elcaiseri / Siamese-Network

Image similarity estimation using a Siamese Network with a triplet loss

Siamese Network

This repository contains an example of using a Siamese Network with a triplet loss for image similarity estimation. A Siamese Network is a type of network architecture that contains two or more identical subnetworks used to generate feature vectors for each input and compare them.

Siamese Networks can be applied to different use cases, like detecting duplicates, finding anomalies, and face recognition.

This example uses a Siamese Network with three identical subnetworks. We will provide three images to the model, where two of them will be similar (anchor and positive samples), and the third will be unrelated (a negative example.) Our goal is for the model to learn to estimate the similarity between images.

For the network to learn, we use a triplet loss function. You can find an introduction to triplet loss in the FaceNet paper by Schroff et al,. 2015.

Siamese Networks can be applied to various use cases, such as detecting duplicates, finding anomalies, and face recognition.

Table of Contents

• Introduction
• Setup
• Load the Dataset
• Preparing the Data
• Setting up the Embedding Generator Model
• Setting up the Siamese Network Model
• Putting Everything Together
• Training
• Inspecting What the Network Has Learned
• Summary
• About

Introduction

A Siamese Network with a triplet loss function is used in this example to estimate the similarity between images. The triplet loss function is defined as follows:

L(A, P, N) = max(‖f(A) - f(P)‖² - ‖f(A) - f(N)‖² + margin, 0)

Setup

This example uses the Totally Looks Like dataset by Rosenfeld et al., 2018.

Load the Dataset

The dataset used in this example is named "Totally Looks Like." It contains images to create triplets (anchor, positive sample, and negative sample) for training.

Preparing the Data

Prepare the data for training the Siamese Network by creating triplets of images.

Setting up the Embedding Generator Model

Create a model that generates embeddings for each image in a triplet.

Setting up the Siamese Network Model

Construct the Siamese Network model using three identical subnetworks.

Putting Everything Together

Combine all the components to form the complete Siamese Network.

Training

Train the Siamese Network using the triplet loss function.

Inspecting What the Network Has Learned

Evaluate the network's performance and inspect the embeddings learned by the model.

Summary

A summary of the key points covered in the example.

About

This example demonstrates image similarity estimation using a Siamese Network with a triplet loss.

© 2023 elcaiseri