This repo contains the ml experiment that was used to find all the similar images in the database.
Please Read PDF document in the repo
All the .ipynb are named as:-
-
Avantari_TASK1_v1: Best performance, Used pretrained MobileNet, Imagesize 224x224x3
-
Avatari_TASK1_v2: Use of simple autoencoder, not the best
-
Avatari_TASK1_v2_3: Use of a bigger autoencoder without dropout layer, reconstruction better than previous version but performance for similarity check still not comparable to pre-trained model.
-
Avantari_AutoEncoder_V0.1: Have tried to create a better performing Autoencoder, And it thus perform better. Please have a look and comment.
- Avantari_TASK2_V1: Used pretrained MobileNet to extract feature then DBSCAN to cluster images.
https://towardsdatascience.com/building-a-similar-images-finder-without-any-training-f69c0db900b5