Describe features of computer vision workloads on Azure practice questions

Optical Character Recognition (OCR)

Why wrong: Azure's OCR (Optical Character Recognition, provided by the Read API) is specialized for extracting printed or handwritten text from images, returning bounding boxes and the recognized text characters. It does not infer image semantics, scene context, or relationships between visual elements, so it would only capture any literal text painted on an artwork rather than describing its subject or style. For artworks without text, OCR would produce no useful caption at all.

Face API

Why wrong: The Face API performs face detection, recognition, and attribute analysis, such as estimating age, emotion, and facial landmarks. It is strictly scoped to human faces and will not generate meaningful output for landscapes, abstract art, or still-life paintings that lack faces. Even when a face appears in an artwork, the Face API merely labels facial attributes and does not construct a descriptive sentence about the overall image content.

Custom Vision (object detection)

Why wrong: Custom Vision's object detection feature requires you to upload and label training images to create a model that recognizes specific objects, outputting bounding boxes and class labels (e.g., 'cat', 'defect'). It does not generate natural-language descriptions or full sentences, and it would need extensive labeled data for every type of artwork in the museum's collection. This is overkill and misaligned with the general-purpose captioning task, so it is not suitable.

Object detection

Why wrong: Optical Character Recognition (OCR) extracts printed or handwritten text from an image and returns the text strings along with their spatial coordinates. It is valuable for digitizing documents, reading signs, or processing menus, but it cannot interpret or describe non-text visual content such as people, scenery, or actions. Since the goal is to generate a description of the image itself, OCR would be relevant only if the image contained meaningful text, and even then it would not describe the image's visual content.

Optical Character Recognition (OCR)

Why wrong: Face Detection locates human faces in an image and optionally returns attributes like age, emotion, or accessories, but it does not generate a general description of the entire scene. A social media image may contain no faces, or faces may be only a minor component, and the output is a set of bounding boxes and attribute values, not a sentence describing what is happening. Consequently, it cannot satisfy the requirement to produce automatic alt text that conveys the overall image content.

Face detection

Why wrong: Object Detection identifies individual objects within an image and surrounds each with a bounding box, a class name, and a confidence score. Although it could enumerate elements such as 'dog' or 'car,' it does not synthesize a narrative that explains relationships between objects, actions, or the overall setting. For a social media platform generating image descriptions, a list of detected objects without grammatical structure would not provide the fluent, complete sentence needed for accessibility.

Optical character recognition (OCR)

Why wrong: OCR via Azure AI Vision Read API extracts machine-readable text from product labels and signs, but it does not detect structural objects in a shelf scene. Even if the model reads a brand name, it cannot tell whether those words correspond to one box or a carton with many boxes, so it cannot reliably count product occurrences.

Prebuilt image analysis (Describe Image)

Why wrong: Prebuilt image analysis (Describe Image) relies on fixed, general-purpose models that produce tags or captions for common categories like grocery store but cannot be retrained for client-specific SKUs. Its object detection is limited to a set of generic item types (person, dog, tv), and it rarely provides a count of a single branded product, especially when multiple instances overlap.

Facial recognition

Why wrong: Facial recognition in Azure Face API is designed to detect, analyze, and identify human faces—locating eyes, nose, and face landmarks—not to identify packaged retail goods. It has no label classes for products such as cereal boxes and no object count mechanism, making it completely inappropriate for shelf monitoring.

Optical Character Recognition (OCR)

Why wrong: Optical Character Recognition is not relevant for manufacturing defect detection because its sole purpose is to extract text and digits from images, such as printed label, serial numbers, or barcodes. It performs no spatial interpretation of product surfaces and cannot distinguish a scratch from a normal texture variation. Since the company's goal is to identify physical flaws rather than read text, OCR would provide no useful output and is therefore an incorrect choice.

Image Classification

Why wrong: Image Classification assigns a single category label to an entire image or the dominant object in the frame; it cannot output bounding boxes or pixel masks to localize defects. If an image contains both a scratch and normal surface, the classifier must choose one overall label, losing all spatial information about where the flaw is or how many flaws exist. This makes it inappropriate for a use case that requires locating and counting multiple defects on a single product, so it is the wrong service.

Optical character recognition (OCR)

Why wrong: OCR, or optical character recognition, is wrong here because it only converts printed or handwritten text from images into machine-readable characters, such as reading a license plate number or a street sign. It has no mechanism for analyzing facial geometry, skin texture, or other biometric signals, so even if a person's face is clearly visible, OCR produces no information about their identity.

