AI-102 Azure AI Engineer Associate - Set 4 - Part 1

Correct Answer: Using efficient embedding models

The Read API is specifically designed for handling difficult handwriting like doctor's notes and uses asynchronous processing to efficiently handle large document volumes with parallel line processing.

Option b is incorrect because the OCR API uses an older recognition model and synchronous execution, making it unsuitable for large-scale processing of complex handwritten documents.

Option c is incorrect because image classification categorizes entire images into predefined classes rather than extracting text content from handwritten documents for digitization purposes.

Option d is incorrect because Custom Vision is designed for training custom classification or object detection models, not for extracting text content from handwritten medical prescriptions.

Correct Answer: Compact domain for edge devices

The Compact domain is specifically optimized for real-time classification on edge devices like mobile phones and IoT devices, making it ideal for on-device factory floor applications.

Option a is incorrect because while the General domain handles broad tasks, it is not optimized for the performance constraints and real-time requirements of mobile edge deployment.

Option b is incorrect because General A1 prioritizes high accuracy with slower inference speed, which contradicts the real-time processing requirements needed for mobile factory applications.

Option d is incorrect because the Retail domain is optimized for shopping items and clothing recognition, not industrial equipment identification in manufacturing environments.

Correct Answer: Landmark detection not supported in version

Version 4.0 of the Image Analysis SDK does not include landmark detection, which remains exclusive to the legacy version 3.2 alongside celebrity recognition features.

Option a is incorrect because version 4.0 explicitly supports object detection capabilities along with people detection, making this functionality fully available in the latest SDK release.

Option b is incorrect because generating captions and dense captions is one of the primary features available in version 4.0 of the Image Analysis SDK.

Option c is incorrect because smart cropping is actually included as a feature in version 4.0 of the Image Analysis SDK along with model customization capabilities.

Correct Answer: ML Assisted Labeling with Smart Labeler

Once sufficient data is labeled, ML Assisted Labeling automatically suggests labels for new images, significantly accelerating the annotation process for additional training data.

Option a is incorrect because Quick Training is a mode that speeds up the model training process itself, not the labeling or annotation of new images.

Option b is incorrect because Advanced Training improves model accuracy through longer training times but does not assist with labeling or annotating additional images in the dataset.

Option c is incorrect because the probability threshold determines prediction confidence requirements but does not provide any assistance in labeling or annotating new training images.

Correct Answer: People detection and object detection

People detection identifies and locates multiple individuals in the footage, while object detection finds and labels the products they interact with using bounding boxes.

Option b is incorrect because image classification assigns entire images to categories and facial detection identifies faces, neither providing the detailed location tracking needed for shopper interactions.

Option c is incorrect because semantic segmentation creates pixel-perfect masks which is unnecessarily detailed here, and OCR extracts text rather than detecting people or physical products.

Option d is incorrect because image analysis provides general descriptions and celebrity recognition identifies famous individuals, neither specifically tracking regular shoppers or their product interactions.

Correct Answer: Multiclass allows only single tag predictions

Multiclass classification is designed to assign only one tag per image, so it must choose either cardinal or blue jay, not both simultaneously.

Option a is incorrect because insufficient training data would result in poor accuracy or confidence scores, not the fundamental limitation of returning only one prediction per image.

Option b is incorrect because a high probability threshold would prevent low-confidence predictions from appearing but would not cause the model to select only one species.

Option d is incorrect because Quick Training affects training speed and slight accuracy differences but does not change the fundamental multiclass behavior of single-tag predictions.

Correct Answer: Read API for mixed content handling

The Read API supports PDF documents, handles both printed and handwritten text, and uses asynchronous processing ideal for processing thousands of large legal documents.

Option b is incorrect because the legacy OCR API only supports images not PDF files, and uses synchronous execution unsuitable for processing thousands of large documents.

Option c is incorrect because the describe image function generates human-readable descriptions of image content rather than extracting actual text from documents for legal review.

Option d is incorrect because Face API is specialized for facial detection and recognition, not for extracting text content or verifying handwritten signatures on documents.

Correct Answer: Caption and confidence score

The describe image in stream function returns a human-readable caption describing the image content and a confidence score percentage indicating the AI's certainty level.

Option a is incorrect because bounding boxes and object labels are outputs from object detection operations, not from the describe image function for general description.

Option b is incorrect because while tags may be generated, semantic segmentation masks are pixel-perfect outlines used in segmentation tasks, not description operations.

Option c is incorrect because pixel coordinates and color histograms are low-level image data not returned by describe image, which focuses on high-level semantic understanding.

Correct Answer: Adult content detection only in version

Adult content detection is a feature exclusive to version 3.2 of the Image Analysis SDK and is not available in the newer version 4.0.

Option a is incorrect because object detection is fully available in version 4.0 of the Image Analysis SDK, along with people detection capabilities.

Option b is incorrect because caption generation is one of the core features in version 4.0, and its availability is unrelated to content moderation functionality.

Option d is incorrect because OCR functionality for reading text from images is available in version 4.0 through the Read API integration.

Correct Answer: Semantic segmentation with pixel masks

Semantic segmentation creates pixel-perfect masks that outline the exact shape of objects, providing the precision needed for accurate vehicle size measurements unlike rectangular bounding boxes.

Option a is incorrect because image classification assigns entire images to vehicle type categories without providing any location information or shape outlines for measurement purposes.

Option b is incorrect because facial detection identifies and locates human faces, which is irrelevant to determining the outline and measurements of vehicle bodies.

Option c is incorrect because OCR extracts text characters from images like license plates but provides no information about the vehicle's physical shape or dimensions.

