跨场景时尚图像的在线提取

Cross-scenario Online Retrieval of Fashion Images

随着智能手机的普及,人们可以随时随地购物,但基于关键字的搜索很难准确检索特定服装款式。当看到想要的物品时,基于内容的在线检索方法可以在不知道确切的文本描述的情况下带来极大的便利性。然而,由于购物网站的图片是在专业的灯光、场景布局下拍摄的,而实时图像在背景、灯光等方面都有所不同,这使得最相似的物品很难匹配。如何在不同的拍摄角度下减少背景噪声干扰,获得准确率的检索结果是一个挑战。目前大规模的时尚图像数据集很容易获取图像特征,机器学习方法可以对数据进行预处理,消除背景干扰,提高不同角度下的检索精度。由于跨场景的不确定性,本文首先使用目标检测算法目标定位,找出需要检索的目标商品,再进行图片分割;其次,使用卷积神经网络对图片进行特征提取;最后,在图片数据库中找出与其最相似的数据图片。本文提出的不同的检索方式可以满足不同用户的不同需求,给予用户更好的体验。

With the popularity of smartphones,people can shop anytime,anywhere,but in the past,keyword-based searches were difficult to accurately retrieve for specific clothing styles. When you see the clothes you want,content-based online retrieval method can bring great convenience without knowing the exact text description. However,since the image of the shopping website is shot in a professional lighting,scene layout,etc.,and the real-time image is different in background,lighting,which makes it difficult to match the most similar item. how to reduce the background noise interference and obtain accurate retrieval results under different shooting angles is a challenge. Nowadays large-scale fashion image data set could be obtained and could be implemented in our research to retrieve abundant features,machine learning method could preprocess the data to remove background interference and improve the retrieval accuracy under different angles. This paper designs a fashion item recommendation system that can provide the best matching products in real time and accurately according to the pictures given. Our goal is to provide innovative models and methods as well as new technologies that contribute to research and future industrial applications.