Image classification

Why wrong: Image classification is incorrect because it assigns one categorical label to an entire image, such as 'crowd,' 'beach,' or 'protest,' based on the dominant visual content. It does not locate individual people, separate faces from the background, or compare any biometric data, so it can never determine who in the crowd is a specific person.

Object detection

Why wrong: Object detection is wrong for this security need because while it can find people and draw bounding boxes around them, it treats each person as an instance of the generic class 'person' and outputs only coordinates and a confidence score. It performs no facial analysis or identity matching, so it cannot distinguish one known individual from another; face identification requires a dedicated recognition step on top of detection.

Azure AI Language

Why wrong: Azure AI Language is incorrect because it is a natural language processing service focused on text analysis, including sentiment, key phrases, and named entities, but it does not include optical character recognition. It can process text once it has been digitized, but it cannot read text from scanned invoice images. Therefore it cannot extract text directly from scanned invoices or receipts.

Custom Vision

Why wrong: Custom Vision is incorrect because it is designed for training custom image classification and object detection models, such as identifying whether an image contains a receipt, but it does not output the text content from the image. It might be used to categorize or locate documents, but it has no OCR text extraction capability. So it cannot extract the text from a scanned invoice or receipt.

Creating image files from raw binary data uploaded to Azure Blob Storage

Why wrong: Incorrect. Converting raw binary data in Azure Blob Storage into image files is a file-format or deserialization task handled by storage SDKs, client libraries, or rendering tools; it does not involve the Florence vision-language model. Even if a blob holds valid image bytes, decoding those bytes to a viewable file does not require AI inference, whereas the advanced Azure AI Vision capabilities extract semantic meaning from the decoded image. Because no understanding or language grounding is performed, this is not a Florence capability.

Generating high-resolution versions of low-resolution input images

Why wrong: Incorrect. This option describes a super-resolution or image-enhancement operation that alters pixel data to increase resolution, which is a classic computer-graphics/image-processing task rather than a semantic vision-language skill. Florence-based Azure AI Vision analyzes image content—identifying what is present, where it is, and how it relates to text—but it does not upscale or synthesize new high-resolution output images. Nothing in dense captioning, grounded detection, or image-text search changes the dimensions or fidelity of the source image.

Automatically generating training image variations through data augmentation

Why wrong: Incorrect. Data augmentation—flipping, rotating, cropping, or recoloring source images to create new training variants—is a machine-learning model-training technique, not a runtime capability offered by Azure AI Vision's Florence foundation model. The advanced vision APIs are pretrained, zero-shot multimodal models that perform inference: they read images and produce language or grounding metadata directly, rather than synthesizing additional example images for a pipeline. Confusing 'generation' here is the trap: Florence's generative aspect produces text (captions/answers), never training images.

AI systems that replace radiologists by autonomously making all diagnoses from scans

Why wrong: Deploying AI to autonomously replace radiologists is categorically wrong because current approved medical imaging AI - including Azure AI Health Insights radiology capabilities - provides decision support, not independent clinical judgment. Regulatory frameworks like FDA and MDR require human oversight and accountability for diagnoses, and the models cannot adequately reason over the whole patient context or communicate uncertainty. Thus these systems assist, rather than supersede, radiologists.

Storing and managing medical images in Azure Blob Storage with HIPAA compliance

Why wrong: Storing images in Azure Blob Storage with HIPAA compliance is data-platform infrastructure and does not apply intelligence to the image content itself. Although secure storage is a prerequisite for medical imaging workloads, it performs no inference, abnormality detection, or clinical decision support, and it is not considered an AI system. This option describes the underlying data layer, not the AI analysis capability the question is seeking.

Automatically scheduling patient appointments based on medical image analysis results

Why wrong: Appointment scheduling is an operational workflow task that could be triggered after an AI finding, but it does not itself analyze radiology or pathology images. Medical imaging AI in Azure AI Health Insights extracts clinical findings from pixel or tissue data, whereas scheduling relies on calendars, patient availability, and queue priorities. Confusing downstream automation with the actual image-interpretation capability would misclassify the purpose of this AI workload.

Free: Designed for production workloads with SLA.

Why wrong: The Free tier does not support production workloads, nor does it offer an SLA, so this statement is incorrect. Production-grade capabilities, including the 99.9% availability SLA and scaling options, are characteristics of the Standard tier (S0). The Free tier's lack of an SLA means any downtime cannot be claimed for service credits, which is why it should never be used for production.

Standard: Limited to evaluation and testing, not for production.

