Content-Based Image Retrieval in P2P Networks with Bag-of-Features

Recently, the Bag-of-Features (BoF) model has emerged as a popular solution to scalable content-based image retrieval (CBIR), due to great success of the Bag-of-Words (BoW) model in textual information processing. While most of the existing algorithms on CBIR in P2P networks focus on indexing high dimensional low level features, we propose to address such an issue by employing the BoF model. However, it is not straightforward due to the fact that the BoF model depends on a global codebook and it is very challenging to create and maintain such a global codebook across the whole P2P network. We design a novel online sampling mechanism to create a codebook with low network cost. Since the number of features in each image is large, compared to a text query generally consisting of several keywords, information exchange between nodes for each query image generates high network cost. In order to further reduce the network cost, we implement two static index pruning policies to limit the document length and the returned term weights. Our comprehensive experimental results show that our proposed approach is able to scale up to medium size networks with performance comparable to the centralized environment.