AI-900 Practice Question: Describe features of computer vision workloads on Azure

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Storing images in Azure Blob Storage with automatic tagging

  Why it's wrong here

  Storing images in Azure Blob Storage is a persistence operation that simply places bytes into a distributed object store, while automatic tagging is typically implemented with computer vision classification models that assign labels like 'dog' or 'beach' based solely on the image content. Multimodal embeddings, however, do not perform storage or generate discrete tags; they produce dense numeric vectors that enable semantic similarity search across image collections. Although you could store the embedding vectors in Azure Cognitive Search to support retrieval, the act of storage and tag generation itself is not what multimodal embeddings accomplish, making this option incorrect.

• ✓

  Enabling natural language image search and finding visually similar images using semantic understanding

  Why this is correct

  This is the core capability enabled by multimodal embeddings, which map images and text into a shared vector space where cosine similarity measures semantic meaning. For example, an embedding model like Azure AI Vision or OpenAI CLIP can encode the query 'red car on a road' and an image of a red car driving on a scenic highway, then return that image as a top hit even if the metadata or labels do not contain those exact words. The same semantic representation also allows finding images visually similar to a reference photo, because near-identical concepts have close vector positions. This combination of cross-modal text-to-image retrieval and image-to-image similarity is precisely what multimodal embeddings are built for.

• ✗

  Automatically resizing images for different screen sizes

  Why it's wrong here

  Automatically resizing images for different screen sizes is a deterministic, pixel-level image processing task that scales dimensions without interpreting the content of the image. Multimodal embeddings, in contrast, map images and text into a shared high-dimensional vector space to capture semantic relationships, such as matching a photo of a 'red car' to the text query 'red car on a road'. Resizing neither requires nor benefits from semantic embeddings because it does not involve understanding the meaning of the image, so it is not a capability of multimodal embeddings.

• ✗

  Detecting copyrighted images in user-uploaded content

  Why it's wrong here

  Detecting copyrighted images in user-uploaded content typically relies on perceptual hashing or fingerprinting that identifies exact or near-exact duplicate copies of known content, regardless of what the image semantically depicts. Multimodal embeddings are designed to return images that are conceptually similar to a text prompt or another image, so a substantially re-edited or semantically altered version of a copyrighted photo might still match conceptually but would evade a hash-based copyright check. Therefore, this task is not the intended use of multimodal embeddings; it is better served by content fingerprinting systems.

Go deeper

Related to this question