In this paper, we studied the semantic features of the free-hand sketches in the research field of SBIR(sketch based image retrieval), and proposed a new approach to dig out the semantic property in sketches and improve the performance of sketches retrieval, which is based on multi-layer semantic feature learning and deep convolutional neural network. Our methods are demonstrated as follow: firstly, we put forward a new conception of multi-layer semantic feature descriptors; secondly, we constructed a corre-sponding multiple layers of deep convolutional neural network to learn the deep features of sketches; thirdly, we combined semantic features of different layers by the feature fusion algorithm to forming the final feature representations and to realize the high retrieval accuracy. The experiment on benchmark Flickr15k dataset proves the efficiency and accuracy of our proposed method.%针对手绘图像检索领域中手绘图像的语义特征, 为了深度发掘手绘图像的语义特征, 并获得高效、准确的检索结果, 提出一种基于多层语义特征和深度卷积网络的融合网络的方法. 首先提出针对手绘图像语义特征的分层的概念, 并构建与多层语义特征相对应的多层深度卷积神经网络来学习不同层次的深度特征, 然后通过特征融合, 实现多层深度语义特征的融合, 形成最终的特征描述子, 达到高精度的检索. 在基准数据库 Flickr15k 上的实验结果表明该方法是可行、有效的.
展开▼
