基于多尺度深度学习的商品图像检索

发布时间：2018-04-30 21:19

本文选题：商品图像检索 + 深度学习　；参考：《计算机研究与发展》2017年08期

【摘要】：商品图像检索的目标是检索与图像内容相符的商品,它是移动视觉搜索在电子商务中的重要应用.商品图像检索的发展,既为用户购物提供便利,又促进了电子商务向移动端发展.图像特征是影响商品图片检索性能的重要因素.复杂的图片背景、同类商品之间的相似性和被拍摄商品尺度的变化,都使得商品图像检索对图像特征提出了更高的要求.提出了一种多尺度深度神经网络,以便于抽取对复杂图片背景和目标物体尺度变化更加鲁棒的图像特征.同时根据商品类别标注信息学习图片之间的相似度.针对在线服务对响应速度的要求,通过压缩模型的深度和宽度控制了计算开销.在一个百万级的商品图片数据集上的对比实验证明:该方法在保持速度的同时提升了查询的准确率.
[Abstract]:The goal of commodity image retrieval is to retrieve goods which are consistent with the content of the image. It is an important application of mobile visual search in electronic commerce. The development of the commodity image retrieval not only provides convenience for the user shopping, but also promotes the development of e-commerce to the mobile terminal. The image feature is an important factor affecting the performance of commodity image retrieval. The image background, the similarity between the same commodities and the changes in the scale of the goods taken, all make the image retrieval of the commodity higher requirements for the image features. A multi-scale deep neural network is proposed to facilitate the extraction of more robust image features of the complex picture background and the target object scale. Do not mark the similarity between the information learning pictures. In view of the demand for the response speed of the online service, the computing overhead is controlled by the depth and width of the compression model. A comparison experiment on a million level commodity picture data set shows that the method raises the accuracy of the query while maintaining the speed.

【作者单位】：复旦大学计算机科学技术学院;上海市智能信息处理重点实验室;
【基金】：国家自然科学基金项目(61673118) 上海市浦江人才计划项目(16PJD009)~~
【分类号】：TP391.41

【参考文献】