Why wrong: The Standard tier is not limited to evaluation and non-production environments—in fact, it is the opposite. Evaluation and low-volume testing are intended for the Free tier, while Standard is explicitly built for production with higher transaction limits and an SLA. Describing Standard as evaluation-only confuses its role with that of the Free tier and would lead to incorrect architectural decisions.

Enterprise: High throughput, advanced features, dedicated capacity.

Why wrong: The Enterprise tier is intended for high-throughput, mission-critical scenarios that require dedicated capacity and advanced features. It offers superior scalability, lower latency, and additional enterprise-grade capabilities such as private endpoints, virtual network integration, and enhanced compliance controls. This tier is selected when the application demands guaranteed performance and isolation beyond what the Standard tier can provide.

Enterprise: Free tier with limited transactions.

Why wrong: The Enterprise tier is not a free tier with limited transactions; that description applies only to the Free tier (F0). Enterprise-grade services are priced for premium, dedicated infrastructure and are chosen for demanding, high-volume workloads. This statement reflects a common misunderstanding of the tier hierarchy, where a higher tier is mistakenly presumed to offer the same constraints as a lower-cost tier.

Object detection

Why wrong: Object detection is not the right fit because it provides a bounding box and a class label for each detected object, and it is designed to find multiple objects within one image. The transportation scenario only asks whether the vehicle type can be identified from the whole image, not where the vehicle is located or how many vehicles appear. Choosing detection would add unnecessary spatial output and require more complex post-processing to collapse the results into a single image-level answer.

Optical Character Recognition (OCR)

Why wrong: OCR is not relevant because it extracts text characters from images, such as license plate numbers, road signs, or printed labels, rather than identifying physical objects by their shape and appearance. While OCR might read a logo or text on a vehicle, it does not understand the vehicle's body type, number of wheels, or other visual characteristics needed to classify it. The intended use case is visual recognition of vehicle morphology, not text extraction.

Semantic segmentation

Why wrong: Semantic segmentation is unsuitable because it applies a class label to every pixel in the image, generating a dense segmentation map rather than a single overall label. This pixel-level detail is useful for autonomous driving or lane detection, but it is far more granular than a transportation company's need to simply identify the type of vehicle shown. Segmentation also emphasizes spatial boundaries and scene composition, so it would not directly answer whether an image is a certain vehicle category without additional aggregation logic.

B. Use the Optical Character Recognition (OCR) API

Why wrong: The OCR API (Read API in Azure AI Vision) is designed to extract printed or handwritten text from images using optical character recognition; it sequences characters and words, not visual surface anomalies. Surface defects like scratches, pits, or color variations have no textual representation, so the OCR service would only return text if your parts were printed with labels or part numbers. It has no concept of texture, geometry, or flaw detection, so using it to find a dent would be a categorical task mismatch.

C. Use the Describe Image API (Image Captioning)

Why wrong: The Describe Image API (also called Image Captioning within Azure AI Vision) uses a neural encoder-decoder architecture to produce a natural-language sentence summarizing the whole scene, such as 'a dark metal object on a white background.' It does not output structured classifications, confidence scores for defect categories, or bounding boxes; it merely describes overall content for use cases like alt-text generation or scene understanding. A caption like 'a scratched part' is not a deterministic defect inspection result, and the model is not trained to recognize manufacturing anomalies.

D. Use the Face API

Why wrong: The Face API (part of Azure AI Vision) is specialized for human faces: it detects frontal and side faces, identifies facial landmarks, and performs tasks like identity verification, attribute analysis (age, emotion), and liveness detection. The underlying models are fine-tuned on face datasets and rely on facial structure such as eyes, nose, and mouth; a metal part or plastic housing contains no face, so the API would not extract any meaningful defect signal. It is not a general-purpose visual anomaly detector and cannot be repurposed for surface inspection.

Image classification

Why wrong: Image classification assigns a single categorical label (for example, 'shipping label' or 'document') to the entire image based on its overall visual content. It uses object-level patterns to distinguish image types, but it does not localize or transcribe individual characters, words, or lines. Classifying a label as a 'shipping label' would not reveal the handwritten destination, so it is insufficient for extracting shipping details.

Object detection

Why wrong: Object detection draws bounding boxes around physical entities in an image, such as parcels, barcodes, pallets, or people, and labels each with a class. It can potentially locate the region of a shipping label, but it does not perform any text recognition or character decoding. Since the logistics company needs the actual handwritten content, object detection cannot turn those bounding boxes into usable address data.

Face detection

Why wrong: Face detection identifies the spatial locations and bounding boxes of human faces in an image, optionally returning attributes like age or emotion, but it cannot interpret any written content. Handwritten shipping labels do not typically contain faces, and even if one appeared, face detection would ignore the text. Therefore, it provides no mechanism to extract the address or recipient information required by the logistics company.

