Vision Solution Types

Every one of the five vision solution types takes the same thing in (an image), so you can never tell them apart by their input, which is exactly the trap the Features of computer vision workloads objective is built on. What separates them is the output: a whole-image label, a box, a pixel mask, recognized text, or a face location. Lock the five input→output signatures below and a stem that describes the result it needs answers itself. The taxonomy below fixes the picture: one shared input feeds five type cards, each labeled with the single output that defines it.

Image classification → one label for the whole image

Microsoft calls image classification "one of the oldest computer vision solutions ... in which a model that has been trained with a large number of images is used to predict a text label based on an image's contents"^[1]. The input is one image; the output is a single text label naming the image's main subject, for example a produce photo on a checkout scale classified as "apple." Critically, classification gives you what the image is about but not how many objects there are or where they sit. There are no coordinates in the answer.

Object detection → per-object label + bounding box

Object detection answers a richer question. Microsoft describes it as a model that "examine[s] multiple regions in an image to find individual objects and their locations,"^[1] returning "which objects were detected, and the specific regions of the image in which they appear, indicated by the coordinates of the rectangular bounding box." So the output is a list: each detected object gets its own class label plus rectangular bounding-box coordinates. Because each object is reported separately, object detection can also count items: the smart-checkout scenario that must identify and locate every item on the scale is object detection, not classification.

Semantic segmentation → per-pixel mask

Semantic segmentation is, in Microsoft's words, "another, more sophisticated way to detect objects in an image ... a model is trained to find objects, and classify individual pixels in the image based on the object to which they belong,"^[1] giving "a much more precise prediction of the location of objects." The output is a per-pixel mask, not a box. Where object detection draws a rectangle that necessarily includes some background, segmentation traces each object pixel-for-pixel, the most precise localization of the five types.

Optical character recognition (OCR) → the text itself + its location

OCR is the odd one out: its output is text, not an object label. Microsoft defines it as machine-learning techniques that "extract printed or handwritten text from images such as posters, street signs, and product labels, as well as from documents like articles, reports, forms, and invoices"^[2]. The Read engine returns "pages, text lines, and words with location and confidence scores." So OCR converts pictures of words into machine-readable text and tells you where on the image that text was found. If the goal phrase is "read the text" or "digitize the receipt," it is OCR.

Facial detection & analysis → face location (+ attributes), not identity

Facial detection finds where the faces are. Microsoft's Face service description is precise: the "Detect API detects human faces in an image and returns the rectangle coordinates of their locations,"^[3] and "optionally, face detection can extract a set of face-related attributes, such as head pose ... and glasses." That is detection (where is the face?) plus analysis (what can be described about it?). It is deliberately not the same as facial recognition, which matches a detected face to a known identity, a separate operation. Detection alone never tells you who the person is.

Once you know each type's output, the differences become a clean hierarchy of spatial precision plus one outlier (OCR) whose output is a different kind of thing entirely. Memorize the progression and the distractors collapse.

The precision ladder: none → box → pixel

The three object-oriented types form a ladder of increasing locational precision, modeled on the way Microsoft Learn's computer-vision overview steps from classification through detection to per-pixel segmentation:

• Image classification, no location at all. Output is just a whole-image label. It cannot say where the subject is or how many subjects there are. ("Whole image → one word.")
• Object detection, coarse location, the rectangle. Output adds bounding-box coordinates per object, so you get where and how many, but the box is a rectangle that may enclose background pixels around the object.
• Semantic segmentation, exact location, the pixel mask. Output classifies every pixel, tracing the object's true outline. This is the most precise of the three and the right answer whenever a scenario needs an object's exact shape, area, or boundary rather than a rough box.

A compact way to hold it: classification = what; object detection = what + where (boxed) + how many; segmentation = what + where (pixel-perfect).

OCR is not on that ladder, its output is text

OCR shares the "reads an image" framing but produces the recognized characters as its result, with location as secondary metadata. The decisive cue is the verb in the stem: "classify / categorize the image" → classification; "find / locate / count objects" → object detection; "trace the exact outline / segment" → segmentation; "read / extract / digitize the text" → OCR. The presence of the word text (or receipts, signs, forms, license plates, handwriting) is the OCR tell.

