Visual Attributes Transfer Research For Remote Sensing Image Classification

Semantic presentation of image is significant in various application domains,such as image annotation,man-machine interaction and large-scale cross-media retrieval.In recent years,as a mid-level image semantic presentation method,visual attributes meet with tremendous success in zero-shot recognition,target exquisite classification and image retrieval.Methods of Image visual attribute learning depend on huge images which annotated relevant attribute labels.However existing image attribute datasets are all composed of close-shot images.With rapid developing of optical remote sensing technology,remote sensing images present more quantity,higher resolution,more plentiful tendencies.It’s possible for semantic presentation in remote sensing images.However attribute labels for semantic learning are hardly collected.The cost of ground survey and user interaction annotation is huge.This paper researches for remote sensing image classification for visual attribute transfer.Explore feasibility of learning remote sensing image semantic presentation for using visual attribute labels of close-shot image.The paper works for theoretical analysis and experiment exploration,mainly work as follow:1)Direct at difference of low-level features in visual attribute transfer,and select which visual attribute to remote sensing images.This paper evaluated low-level features and selected visual attributes.Through quantitatively comparing results of remote sensing image classification,evaluated effectiveness of various low-level features for visual attribute learning.Based on discrimination of human semantic comprehension and visual attribute models,we selected a series of visual attributes which are applicable for remote sensing images,laid foundation for following research.2)Direct at discrepancy for predicting remote sensing images by visual attribute models based close-shot images,the paper proposed unsupervised subspace transform based cross-perspective scene attribute transfer method.Through aligning subspaces of resource and target domains,reduced discrepancy between feature distributions of close-shot image and remote sensing image.Experiments analyzed influences of different subspaces for visual attribute transfer.Result of experiment showed visual attribute transfer method in aligned subspaces improved remote sensing image classification accuracies,reduced influences for attribute model generalization ability.3)Direct at unsupervised subspace transform lacking information of visual attribute labels,the paper proposed visual attribute transfer method based Fisher distinguish subspace projection.Through quantifying feature distribution discrepancy ofresource and target domains,the method not only improved remote sensing image classification accuracies,but also made prediction more precise.Experiment compared the effect of the method on subspaces,calculated Bregman distances of different visual attributes.The results proved scientific basis and effectiveness.Through processing by our method,accuracies of remote sensing image classification and prediction have obvious improvement,proved effectiveness of the method.