Do image retrieval using multimodal embeddings (version 4.0)

Artikkeli
10/16/2024

The Multimodal embeddings APIs enable the vectorization of images and text queries. They convert images to coordinates in a multi-dimensional vector space. Then, incoming text queries can also be converted to vectors, and images can be matched to the text based on semantic closeness. This allows the user to search a set of images using text, without the need to use image tags or other metadata. Semantic closeness often produces better results in search.

The 2024-02-01 API includes a multi-lingual model that supports text search in 102 languages. The original English-only model is still available, but it cannot be combined with the new model in the same search index. If you vectorized text and images using the English-only model, these vectors won’t be compatible with multi-lingual text and image vectors.

Important

These APIs are only available in certain geographic regions. See Region availability.

Prerequisites

Azure subscription - Create one for free
Once you have your Azure subscription, create a Computer Vision resource in the Azure portal to get your key and endpoint. Be sure to create it in one of the supported geographic regions: see Region availability.
- After it deploys, select Go to resource. Copy the key and endpoint to a temporary location to use later on.

Try out Multimodal embeddings

You can try out the Multimodal embeddings feature quickly and easily in your browser using Vision Studio.

Important

The Vision Studio experience is limited to 500 images. To use a larger image set, create your own search application using the APIs in this guide.

Try Vision Studio

Call the Vectorize Image API

The retrieval:vectorizeImage API lets you convert an image's data to a vector. To call it, make the following changes to the cURL command below:

Replace <endpoint> with your Azure AI Vision endpoint.
Replace <subscription-key> with your Azure AI Vision key.
In the request body, set "url" to the URL of a remote image you want to use.
Optionally, change the model-version parameter to an older version. 2022-04-11 is the legacy model that supports only English text. Images and text that are vectorized with a certain model aren't compatible with other models, so be sure to use the same model for both.

curl.exe -v -X POST "<endpoint>/computervision/retrieval:vectorizeImage?api-version=2024-02-01&model-version=2023-04-15" -H "Content-Type: application/json" -H "Ocp-Apim-Subscription-Key: <subscription-key>" --data-ascii "
{
'url':'https://learn.microsoft.com/azure/ai-services/computer-vision/media/quickstarts/presentation.png'
}"

To vectorize a local image, you'd put the binary data in the HTTP request body.

The API call returns a vector JSON object, which defines the image's coordinates in the high-dimensional vector space.

{ 
  "modelVersion": "2022-04-11", 
  "vector": [ -0.09442752, -0.00067171326, -0.010985051, ... ] 
}

Call the Vectorize Text API

The retrieval:vectorizeText API lets you convert a text string to a vector. To call it, make the following changes to the cURL command below:

Replace <endpoint> with your Azure AI Vision endpoint.
Replace <subscription-key> with your Azure AI Vision key.
In the request body, set "text" to the example search term you want to use.
Optionally, change the model-version parameter to an older version. 2022-04-11 is the legacy model that supports only English text. Images and text that are vectorized with a certain model aren't compatible with other models, so be sure to use the same model for both.

curl.exe -v -X POST "<endpoint>/computervision/retrieval:vectorizeText?api-version=2024-02-01&model-version=2023-04-15" -H "Content-Type: application/json" -H "Ocp-Apim-Subscription-Key: <subscription-key>" --data-ascii "
{
'text':'cat jumping'
}"

The API call returns a vector JSON object, which defines the text string's coordinates in the high-dimensional vector space.

{ 
  "modelVersion": "2022-04-11", 
  "vector": [ -0.09442752, -0.00067171326, -0.010985051, ... ] 
}

Calculate vector similarity

Cosine similarity is a method for measuring the similarity of two vectors. In an image retrieval scenario, you'll compare the search query vector with each image's vector. Images that are above a certain threshold of similarity can then be returned as search results.

The following example code calculates the cosine similarity between two vectors. It's up to you to decide what similarity threshold to use for returning images as search results.

C#
Python

public static float GetCosineSimilarity(float[] vector1, float[] vector2)
{ 
    float dotProduct = 0; 
    int length = Math.Min(vector1.Length, vector2.Length); 
    for (int i = 0; i < length; i++) 
    { 
        dotProduct += vector1[i] * vector2[i]; 
    } 
    float magnitude1 = Math.Sqrt(vector1.Select(x => x * x).Sum());
    float magnitude2 = Math.Sqrt(vector2.Select(x => x * x).Sum());
    
    return dotProduct / (magnitude1 * magnitude2);
}

import numpy as np

def cosine_similarity(vector1, vector2):
    return np.dot(vector1, vector2) / (np.linalg.norm(vector1) * np.linalg.norm(vector2))

Next step

Image retrieval concepts

Jaa