Compact Root Bilinear CNNs for Content-Based Image Retrieval
Ahmad Alzu’bi, Abbes Amira, Naeem Ramzan · 2016
Convolutional Neural Networks (CNNs) have recently demonstrated a superior performance in computer vision applications; including image retrieval. This paper introduces a bilinear CNN-based model for the first time in the context of Content-Based Image Retrieval (CBIR). The proposed architecture consists of two feature extractors using a pre-trained deep CNN model fine-tuned for image retrieval task to generate a Compact Root Bilinear CNN (CRB-CNN) architecture. Image features are directly extracted from the activations of convolutional layers then pooled at image locations. Additionally, the output size of bilinear features is largely reduced to a compact but high descriminative image representation using kernal-based low-dimensional projection and pooling, which is a fundamental improvement in the retrieval performance in terms of search speed and memory size. An end-to-end training is applied by back-probagation to learn the parameters of the final CRB-CNN. Experimental results reported on the standard Holidays image dataset show the efficiency of the architecture at extracting and learning even complex features for CBIR tasks. Specifically, using a vector of 64-dimension, it achieves 95.13% mAP accuracy and outperforms the best results of state-of-the-art approaches.