, 80, 300, 250] and a cat at [400, 90, 550, 280]." If the exam asks about locating objects, it's detection.
Detect and locate 4 people in an image, each with its own bounding box and confidence score.
+1 more...
Semantic segmentation
Assigns a class to EVERY PIXEL. All pixels of the same object type get the same color/class, without distinguishing individual instances.
All cars = blue. All road = gray. All sky = light blue. (No matter how many cars there are)
Instance segmentation
Same as semantic + distinguishes individual instances of the same type.
Car 1 = light blue. Car 2 = dark blue. Car 3 = medium blue. (Each car has its own mask)
Real-world use cases
Autonomous vehicles — understanding exactly the boundaries of the road and obstacles
Medical imaging — segmenting a tumor from healthy tissue
Photo editing — automatically separating the subject from the background
Agriculture — identifying affected crop areas in satellite imagery
OCR extracts text from images and scanned documents. It converts visual content (a photo of an invoice, a printed document, a sign) into digital text that a computer can process.
Use cases
Digitizing invoices and receipts for automatic accounting
Reading vehicle license plates in parking lots or toll roads
Processing physical forms (applications, contracts)
Making scanned documents searchable
Extracting data from business cards
In Azure
Azure AI Vision — Read API (OCR)
For images and PDFs. Extracts text with its position, detected language, and confidence. Optimized for dense documents.
Azure AI Document Intelligence
OCR + form understanding. Extracts specific fields from invoices, receipts, ID cards. Understands the document's STRUCTURE.
Critical AI-901 distinction
This difference comes up frequently on the exam. Memorizing it is mandatory.
Facial detection
Identifies the presence and location of faces. Answers: "Are there faces? Where?"
Output: bounding box coordinates for each detected face.
Doesn't know who they are
Facial recognition
Identifies who a person is by comparing against a database of known people. Answers: "Who is this?"
Output: the person's identity (with a confidence score).
Knows who they are
Other capabilities of Azure's Face API
Verification
Are this photo and that other photo the same person?
Grouping
Grouping photos of unidentified people by similarity
Attributes
Estimated age, expression, head orientation, glasses
The AI-901 doesn't require knowing the math behind CNNs, but it does require understanding the concept of how they learn.
Convolutional layers
Apply filters that detect local patterns: first edges and corners, then textures, then simple shapes, then complex structures.
Pooling
Reduce the spatial size while preserving the most important features. They make the model more efficient and give it some invariance to translation.
Fully connected layers
At the end of the network, they combine all the learned features to make the final prediction (classification).
Transfer Learning
Instead of training from scratch (expensive), a CNN pre-trained on millions of images (ResNet, VGG) is used and fine-tuned for the specific task with little data.
Did you get this topic?
Test what you just learned
A supermarket wants a system that automatically detects when a product is misplaced on a shelf, identifying exactly which product is out of place and at what position in the image. Which computer vision task is most appropriate?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.