面向移动购物平台跨场景的服装图片检索理论与方法研究

The problem of cross-scenario clothing retrieval (CSCR) is that, given a daily human photo captured, finding similar clothing in online shops. However, this is a challenging problem due to the large discrepancy between online shopping images, usually taken in ideal lighting/pose/background conditions, and user photos captured in uncontrolled conditions. The clothing retrieval system still exist some problems. The training procedure of existing usually slower speed; the annotation of training data have noisy labels; there are large discrepancies between daily photo scenario and online shopping scenario. This project aims to address this problem by investigate the theory and methods of CSCR. The main contents are as follows: (1) based on methods of sample selection and approximate calculation, we design a fast DCNN training algorithms; (2) train a robust DCNN model with noisy labels; (3) cross domain image feature mapping. This project will be verify in online shopping platform. It is a very powerful and important technology in the fields of electronic commerce. We can also extend the proposed theory and methods to similarity domains, such as CBIR (content-based image retrieval ).

本项目研究面向移动购物平台的跨场景服装图像检索技术，即通过图像特征提取和比对，从服装数据库中检索出感兴趣的服装，其难点在于待检索图像（手机拍照）和服装数据库中图像（通常是高清照片）在拍摄角度、光照、背景、分辨率等方面存在较大差异（称为跨场景），导致现有系统准确率较低。基于深度卷积神经网络（DCNN）和海量训练数据，针对现有框架下，跨场景服装检索系统中存在的模型训练时间长、训练数据中存在大量噪声等问题，本项目主要研究：（1）基于样本选择和近似运算的快速模型训练；（2）利用含有噪声的服装数据训练鲁棒模型的方法；（3）跨场景服装图像的特征映射。本项目所研究的理论和方法将在真实互联网购物平台数据上进行验证。本项目所研究的技术在电子商务领域具有广阔的应用前景，可以推广到类似的基于内容的图像检索系统。

随着拍摄设备的普及，图像数据呈现爆炸性的增长。跨场景的图像检索技术具有非常大的研究和实用价值，在安防、娱乐以及社会民生等各领域都有着非常广泛的应用。本项目研究了移动购物平台中跨场景服装检索的理论和关键技术，通过系统深入的研究，本项目的主要研究贡献在于：（1）利用互联网海量且存在噪声的数据，研究了基于弱标注的特征学习方法；（2）基于深度卷积神经网络（DCNN）跨场景的模型自适应方法研究，提出了一种半监督的跨域图像识别方法，将特征学习和自适应统一到卷积神经网络框架中；（3）提出了一种基于KL相对熵的移位操作DCNN量化方法，能将32位浮点数的DCNN模型几乎性能无损的转换成8位整数运算，并在低功耗FPGA硬件平台上实现。