The traditional methods on sketch based image retrieval (SBIR) is mainly based on the hand-crafted de-scriptors such as HOG and SIFT .Considering the limitations of the traditional hand-crafted descriptors ,we propose a novel approach based on the abstract-level transform and the convolutional neural network (CNN) .Our work is re-alized by the following steps :1) Extracting the boundary probability images from the database images ;2) Conver-ting the boundary probability images into abstract-level images ;3) Inputting the abstract-level images into the net-works and extracting the hidden layers' output vectors ;4) Combining different hidden layers' output vectors as the final descriptor for retrieval .We evaluate our proposed retrieval strategy on Flickr15K datasets .The main contribu-tions of our work are the preprocessing based on the boundary probability detector and the abstract-level transform ation ,furthermore ,proposing an improved combination of deep features .Results show that the proposal achieves significant improvements .%针对人工设计的描述子(HOG、SIFT等)在基于手绘的图像检索(Sketch Based Image Retrieval,SBIR)领域的局限性,提出了一种融合抽象层级变换和卷积神经网络构建联合深度特征描述子的手绘图像检索方法.首先,提取常规图像的边缘概率图,在此基础上进行不同抽象层级的图像变换,将抽象层级变换图像输入到深度神经网络并提取不同隐层的输出向量,最后,联合不同隐层的输出向量作为手绘图像检索的特征描述子(即联合深度特征描述子).在Flickr15k数据库上对本方法进行了实验验证,结果表明:融合抽象层级变换和联合深度特征描述子的检索效果相较HOG、SIFT等传统方法有显著提高.本方法从图像预处理和特征描述子构建2个方面,对SBIR问题进行了改进,具有更高的准确率.
展开▼
