针对复杂场景服饰图像搜索及推荐关键技术研究

With the exponential growth of e-commerce and social image sharing websites, on-line clothing shopping has great appeal and commercial value. However, the factors, such as the difficulty of accurate description in text words,the existence of complex backgrounds and fashion models, the photometric and geometric transformations, significantly affect the accurancy of clothing image search, which make it a challenging research issue. This project focuses on key research on clothing image search an recommendation with complex scenes. Bases on the characteristics of clothing images, we analyze and model the properties of clothing and human body, and integrate them into later stages for boosting the performance. We propose multiple clothing extraction strategies for clothing images with complex scenes, including semi-supervised, automatic, and multie-dimensional co-segmentation based approaches to extract clothing objects and remove background noises. In addition, clothing image similarity measure and optimization are conducted to improve the accuracy and efficiency of clothing image search. Moreover, with the asisstance of automatic attribute based clothing description, we propose a clothing recommendation based on probability latent relationship model by taking into account of user preference and context. Combining the economic theories and clothing similarity analysis, a marginal net utility based clothing recommendation is further proposed to enhance the search experience.

电子商务和社会化图像分享网站的高速增长，使得在线服饰购物极具吸引力和商业价值。然而，服饰图像难于用文字描述，复杂背景和时尚模特的出现、光照变化及角度等诸多因素，极大地影响了服饰图像搜索的准确性，使之成为一个极具挑战性的研究课题。本课题旨在研究基于复杂场景的服饰图像搜索和推荐的关键技术。针对服饰购物图像的特点，从不同角度对服饰属性和人体模型进行分析与建模，并将其集成到传统算法以改进性能。针对复杂场景的购物图像，进行半监督、全自动以及多维度协同分割等多种策略的服饰目标提取研究，获取图像中的服饰目标并去除噪音。在此基础上，探索图像相似度测量和优化研究，以提高服饰图像检索的准确度和性能。另外，结合服饰属性的自动描述，并且考虑用户的兴趣偏好和当前情境上下文，探索基于概率潜在关系模型的服饰推荐。同时，结合经济学原理和服饰相似度，研究基于边际净效用的服饰推荐，改善用户的服饰搜索体验。

电子商务和社会化图像分享网站的高速增长，使得在线服饰购物极具吸引力和商业价值。然而，服饰图像难于用文字描述，复杂背景和时尚模特的出现、光照变化及角度等诸多因素，极大地影响了服饰图像搜索的准确性，使之成为一个极具挑战性的研究课题。针对具有复杂场景的服饰图像，围绕服饰图像的分析理解及建模、服饰图像的分割及目标自动提取、服饰图像的精准测量及属性描述、服饰和商品推荐四个方面展开深入探索。探索多种策略的多服饰目标自动提取，服饰分割以及服饰属性标注、商品匹配和推荐的算法。旨在提升复杂场景下服饰图像搜索和商品推荐的准确性，弥补了现有购物搜索引擎的不足。研究工作已成功应用于阿里巴巴服饰搜索和在线视频广告推荐系统。..项目在国际顶级期刊IEEE Trans. on Image Processing, IEEE Trans. on Multimedia, IEEE Transactions on Medical Imaging, IEEE Trans. on Image Processing, IEEE Trans. on Human-Machine Systems, Pattern Recognition等顶级期刊, 在国际顶级会议CVPR和ACM MM发表或录用17篇SCI论文，7篇EI高水平论文，获得授权发明专利7项。同时, 作为第二完成人于2016年获得教育部自然科学奖二等奖，作为第一完成人于2017年获得了河南省科技进步奖二等奖。