Image-Similarity

# Image-Similarity **Show it a picture. It finds the ones that look alike, using Vision Transformer embeddings.**     

What it does

A Streamlit app for content-based image search. Upload a query image and the app ranks every image in the local images/ folder by visual similarity, using feature vectors from a pretrained ViT-B/16 model on TensorFlow Hub and cosine similarity to compare them.

• Upload a query image (JPG, JPEG, PNG, BMP, TIFF)
• Choose how many results to show (top-k, up to 10) and a minimum similarity threshold
• See the top matches with their scores, colour-coded high / medium / low
• Grow the dataset from the sidebar by uploading new images, then hit Refresh Dataset

The repo includes a small sample gallery of ~47 images in images/ so it works out of the box.

How it works

flowchart LR
    subgraph Index[On startup - cached]
        A[images/*] --> B[Resize 224x224<br/>scale to 0-1, force RGB]
        B --> C[ViT-B/16<br/>feature extractor]
        C --> D[(Embedding matrix)]
    end
    Q[Query upload] --> P[Same preprocessing]
    P --> V[ViT-B/16]
    V --> S{Cosine similarity}
    D --> S
    S --> R[Threshold + sort + top-k]
    R --> UI[Streamlit results grid]

1. The model sayakpaul/vit_b16_fe/1 is loaded once from TensorFlow Hub (st.cache_resource).
2. Every image in images/ is preprocessed and embedded; the result is cached with st.cache_data.
3. The query image goes through the same pipeline.
4. Scikit-learn's cosine_similarity scores it against all dataset embeddings.
5. Results above the threshold are sorted and the top-k are displayed. If nothing clears the threshold, the best matches are shown anyway.

Tech stack

• Model: Vision Transformer ViT-B/16 feature extractor (TensorFlow Hub)
• Framework: TensorFlow, TensorFlow Hub
• Similarity: scikit-learn cosine_similarity, NumPy
• Imaging: Pillow
• UI: Streamlit

Project structure

Image-Similarity/
├── app.py        # the whole app: model loading, embedding, search, UI
├── images/       # searchable image gallery (add your own here)
└── demo.png      # screenshot used above

Quickstart

git clone https://github.com/nabeelalikhan0/Image-Similarity.git
cd Image-Similarity

python -m venv .venv
# Windows: .venv\Scripts\activate    macOS/Linux: source .venv/bin/activate
pip install streamlit tensorflow tensorflow-hub scikit-learn numpy pillow

streamlit run app.py

On first launch the ViT weights are downloaded from TensorFlow Hub and every image in images/ is embedded, so the first load takes a moment. After that, the cache makes searches fast.

To use your own gallery, drop images into images/ (or upload them from the sidebar) and click Refresh Dataset.

Notes

• Embeddings are recomputed from the images when the cache is cleared; they are not saved to disk.
• Similarity is purely visual (global ViT features), so it matches on overall appearance and composition rather than on labels or text.