CNN Reverse Image Search

This project was inspired by pyimagesearch's tutorial on building an image search engine by using the images' histogram as a feature vector and then measure the chi-squared distance between a given query image and the indexed database.

From my studies on deep learning I knew that a CNN also creates a feature vector, which is then given to classification layers. I imagined I could create a similar algorithm using a CNN's feature vector instead of an histogram, as it will contain more information that just the color. This is the result.

The algorithm

1. With getFeatures.py, use a ResNet50 model to extract a feature vector for each image and dump them to a file.
2. With query.py, extract the features of a given picture and send them to tools/Search.py to measure the euclidean distance between the query vector and the vectors on the dataset. Sort them and show the five closest results.

How to use

To run this script first install the dependencies

pip install -r requirements.txt

Then drop your image database on an "images" folder and run getFeatures.py to index all the pictures. You can now send queries to query.py --images [path-to-image].