Logo Similarity Retrieval via Synthetic Data and Deep Metric Learning

This project presents a scalable, weakly-supervised pipeline for logo recognition and visual similarity search. The system is designed to handle large-scale, unlabeled logo datasets by leveraging synthetic data generation, multimodal embeddings, and deep metric learning using a Triplet Network.

🧠 Overview

The core idea is to learn an embedding space where similar logos are projected close to each other, enabling high-performance retrieval even for previously unseen logos.

The pipeline consists of the following major stages:

1. Synthetic Dataset Creation:
   Over 1.7 million logo images are generated along with text prompts describing their style and concept.

2. Multimodal Embedding Construction:
   Visual features are extracted via a frozen ResNet50; text prompts are embedded using MiniLM. Both are concatenated to form a 2432D vector.

3. Clustering:
   UMAP reduces the embedding to 600D, followed by HDBSCAN clustering to derive over 209,000 pseudo-categories as weak labels.

4. Triplet Network Training:
   Triplets are sampled using pseudo-labels. The model is trained to learn a 256D normalized embedding via Triplet Loss.

5. Inference and Retrieval:
   Embeddings from the trained model are queried against a FAISS index to retrieve similar logos.

🔧 Pipeline Architecture

Data Preparation Pipeline

Triplet Network Architecture

Inference-time Retrieval

---
config:
  theme: redux
---

flowchart TB
    %%== Simplified Inference Architecture: TripletNet Retrieval ==%%


    A@{shape: lean-r, label: "User Input:\nQuery Logo Image"}


    B@{shape: subproc, label: "Trained TripletNet\n(CNN + Projection + Norm)"}
    C@{shape: rect, label: "Query Embedding (Vector)"}

    D@{shape: cyl, label: "Embeddings Database\n(Precomputed Vectors)"}
    E@{shape: subproc, label: "FAISS Index"}

    C --> F@{shape: diamond, label: "Find Top-K Nearest Neighbors"}
    F --> G@{shape: rect, label: "Retrieve Matching Logos"}
    G --> H@{shape: curv-trap, label: "Display Results to User\n(Grid or Ranked View)"}
    H --> I@{shape: dbl-circ, label: "End of Inference"}


    A --> B --> C
    D --> E
    C --> E
    E --> F

  
    classDef inputStyle fill:#e1f5fe,stroke:#0288d1,stroke-width:1.5px;
    classDef modelStyle fill:#ede7f6,stroke:#512da8,stroke-width:1.5px;
    classDef embedStyle fill:#fff3e0,stroke:#f57c00,stroke-width:1.5px;
    classDef dbStyle fill:#f9f,stroke:#333,stroke-width:2px;
    classDef searchStyle fill:#f3e5f5,stroke:#7b1fa2,stroke-width:1.5px;
    classDef displayStyle fill:#fff8e1,stroke:#f9a825,stroke-width:2px;

    class A inputStyle;
    class B modelStyle;
    class C embedStyle;
    class D,E dbStyle;
    class F,G searchStyle;
    class H,I displayStyle;

Loading

📈 Training Stability

Loss curves during training reveal convergence behavior for each backbone:

ResNet50

EfficientNet-B0

VGG16

📝 Summary

This project demonstrates the potential of combining synthetic data, unsupervised clustering, and triplet-based metric learning to build a practical, label-free logo similarity system that can scale to millions of images.

For full documentation and academic report, see the docs/ directory.