Facial detection vs facial recognition, the most exploited distinction

The blueprint bundles facial detection and analysis as one solution type, and the most common AI-900 trap is to slide from there into facial recognition:

• Detection = locate faces, return their rectangle coordinates. Answers "is there a face, and where?"
• Analysis = describe attributes (head pose, glasses, etc.). Answers "what can be described about it?"
• Recognition = match a detected face to a specific known identity (verification = "are these two the same person?"; identification = "who, out of these many, is this?"). Answers "who is it?"

Microsoft frames detection as the prerequisite first step ("Face detection is required as a first step in all the other scenarios"^[3]) while keeping recognition (identify/verify) a distinct operation. The falsifiable rule: face detection alone never identifies the person. A stem that says "detect that a face is present and where" is detection; one that says "confirm this is employee Jane" or "match against a watchlist" is recognition, which carries additional Limited Access / Responsible-AI gating.

Quick reference

Solution type	Output	Tell-tale verb / cue
Image classification	A single label for the whole image	"classify / categorize this image"
Object detection	Per-object label + bounding-box coordinates (can count)	"find / locate / count the objects"
Semantic segmentation	Per-pixel mask classifying each pixel by its object	"exact outline / shape / segment"
OCR	Recognized text + where it appears	"read / extract / digitize the text"
Facial detection & analysis	Face location (rectangle) + optional attributes	"find / describe faces" (not "who is it?")

Vision-solution-type questions on AI-900 are scenario → type mappings. The stem paints a business situation and the four options are solution types (or a mix of types and Azure services). Below are the recurring shapes, why the right answer is right, and why the favorite distractors are wrong, so you score on recognition under time pressure.

Pattern 1: "sort / categorize images" → image classification

Stems: "automatically file each uploaded product photo into a category," "decide whether a photo is of a cat or a dog." One label for the whole image, no need to know location or count → image classification. The distractor object detection is wrong here because nothing in the stem asks where or how many: paying for bounding boxes you don't need is the over-engineering trap.

Pattern 2: "find AND locate / count multiple items" → object detection

Stems: "scan the checkout and identify each item and where it is," "count how many cars are in the parking-lot photo and box each one." The give-aways are multiple objects, locations, and counting in one image → object detection, because it returns a class label plus bounding-box coordinates per object. The distractor image classification fails because it yields only one whole-image label and cannot count or localize.

Pattern 3: "exact outline / precise area / pixel-level" → semantic segmentation

Stems: "highlight the exact pixels belonging to the tumor," "measure the precise area each crop covers in the aerial photo," "separate foreground object from background pixel-by-pixel." The words exact, outline, area, or pixel point to semantic segmentation. The distractor object detection is the tempting near-miss (it also localizes) but a rectangle can't give an exact shape or area; only a per-pixel mask can.

Pattern 4: "read / extract text" → OCR

Stems: "digitize printed receipts," "read the number from a license plate," "extract handwritten notes from scanned forms," "capture the text on street signs." Any goal that converts pictures of words into machine-readable text → OCR (the Read capability). Two distractors recur: (a) image classification/object detection, wrong because the desired output is the text content, not an object label or box; (b) Document Intelligence / form processing, a tempting trap when the source is a form, but plain "read the text" is OCR; only "pull specific fields / key-value pairs / tables" rises to document intelligence (a different AI workload).

Pattern 5: "find or describe faces" → facial detection & analysis (NOT recognition)

Stems: "detect how many faces are in the photo and where," "blur every face for privacy," "describe whether subjects are wearing glasses." Locating or describing faces without naming them → facial detection & analysis. The signature distractor is facial recognition, the detection-vs-recognition line drawn in How they differ above: it is wrong here because no stem asks who the person is. The reading-comprehension tell is whether the stem names or verifies an identity: "is this person Jane / match the watchlist" makes it recognition; "is there a face / blur the faces" keeps it detection.

Pattern 6: the scope traps, wrong layer or wrong domain

Two distractor families try to pull you off the solution-type question entirely:

• "Which Azure service?" Some options name resources (Azure AI Vision, Azure AI Face, Custom Vision). If the question asks for the solution type / task, the service name is the wrong layer: service selection is a separate subtopic.
• "Generate a new image" Creating an original image from a prompt is generative AI, not a vision analysis task. Any stem about producing new images belongs to the generative-AI domain, never to these five analysis types.

Master these six shapes and the input→output table from How they differ above, and you can resolve essentially any vision-solution-type stem by matching the requested output to the type that emits it.