Relative Attribute Learning Based On Cross-image Representation

Relative attribute learning aims to analyze the relative strength of images on a certain attribute.Attributes are properties observable in images that have human-designated names.They are valu-able as a new semantic cue for image description,and are widely used in object recognition,image retrieval,image caption and zero-shot learning.Relative attribute learning has received extensive attention from academia and industry.The existing.relative attribute learning efforts are mainly based on the learning-to-rank framework,mapping the single-image representation to the score of attribute strength,ignoring the relationship between image features and attributes.Based on the above analysis,we conduct our study in the following two aspects:on the one hand,learning cross-image representation,mining the relationship between features and attributes and mapping the cross-image features to relative attribute;on the other hand,introducing channel-wise attention mechanism,mining differences in the degree of correlation between features and attributes and automatically discovering and focusing on valuable image features.Specifically,the main contributions of this paper are reflected in the following two aspects:First,this paper proposes a relative attribute learning network based on cross-image represen-tation features.The model makes full use of the image features extracted from the deep convo-lutional neural network to perform cross-image feature learning and establishes the mapping from cross-image feature space to relative attribute space,which realizes the joint optimization of image feature extraction and relative attribute prediction.Second,this paper proposes a relative attribute learning network based on attention mechanism and cross-image features.The model reveals and exploits the difference in the degree of correlation between image features and attributes,and introduces a attention mechanism in the image feature dimension.Through the attention module,the model automatically discovers and pays attention to the image features related to attribute learning,and enhances the feature learning capcity of the model and the ability to capture the relationship between features and attributes.