Image Similarity Ranking using Deep CNN

Goal: find top n and bottom n images most similar with respect to a query image

Results Visualization

• Example 1

• Example 2

Components

• a triplet sampling module
• 3 copies of a deep CNN image encoder to convert images into the corresponding latent vectors
• an image ranking layer; use max margin triplet loss measured in Euclidean distance in latent space
• a visualization module to plot the top 10 and bottom 10 similar images w.r.t. a query image