I used a lot VGG network(with transfer learning) and then extract visual features from last layers to use them for similarity search. This approach works fine, however the VGG network is very slow and quite big. I’m looking for similar performance but much smaller/simpler/effective model which can be deployed on cpu.
There are some small but effective networks as Mobilenet V1/V2 which works great for classification tasks. But I don’t have good results with them in similarity search.
Do you know any other architectures which can be helpful?