中文摘要：

随着触屏设备的广泛普及, 基于手绘的图像检索技术受到了越来越多的关注. 针对HOG, SIFT, RST-HELO等传统图像描述子在基于手绘图像检索领域的局限性, 提出一种基于卷积神经网络构建隐层图词包特征描述子的手绘图像检索方法. 首先对数据库图像提取边缘概率图; 其次在边缘概率图上计算Chamfer 距离变换图像, 通过对距离变换图像提取隐层图词包特征构建预分类检索. 在Flickr15K 数据集上对文中方法进行了实验, 证明基于隐层图词包特征描述子的检索效果比RST-HELO 等传统方法有了显著提高; 从图像预处理和特征描述子两方面对SBIR 进行了改进, 实验结果表明, 文中方法具有更高的准确率.

英文摘要：

With the popularity of touch-screen devices, sketch based image retrieval has attracted more and moreattention. Considering the limitation of the traditional descriptors such as HOG, SIFT and RST-HELO, we proposeda novel feature descriptors based on the bag of mid maps of the convolutional neural network. Our work isrealized by the following steps： 1） Extracting the boundary probability images of the database; 2） Converting theboundary probability images into Chamfer distance images; 3） Generating the bag of mid maps descriptor for finalretrieval. We evaluated our proposed descriptor and retrieval strategy on the Flickr15K data sets. The maincontribution of our work is the preprocessing based on the boundary probability detector and the Chamfer distancetransform and proposing a novel bag of mid maps descriptor. Results show that the proposal achieves significantimprovements.