Correct Answer: Food domain with multiclass classification

The Food domain is specifically optimized for dishes and menus, while multiclass classification assigns each food photo to one specific dish category.

Option b is incorrect because while General A2 offers good accuracy, it is not optimized for food recognition, and object detection finds item locations rather than categorizing images.

Option c is incorrect because the Landmarks domain is designed for natural and artificial landmarks, not food items, and multilabel would allow multiple categories per image unnecessarily.

Option d is incorrect because the Retail domain is optimized for shopping items and clothing, not food dishes, making it suboptimal for restaurant menu categorization.

Correct Answer: Model B with higher recall

Recall measures sensitivity, indicating how many actual objects the model successfully finds, so Model B's ninety percent recall means it finds more relevant objects.

Option a is incorrect because precision measures exactness of predictions that were made, not the ability to find all relevant objects, which is measured by recall.

Option b is incorrect because the models have opposite strengths, with Model A being more precise but missing more objects, while Model B finds more objects overall.

Option d is incorrect because both models show reasonable performance levels, with scores in the seventy to ninety percent range indicating adequate training, just different strengths.

Correct Answer: Vision Studio wraps underlying APIs with UI

Azure AI Vision Studio is a no-code playground that provides a UI wrapper for various underlying APIs like Computer Vision API and Face API.

Option a is incorrect because Vision Studio is explicitly not a standalone SDK but rather a user interface that provides access to existing APIs without coding.

Option b is incorrect because Vision Studio is specifically designed as a no-code or low-code environment that eliminates the need for custom code deployment.

Option c is incorrect because Vision Studio does not replace the Computer Vision API but rather provides a graphical interface to access and test its functionality.

Correct Answer: OCR API with synchronous execution

The legacy OCR API with synchronous execution is ideal for small amounts of clean printed text, providing immediate results without the overhead of asynchronous processing.

Option b is incorrect because the Read API's asynchronous processing and advanced capabilities are designed for large documents or difficult handwriting, creating unnecessary complexity for simple labels.

Option c is incorrect because Custom Vision trains classification or object detection models for visual recognition tasks, not for extracting text characters from printed labels.

Option d is incorrect because semantic segmentation creates pixel-perfect masks around objects or regions but does not extract or digitize the actual text content from images.

Correct Answer: Multiclass classification for single status

Multiclass classification ensures each image receives exactly one tag, perfect for assigning a single status label like damaged or intact to each box.

Option a is incorrect because object detection finds and locates multiple objects within images using bounding boxes, rather than classifying the overall condition of boxes.

Option b is incorrect because multilabel classification allows multiple tags per image, which contradicts the requirement that each box receive only one status label.

Option d is incorrect because semantic segmentation creates pixel-perfect outlines of objects or regions but does not classify the overall status condition of boxes.

Correct Answer: Face API for facial recognition

The Face API is specifically designed and optimized for facial detection and recognition tasks, making it the appropriate service for identifying authorized personnel.

Option a is incorrect because Azure AI Custom Vision is for training custom classification or object detection models on user-provided data, not specialized facial recognition.

Option b is incorrect because Azure AI Vision provides general image analysis, descriptions, and tags but is not specialized for the facial recognition requirements of security systems.

Option d is incorrect because the OCR service extracts text from images like badges but does not provide facial recognition capabilities to identify specific individuals.

Correct Answer: SDK, matplotlib, and authentication credentials

You must install the vision SDK, import matplotlib for visualizing images or borders, and authenticate using CognitiveServicesCredentials with your key and endpoint.

Option a is incorrect because TensorFlow, Keras, and NumPy are general machine learning libraries not specifically required for using Azure's pre-built Computer Vision SDK.

Option c is incorrect because Docker and Kubernetes are containerization and orchestration tools for deployment, not prerequisites for basic SDK usage and image analysis.

Option d is incorrect because Flask and Django are web application frameworks for building websites, not required components for using the Computer Vision SDK directly.

Correct Answer: Multilabel classification for multiple tags

Multilabel classification allows a single image to receive multiple tags simultaneously, enabling the dress to be tagged as both red and formal.

Option a is incorrect because multiclass classification restricts each image to exactly one tag, preventing the simultaneous assignment of multiple attributes like color and style.

Option b is incorrect because object detection identifies and locates items within images using bounding boxes, rather than assigning multiple descriptive attribute tags.

Option c is incorrect because semantic segmentation creates pixel-perfect outlines of objects but does not assign multiple categorical tags or attributes to images.

Correct Answer: Parallel processing per line for speed

The Read API parallelizes tasks per line, making it significantly faster for processing large documents with varying quality and sizes compared to legacy sequential processing.

Option a is incorrect because synchronous execution is actually a characteristic of the legacy OCR API, while the Read API uses asynchronous processing for efficiency.

Option b is incorrect because celebrity recognition is an image analysis feature in version 3.2 SDK, completely unrelated to optical character recognition functionality.

Option d is incorrect because facial landmark detection is a feature of the Face API for facial analysis, not related to extracting text from documents.

Correct Answer: Insufficient data requires fifty images minimum

Custom Vision typically requires at least fifty images per tag for effective training, so forty images falls below the recommended minimum for achieving good accuracy.

Option a is incorrect because domain selection affects optimization for specific use cases, but using a suboptimal domain would not typically cause poor accuracy with adequate data.

Option b is incorrect because while Advanced Training improves accuracy over Quick Training, it cannot overcome fundamental data insufficiency when below the recommended image count.

Option d is incorrect because a low probability threshold would increase the number of predictions made regardless of confidence, not reduce overall model accuracy during evaluation.