Image Classification

Why wrong: Image classification assigns one or more pre-defined class labels to an entire image, such as 'cardboard box on conveyor belt' or 'labeled package', but it does not inspect the individual pixels that form each character. It uses a trained model that scores the image against a set of categories, and the output is only the category name and confidence score. Consequently, even if the model correctly recognizes that the image shows a labeled box, it cannot produce the exact serial number string from that label.

Object Detection

Why wrong: Object detection returns the locations of objects in an image by drawing bounding boxes around recognized entities and outputting the object class, along with confidence scores. While it could find the box or even the label as an object, its output consists of coordinates and class names ('box', 'label'), not the textual characters inside those boxes. It cannot decode the OCR-like glyphs in the serial number, so it is unsuitable for extracting the identifier.

Image Analysis (captioning and tagging)

Why wrong: Image Analysis with captioning and tagging generates human-friendly descriptions of the image and returns semantic tags such as 'box', 'text', 'conveyor belt', and 'warehouse', rather than the discrete characters printed on the label. Captioning produces a single-sentence summary like 'a cardboard box on a conveyor belt' and is designed for accessibility or search, not for character-level transcription. This option fails the requirement because it never segments the label into letters and digits, so it cannot supply the exact serial number needed for inventory tracking.

Azure Computer Vision - Image Analysis

Why wrong: Azure Computer Vision - Image Analysis is a general-purpose visual understanding service. It can detect faces in an image and return attributes like age, emotion, or whether a person is wearing glasses, but it does not create an identity signature or compare two separate face images to determine whether they belong to the same person. Therefore, it lacks the core 1:1 face verification capability required here.

Azure Custom Vision

Why wrong: Azure Custom Vision is a model training service for building custom image classifiers or object detectors from your own labeled datasets. While you could theoretically train a model to recognize a specific person's face, it does not offer a prebuilt, turnkey face verification endpoint with facial landmark detection, similarity scoring, and confidence thresholds. It also requires extensive training data and management, making it unsuitable for out-of-the-box identity verification against an ID photo.

Azure Form Recognizer

Why wrong: Azure Form Recognizer is a document intelligence service that extracts text, tables, and key-value pairs from forms and documents using OCR and layout analysis. It processes document structure and printed or handwritten content but does not analyze facial features or compare images of people. Hence, it is completely unrelated to the face verification requirement.

Image classification

Why wrong: Image classification assigns a single category label to the entire image, such as 'warehouse shelf' or 'containers', without any spatial localization of individual objects. It treats the whole image as one unit and cannot answer 'where' a box is. Therefore it fails the core requirement of locating and framing each box, which demands object-level spatial output.

Semantic segmentation

Why wrong: Semantic segmentation classifies every pixel in the image into a semantic category, producing a dense pixel-level mask where each box's exact silhouette is colored or labeled. While this gives far more precise boundaries than a rectangle, it does not naturally return a bounding box, which is what the requirement asks for. In a warehouse setting, it would be overkill and require additional post-processing to derive rectangular frames from the segmentation mask.

Optical Character Recognition (OCR)

Why wrong: Optical Character Recognition (OCR) extracts machine-encoded text from images, typically using detection and recognition models to find text regions and transcribe characters. It is entirely unsuited to locating physical objects such as boxes because it operates on text glyphs and their positions, not on object instances. Even if a box has a label with text, OCR would locate the text, not the box's extent.

Object detection

Why wrong: Object detection identifies objects and localizes them with rectangular bounding boxes, along with a class label and confidence score. It is well suited for counting equipment or detecting its presence, but a bounding box always includes some background and may merge multiple nearby objects into one region. It does not provide pixel-level outlines, so it cannot accurately measure the exact footprint or shape of construction assets.

Image classification

Why wrong: Image classification assigns a single class label to the entire image, such as 'construction site' or 'aerial view', without locating any objects. It discards all spatial information, so it cannot determine where equipment appears, how many objects are present, or what their boundaries are. This makes it too coarse for drone-based survey tasks that need to isolate and analyze individual assets.

Optical Character Recognition (OCR)

Why wrong: OCR extracts text from images by recognizing printed or handwritten characters, and it returns the text content and its location. In a drone survey, OCR could read a serial number on a piece of equipment or a sign on a building, but it has no mechanism to identify or segment non-text objects like excavators or cranes. Therefore it is fundamentally unsuitable for outlining equipment or understanding the spatial layout of a construction site.

