SKETCH-BASED IMAGE RETRIEVAL VIA A SEMI-HETEROGENEOUS CROSS-DOMAIN NETWORK

We propose a novel semi-heterogeneous network for sketch-based image retrieval (SBIR). By exploring different feature-extracting network structures and data augmentation algorithms, we design a high-performance deep-network based method for SBIR. Our work consists of three stages: 1) we propose a novel deep SBIR optimization model, termed a semi-heterogeneous network, to capture the cross-view similarities between different categories; 2) we develop a novel feature extraction method to find a cross-domain representation which contains the line information of sketches while retaining the original information of images; 3) we explore a more comprehensive structure though different parameter and network settings for further performance improvement. Based on our experiments on a widely used dataset, our approach significantly outperforms state-of-the-art methods.