Visual Search in Rust

Visual Search in Rust is a single responsibility server/library performing similar images queries. It works by extracting features using a selected deep learning model and indexing them using an approximate nearest neighbors algorithm.

Examples

Below are examples of search results using a dataset of ecommerce images. Each collection has about 500-600 images.

Features

• Ability to extract features from any ONNX model (https://github.com/onnx/models/tree/master/vision/classification)
• Image transformation pipeline written fully in Rust
• Supports indexing local image files (bytes) or remote (URL)
• Standalone server for image similarity search (using approximate nearest neighbors algorithm)
• Use as a server or as a library
• Multithreaded and async indexing
• Python SDK

See example how to use the SDK

How it works

visual-search wraps ONNX format and creates a structure that includes:

• Url of the model (in this case ONNX model from the Microsoft repository)
• Image transformation pipeline that is necessary to process the image
• Layer name to extract features from (it is almost always last but one layer)

As far as we know this structure should be able to define any model from the ONNX repository. From the model we extract image features and index them in a predefined collection of images.

let model_config = ModelConfig {
    model_name: "SqueezeNet".into(),
    model_url: "https://github.com/onnx/models/raw/master/vision/classification/squeezenet/model/squeezenet1.1-7.onnx".into(),
    image_transformation: TransformationPipeline {
        steps: vec![
            ResizeRGBImageAspectRatio { image_size: ImageSize { width: 224, height: 224 }, scale: 87.5, filter: FilterType::Nearest }.into(),
            CenterCrop { crop_size: ImageSize {width: 224, height: 224} }.into(),
            ToArray {}.into(),
            Normalization { sub: [0.485, 0.456, 0.406], div: [0.229, 0.224, 0.225], zeroone: true }.into(),
            ToTensor {}.into(),
        ]
    },
    image_size: ImageSize { width: 224, height: 224 },
    layer_name: Some("squeezenet0_pool3_fwd".to_string()),
    channels: Channels::CWH
}

Installation

From source:

1. Clone this repository
2. Run cargo build --release
3. Run server target/release/image-embedding-rust --config config/config.toml

For production remember to change the bearer token in config.toml

Benchmark

It takes 100 seconds to index 1000 images using MobileNetV2 backbone model using 4 workers.

Searching for a single image takes 150 milliseconds.

To do

• [ ] persistence (right now the server is fully in-memory)
• [ ] logging
• [ ] clean all warnings