Image Classification

Why wrong: Image Classification, including the Describe and Tag features in Azure Computer Vision, assigns high-level labels or captions to the entire image based on dominant scene-level content, such as 'city', 'outdoor', or 'appliance'. It does not locate or identify specific brand entities within an image; multiple logos would be collapsed into a single scene label, and no bounding-box or brand-name output is provided. This holistic, non-localized approach makes it unsuitable for detecting brands.

Object Detection

Why wrong: Object Detection in Azure Computer Vision localizes generic items such as persons, cars, or bicycles by drawing bounding boxes and assigning a label from a fixed set of common object categories. A trademarked logo is not part of that generic object vocabulary, so the model would miss it or label it with a vague, incorrect category. Brand Detection is a separate, purpose-built model optimized specifically for logo recognition, which is why Object Detection is not the right choice.

Optical Character Recognition

Why wrong: Optical Character Recognition (OCR) is designed to extract printed or handwritten text sequences from an image, such as street signs or documents, and outputs the raw text along with its location. It does not understand the semantic meaning of a logo: a logo is a graphical design, potentially text-free, and OCR would only capture any embedded letters without identifying the brand. Even when a brand name appears in the logo text, OCR provides characters rather than brand identity, so it is insufficient for brand monitoring.

Object Detection

Why wrong: Object Detection predicts a bounding box and class label for each recognized object, such as 'sofa' or 'window,' but every box is a coarse axis-aligned rectangle. Walls and doorways are long, thin, irregular structures that do not fit inside tight rectangles, and overlapping boxes give no information about adjacency or connectivity between spaces. The result lacks the per-pixel boundaries required to trace rooms accurately.

Image Classification

Why wrong: Image Classification assigns a single label or a set of class probabilities to the whole image, answering 'what is in this photo?' rather than 'where are the structural elements?' It might correctly call a photo 'living room,' but it cannot output pixel coordinates, wall lengths, or doorway positions. Without any spatial output, it cannot generate a measurable floor plan layout.

Optical Character Recognition (OCR)

Why wrong: OCR (Optical Character Recognition) detects and transcribes text characters within an image, returning strings such as room numbers, dimensions, or labels on a blueprint. Although a floor plan might contain text that OCR could read, the model has no understanding of walls, doors, or furniture shapes. Extracting text cannot produce a geometric or structural representation of the apartment layout.

Azure Face API

Why wrong: Azure Face API is incorrect for this task because it is a specialized service for detecting, analyzing, and recognizing human faces in images. It performs operations like facial attribute extraction, emotion detection, and face verification, but it has no OCR capabilities to read or extract text from documents. Scanned invoices contain no faces, so this service would be completely unresponsive to text extraction requirements.

Azure Custom Vision

Why wrong: Azure Custom Vision is incorrect because it is a model-training service for image classification and object detection; it enables users to build custom models that classify entire images or identify object locations, but it does not perform text extraction. It has no built-in OCR capability and cannot read or understand the content of a scanned invoice. Any text extraction would require an external OCR engine, making this service unsuitable as the primary solution for invoice text extraction.

Azure Video Indexer

Why wrong: Azure Video Indexer is incorrect because it is designed to process video content, extracting insights such as spoken language via speech-to-text, on-screen faces, and scene descriptions. Although it can perform OCR on text visible within video frames, it is not meant for static scanned invoices and lacks the invoice-specific layout analysis and structured data extraction that document-focused OCR tools provide. Using it would require uploading invoices as video, which is neither practical nor optimized for this scenario.

Azure Face API

Why wrong: Azure Face API specializes in detecting, identifying, and analyzing human faces, including attributes like emotion and age, as well as face comparison for identity verification. It is not a general-purpose object classifier and cannot recognize bottle caps or evaluate manufacturing quality. Using Face API for this task would be a domain mismatch, as it is optimized for facial data, not arbitrary industrial objects.

Azure Form Recognizer

Why wrong: Azure Form Recognizer is a document-processing service that extracts structured data from forms, invoices, receipts, and similar documents using layout analysis and key-value pair extraction. It does not perform object classification or detect physical defects in manufacturing items. Since the task is to classify bottle caps visually rather than parse a document, Form Recognizer is not applicable.

Azure OCR (Read API)

Why wrong: Azure OCR (Read API) is designed specifically for optical character recognition, pulling printed and handwritten text out of images and converting it into machine-readable text. It cannot judge the condition of a bottle cap or distinguish acceptable from defective items, as it ignores all non-text visual features. While it might read text printed on a label, it provides no information about the cap's physical integrity.