Back to Azure AI Engineer Associate (AI-102)

Implement computer vision solutions

Analyze images and videos using Azure AI Vision, custom models, and Video Indexer.

Encompasses analyzing images including selecting visual features, detecting objects, generating tags, extracting and converting text using Azure Vision in Foundry Tools. Covers implementing custom vision models including image classification and object detection with proper labeling, training, evaluation, and deployment. Also includes analyzing videos using Azure AI Video Indexer and spatial analysis for detecting presence and movement.
5 minutes 5 Questions

Implementing computer vision solutions in Azure involves leveraging Azure AI Vision services to analyze, process, and extract meaningful information from images and videos. As an Azure AI Engineer, you need to understand several key components and services. **Azure AI Vision Service** is the prima…

Concepts covered: Labeling images for custom models, Selecting visual features for image processing, Detecting objects and generating image tags, Including image analysis features in requests, Interpreting image processing responses, Extracting text from images with Azure Vision, Converting handwritten text with Azure Vision, Choosing between classification and object detection, Training custom image models, Evaluating custom vision model metrics, Publishing and consuming custom vision models, Building custom vision models code first, Using Azure AI Video Indexer for insights, Using spatial analysis for presence detection

Test mode:
Start practice test
AI-102 - Implement computer vision solutions Example Questions

Test your knowledge of Implement computer vision solutions

Question 1

What does the OCCUPIED state refer to in Azure AI Vision spatial analysis presence detection operations?

Correct Answer: A zone status indicating one or more people are currently detected within the designated area boundaries

The OCCUPIED state in Azure AI Vision spatial analysis presence detection operations indicates that one or more people are currently detected within the designated area boundaries. This is the fundamental purpose of presence detection - to determine whether a defined zone or area contains people at any given moment.

In spatial analysis, zones are virtual boundaries defined within the camera's field of view. The system continuously monitors these zones and reports their status. When the spatial analysis algorithms detect one or more persons within a zone's boundaries, it transitions to the OCCUPIED state. Conversely, when no people are detected in the zone, it would typically be in an UNOCCUPIED or EMPTY state.

This binary or multi-state zone status is essential for various business applications such as:
- Monitoring meeting room occupancy
- Tracking retail space utilization
- Managing queue lengths
- Ensuring compliance with occupancy limits
- Optimizing facility usage

The other options describe concepts that are not related to the OCCUPIED state definition:

One option incorrectly suggests it relates to camera tracking capacity limits, which would be a technical constraint of the system rather than a zone status indicator.

Another option incorrectly associates it with movement pattern identification and analytics processing, which are separate analytical functions beyond basic presence detection.

The final option incorrectly defines it as relating to physical objects or furniture restricting space, which confuses the concept of zone occupancy by people with physical space constraints from inanimate objects.

Question 2

A government agency is modernizing its archive system by digitizing historical handwritten census records from the 1950s. The records are stored as black-and-white microfilm images that have been converted to digital format. The IT team has set up an Azure Computer Vision resource in the East US region and successfully tested the Read API with a few sample images. However, when they attempt to process a batch of 500 images programmatically, they encounter intermittent failures where some requests return an error indicating the service endpoint cannot be reached, while others succeed. The team has verified that their Azure subscription has sufficient quota, the API keys are valid, and the network connectivity is stable. Upon closer inspection, they notice that the endpoint URL they are using in their code is 'https://eastus.api.cognitive.microsoft.com/vision/v3.2/read/analyze'. What is the most likely cause of these intermittent connection failures?

Correct Answer: The endpoint URL format is incorrect and should use the region-specific format 'https://<resource-name>.cognitiveservices.azure.com/vision/v3.2/read/analyze' with the actual resource name

The correct answer is that the endpoint URL format is incorrect and should use the region-specific format with the actual resource name.

When you create an Azure Computer Vision resource, Azure assigns it a unique resource name and generates a specific endpoint URL in the format: https://.cognitiveservices.azure.com/. This is the correct endpoint format that should be used for all API calls.

