The Research On Image Relative Attribute Learning Method

With the development of computer technology, it is more convenient for people to release and get information rapidly. Although a tremendous amount of information enriches the life of people, it is a big challenge for people to realise the organization, inquiry and analysis of the information. Effective techniques for image understanding have been urgent needs for lots of applications. The visual attributes, as a mid-level bridge between low feature data and high level semantics, have been successfully applied in objection recognition, image classification, image retrieval, etc. The research on attributes can be divided into two branches that are binary attributes and relative attributes. Binary attributes have limitations on description of images and cannot provide the accurate information. However, relative attributes are more intuitive to describe images based on the relative strengths and provide much more information than binary attributes since they vary across a spectrum of visual properties. For example, which image is more natural in scene images, which image is sportier in dataset of goods, etc. It is a hot topic that how to improve the accuracy about the model of reative attribute learning because it is the foundation of the relative attribute applications. In this paper, mining of pairwise comparisons labeling and key feature preservation as the two main issues are studied and the detail is as follows.First, the robustness of relative attribute learning depends on the labeled comparative image pairs. However, manually labeling pairwise comparisons is a labor intensive and time-consuming task. In this paper, a semi-supervised learning approach based on group sparse is proposed to discover pairwise comparisons automatically. Firstly, we generate an initial level division of the labeled training images for the basic of new constraints. Then, group sparse classification for the unlabeled images is introduced by embedding the level information into the dictionary. The semi-supervised process is conducted by selecting samples which have minimum reconstruction errors and adding new constraints to the model by comparing the selected ones with the samples in dictionary. Experiments on three public datasets (OSR, Pub fig and Shoes) demonstrate the effectiveness of our proposed approach and the method of active learning for relative attribute is compared. The result shows our method has merits both on efficiency and accuracy.Second, the extracted features are needed when relative attribute learning. There is negative information about features and not each dimension of the features plays a positive role for relative attribute learning. In order to solve this proplem, the paper proposes a key feature preservation method based on rearrangement inequality. It assumes that the model with key feature is basically consistent with the model with original feature in the aspect of images ranking. According to this assumption, the mathematical model of the key feature preservation is proposed in this paper. Meanwhile, the method is proposed to solve the model. Experiments on three public datasets (OSR, LFW-10, UT-Zap50K) demonstrate the effectiveness of our proposed method. The result shows that compared with the original feature, whether single feature or cascade of multiple features, the use of the key features can improve the model accuracy.