Learning Visual Attributes For Image's Label Analysis

With more and more popular of portable devices,such as cell phone and camera,which could obtain visual data easliy,visual data in the network is explosive growth,according to statistics of official,Facebook have 6 billion pictures uploaded each Month,there will be dozens of hours video upload in Youku and other sites of video every minute,and a lot of visual data have been stored in some web albums like Flickr,search engines and portals,etc.To better deal with and use these data to promote our daily life,computer vision of the study ushered in new boom.On the research of computer vision,many classifiers could recognized test images easily by feed with amount of labeded data during training,however,for create label,many human work needed.In this paper,we studied to predict visual attribute deeply for attribute-based zero-shot learning,in which we can refer object classes even though there are no training data is available.Compared to standard classification method,zero-shot learning need labeled image rarely.The precision of attribute prediction playing a role of attribute-based zero-shot learning,but in general attributes learning,some of them learning attribute classifiers independent,some map the attribute relations into object function without any restriction,which will make mistake in attribute prediction.Moreover,it's critical to learning attribute correctly for zero-shot learning,how to learn more robust attribute predictor is very important thereby.The similarity between attribute predicted and attribute baseline is also mainly factor for zero-shot learning.Most of our research are related to those two problem in follow content.We summarized our contributions and mainly research at first.(a)We proposed a more effective method for decorrelating semantic visual attributes.(b)We leveraged hamming distance to calculate similarity among images,which more effective to explore the intrinsical relations and improve the performance of classification finally.(c)we also show that GAN could train visual attributes by adaptation.