Sketch Based Image Retrieval with Adversarial Network

Abstract

Sketch retrieval is a specific cross-domain retrieval task. The core of sketch retrieval is to learn a common feature subspace, where the features of sketches and natural images can be both discriminative and domain-invariant. However, similarity constraints can impair the performance of the feature extractor, resulting in unsatisfactory retrieval accuracy. For this problem, we propose a novel sketch based image retrieval method based on adversarial network. Our method is demonstrated as follows: Firstly, we train the sketch image network and natural image network to improve the ability of classification; secondly, we train the adversarial network to promote the feature fusion of sketches, of which the network is constituted by feature extractor and domain classifier; thirdly, we use the deep convolutional neural network to extract the deep feature to achieve retrieval. Experiments on retrieval show positive results.