A Beginner’s Face-Image Similarity Demo With OpenCV and pgvector

Matthew Berman demonstrates a small image-matching pipeline in Python: detect faces in a group photo, turn the crops into vectors, store them in PostgreSQL and search using a different photo. The January 2024 video is sponsored by Aiven and follows its [published tutorial](https://aiven.io/developer/find-faces-with-pgvector). The accompanying [Colab notebook](https://colab.research.google.com/drive/19ulJqMQqk4PfcTx1v3C3cxjvzokrKgZS?usp=sharing) contains the demonstration.

The example assembles existing pretrained components, rather than training a face-recognition model from scratch. OpenCV detects candidate faces; an image embedding model encodes the crops; pgvector orders stored vectors by distance from a query image. The result is a similarity ranking, not a verified identity.

To introduce vectors, Berman compares the sentences “I love parks” and “I love croissants” using binary indicators for whether each word occurs. Both share two words and differ on the third. This is a simple bag-of-words illustration of numerical comparison, not how modern learned image embeddings are necessarily constructed.

Distances between vectors give a way to order items by similarity. The actual image pipeline uses a learned representation of visual content, rather than those four word-presence flags. Similarity depends on the model, preprocessing and distance measure; a short vector explanation is only an intuition for the technique.

In Colab, Berman installs the libraries and loads a pretrained OpenCV Haar cascade for frontal-face detection. He downloads the `haarcascade_frontalface_default.xml` classifier, uploads it into the notebook environment and supplies its file path. He then uploads a group photo of the cast of *It’s Always Sunny in Philadelphia* as the source image.

The notebook reads the image, prepares a grayscale version and calls `detectMultiScale` to find face rectangles. In the demonstration, Berman changes `minNeighbors` to 5 and uses a minimum size of 100 by 100 pixels. OpenCV defines `minNeighbors` as the number of neighboring candidate rectangles required to retain a detection, not a count of nearby people.

For each detected rectangle, the code crops the image and saves a JPEG in a `stored-faces` folder. Berman first creates that folder, reruns the cell and finds five crops. He notes that detection is imperfect and that changing parameters can affect results; missed faces cannot be returned by later searches.

The next stage is a PostgreSQL database with the pgvector extension. Aiven, the sponsor, provides the hosted service in the demonstration. Berman creates a project and a PostgreSQL service using the trial credits available at the time; trial offers and pricing can change, so the video’s 2024 figures should not be assumed current.

Once the service is running, Berman connects with `psql`, enables the extension using `CREATE EXTENSION vector;`, and creates a `pictures` table with a filename and a 768-dimensional vector column. The dimension matches the embedding output in the example. A service URI contains credentials and should be kept out of shared notebooks, logs and screenshots.

Back in Colab, the code loads each cropped face, uses `imgbeddings` to produce an image vector, and inserts the filename and vector into PostgreSQL. The Aiven tutorial identifies the embedding approach as OpenAI CLIP-based image recognition. A general image embedding is not by itself a validated biometric identification system.

Berman uploads a separate image of actor Charlie Day as the query, loads it and calculates its embedding with the same tool. Using compatible embeddings for both the stored crops and the query is necessary for the distance comparison to be meaningful.

The query orders the stored records by vector distance and returns the closest candidates. In this demonstration, the result corresponds to the same actor in the original group image. The source tutorial uses Euclidean distance via pgvector; nearest-neighbor ordering still returns a closest item even when no correct match is present unless an appropriate rejection rule is added.

A successful match in one small example does not establish a reliable accuracy rate, identity verification, or performance across lighting, pose and different people. Face detection can miss or falsely locate faces, and image similarity can produce false matches. Human review and a representative evaluation are needed before relying on the output.

Berman encourages viewers to extend the example and thanks Aiven for sponsoring the video. He mentions a referral offer and credits available in 2024; check current terms before provisioning a service. Images and face-derived vectors can be sensitive, so use images with appropriate permission, protect access and delete stored material when no longer needed.