Image Similarity Search

Using FAISS index to reverse search images

Setup

• To run the repo create a python virtual environment and install the packanges present in req.txt. using the command python3 -m pip install -r req.txt.
• Create folders /index and /images, to save the training images and trained indexes.
• If the dataset is from s3 bucket the code can be used as is. If not source images from local directory using PIL library.

Working

• We encode the images using ResNet-18 trained on ImageNet1K, to encode the images.
• We then use this encodings to train the FAISS index. For more info on usage of FAISS refer https://www.pinecone.io/learn/series/faiss/faiss-tutorial/.
• Using this index we perform similarity search that can search more than a million records in fewer milli seconds.