Large-scale-Query-text Clustering Via Weakly-supervised Deep Learning

More and more researchers utilize click-data to represent images,in order to solve the problem of semantic gap in computer vision tasks.Click-data,which is collected from search engines,is a kind of user feedback information.It is used to reflect the correlation between images and query-texts.Images can be represented as query-clickvectors based click-data.Query-click-vectors are of rich semantic but at the weakness of sparse and large scale.To deal with the redundance and sparseness in click features,we address to merge large-scale query-texts before using merged queries to represent images.A compact query-frequency-vector for images is constructed to achieve accurate fine-grained image recognition,in order to overcome semantic gap better.The paper focuses on the approach of query-text clustering via click-data,which contains query-text representation and clustering methods as below:(1)This paper proposes to represent a query-text pair as a smooth image-click graph.Firstly,we represent query-text pairs as regular image-click matrixes.And then to solve the sparseness in click feature,we propose a novel 2-D click propagation to transform image-click matrixes into image-click graphs.Compared with original image-click vectors,image-click graphs are of powerful text representations and smooth.(2)This paper proposes a query-text clustering method based on weaklysupervised deep learning schemes.Due to the difference of image-click graph and traditional visual feature,we build a deep query-text clustering model with shallow training networks for the image-click graphs.Under the minimum mean square error within categories,a hierarchical deep-click-feature is learned.In addition,we introduce a weight vector to deal with heavy noise in query-texts,which can measure the reliability of query-text.And then we update the weight vector based on weaklysupervised learning method.Queries of higher quality can be selected automatically during optimizing the deep model and weight vector iteratively,we learn more compact and more hierarchical query-text deep click feature finally.In this paper,we evaluate our method on the public Clickture-Dog and ClicktureBird data sets.The experimental results show that:(1)representing each query-text as an image-click graph helps to deal with the non-smoothness and sparseness in the original click vectors,and it is essential to design a deep model for user-click data;(2)weakly-supervised learning method contributes to further deal with heavy noise in clicked query-text set;(3)the deep query click feature learned helps to improve the image recognition performance,which accelerate the solution of semantic gap.