As the other comment suggests, you can use some kind of dense vector representation to search for nearest neighbours. I think the most effective method would be to use the latent vectors learned by an autoencoder. There is a great tutorial on how you could achieve this.