The endpoint URL being used in the scenario (https://eastus.api.cognitive.microsoft.com/vision/v3.2/read/analyze) represents an older, generic regional endpoint format that may not reliably route to the specific resource instance. This explains why the team experiences intermittent failures - some requests might accidentally route to valid endpoints while others fail because this generic format doesn't consistently resolve to their specific resource.

The intermittent nature of the failures is a key indicator: if quota, authentication, or network issues were the problem, failures would be consistent rather than sporadic. The fact that some requests succeed while others fail points to an endpoint routing issue.

The other options are incorrect for the following reasons:

The suggestion about using a global endpoint with automatic failover is incorrect because Azure Cognitive Services doesn't provide such a global endpoint format. Resources are region-specific and must be accessed through their assigned endpoints.

The claim that API version v3.2 has been deprecated for batch processing is false. Version deprecation would result in consistent failures, not intermittent ones, and there is no such restriction on v3.2 for batch processing scenarios.

The suggestion to include the subscription key as a URL parameter is incorrect because Azure Cognitive Services authentication uses either the Ocp-Apim-Subscription-Key header or Azure AD tokens, not URL parameters. Additionally, this would cause consistent authentication failures, not intermittent connection failures.

Question 3

A wildlife conservation organization is developing a custom image classification model using Azure Custom Vision to identify endangered species from camera trap footage across 45 remote forest locations. They have successfully trained a model using 22,000 images collected over 18 months, achieving 93% accuracy during validation. The model was trained using the General (compact) domain to enable edge deployment on solar-powered devices with limited connectivity. Three months after deployment, field researchers report that the model performs excellently on species captured during daylight hours but struggles significantly with nighttime infrared images, which constitute 60% of actual camera trap footage. The original training dataset contained only 15% nighttime images. The organization has now collected 8,000 additional nighttime images with proper species labels. They need to update the deployed model across all 45 locations while maintaining the compact model size for edge deployment and preserving the strong daytime performance. What approach should the AI engineer take to incorporate the new nighttime data?

Correct Answer: Combine the original 22,000 images with the 8,000 new nighttime images to create a balanced dataset of 30,000 images, then retrain the model from scratch using the General (compact) domain to ensure optimal feature extraction across both lighting conditions and export the new iteration for edge deployment

The correct approach is to combine the original 22,000 images with the 8,000 new nighttime images to create a balanced dataset of 30,000 images, then retrain the model from scratch using the General (compact) domain.

This is the best solution because:

  1. Dataset Balance: The combined dataset will have approximately 35% nighttime images (8,000 nighttime + 3,300 original nighttime = 11,300 out of 30,000), which better represents the actual 60% nighttime usage pattern and is a significant improvement over the original 15%.

  2. Maintains Compact Format: Training from scratch with the General (compact) domain ensures the model remains deployable on edge devices with limited resources, which is critical for solar-powered devices in remote locations.

  3. Optimal Feature Learning: Retraining from scratch allows the model to learn optimal features across both lighting conditions simultaneously, rather than trying to patch on new features incrementally.

  4. Preserves Daytime Performance: Including all original daytime images ensures the model doesn't lose its strong daytime classification capabilities while gaining nighttime proficiency.

Why the other approaches are suboptimal:

The dual-model architecture approach would double the computational and storage requirements at each edge location, which contradicts the constraint of limited resources on solar-powered devices. Managing two models also adds complexity and potential failure points.

The incremental training approach using Quick Training is not a feature that Custom Vision provides for properly incorporating substantial new data with different characteristics. Custom Vision doesn't support true incremental learning that would preserve all previous knowledge while adapting to significantly different data distributions.

The cloud-based deployment approach abandons the edge deployment requirement entirely. Establishing satellite connectivity at 45 remote forest locations would be prohibitively expensive and unreliable, defeating the purpose of the original compact model design for edge deployment with limited connectivity.

More Implement computer vision solutions questions
554 questions (total)
Start 100 question test