Image Similarity Search Pipeline using EvaDB on Images#

In this use case, we want to search similar images based on an image provided by the user. To implement this use case, we leverage EvaDB’s capability of easily expressing feature extraction pipeline. Additionaly, we also leverage EvaDB’s capability of building a similarity search index and searching the index to locate similar images through FAISS library.

For this use case, we use a reddit image dataset that can be downloaded from Here. We populate a table in the database that contains all images.

1. Connect to EvaDB#

import evadb
cursor = evadb.connect().cursor()

2. Register SIFT as Function#

    CREATE UDF IF NOT EXISTS SiftFeatureExtractor
    IMPL  'evadb/udfs/'

3. Search Similar Images#

To locate images that have similar appearance, we will first build an index based on embeddings of images. Then, for the given image, EvaDB can find similar images by searching in the index.

Build Index using FAISS#

The below query creates a new index on the projected column SiftFeatureExtractor(data) from the reddit_dataset table.

    CREATE INDEX reddit_sift_image_index
    ON reddit_dataset (SiftFeatureExtractor(data))

Search Index for a Given Image#

EvaDB leverages the ORDER BY ... LIMIT ... SQL syntax to retrieve the top 5 similar images. In this example, Similarity(x, y) is a built-in function to calculate distance between x and y. In current version, x is a single tuple and y is a column that contains multiple tuples. By default EvaDB does pairwise distance calculation between x and all tuples from y. In this case, EvaDB leverages the index that we have already built.

query = cursor.query("""
    SELECT name FROM reddit_dataset ORDER BY
    LIMIT 5

The DataFrame contains the top 5 similar images.

|             |
| reddit-images/g1074_d4mxztt.jpg |
| reddit-images/g348_d7ju7dq.jpg  |
| reddit-images/g1209_ct6bf1n.jpg |
| reddit-images/g1190_cln9xzr.jpg |
| reddit-images/g1190_clna2x2.jpg |

Check out our Jupyter Notebook for working example. We also demonstrate more complicated features of EvaDB for similarity search.