This project implements a fashion runway search engine that allows users to search for runway images using text queries (e.g., "dress with floral pattern"). The system leverages the CLIP model for semantic text-to-image matching and FAISS for efficient similarity search. Background-removed images are used for feature extraction to focus on clothing, while original runway images are displayed in the results.

The search functionality combines CLIP (Contrastive Language-Image Pretraining) and FAISS (Facebook AI Similarity Search) to deliver fast and accurate results. Here's how it works:

1. Image Feature Extraction:

   • Images with removed backgrounds are processed using the CLIP model (CLIP-ViT-L-14-laion2B-s32B-b82K).
   • Each image is converted to a 768-dimensional feature embedding, normalized for consistent similarity comparisons.
   • Features are extracted once and cached (features.npy, paths.pkl) to avoid redundant computation.

2. FAISS Indexing:

   • The extracted image features are added to a FAISS IndexFlatL2 index, which uses L2 distance for efficient nearest-neighbor search.
   • The index enables fast retrieval of the most similar images to a given query, even with large datasets.

3. Query Processing:

   • A user submits a text query via the FastAPI /search endpoint.
   • The query is encoded into a 768-dimensional CLIP embedding using the CLIP model's text encoder and normalized.

4. Similarity Search:

   • The query embedding is searched against the FAISS index to find the top k closest image embeddings.
   • FAISS returns the indices of the matching images, which are mapped to their corresponding image paths.

5. Result Delivery:

   • The matching image paths (from background-removed images) are converted to URLs pointing to the original runway images (in fw25).
   • The FastAPI server returns a JSON response with a list of image URLs.
   • The frontend fetches and displays the original runway images.

• CLIP: Encodes text and images into a shared embedding space, enabling semantic matching between natural language queries and visual content.
• FAISS: Provides highly efficient similarity search, making it scalable for large image collections.
• Background Removal: Using background-removed images for feature extraction ensures the model focuses on clothing, avoiding irrelevant matches based on runway backgrounds.

• Query: "fur coat"
• Process:
  • CLIP encodes the query into a feature vector.
  • FAISS finds the images with the closest feature vectors.
  • The system returns URLs to the original runway images.
• Result: A list of runway images featuring fur coats.