Image Re-ranking Based On Holistic And Local DCNN Features

Image re-ranking plays an essential role in content-based image retrieval.The search engine returns a raw ranking list by matching the searching keywords with im-ages'textual information.It simply involves textual information and there is great po-tential of image' s visual information and the visual relationship among the image list.Thus image re-ranking is then involved to optimize the ranking list by analyzing the visual information.Two key aspects of image re-ranking are the strength of re-ranking model and the descriptive power of image feature.Graph-based methods have shown promising performance in image re-ranking.These methods treat each image as a vertex while the weight of the edge that connects two images indicates the adjacency between two images.Pairwise affinities are prop-agated through the whole graph iteratively and let images which share higher visual consistency with others achieve higher ranking scores.It seems that most graph-based re-ranking methods mainly focus on capturing the intrinsic manifold structure of the graph and evaluating images' visual consistency.This raises a problem that the influ-ence of query image or image with higher credibility does not get as much attention as it deserves.In this paper,we proposed an improved VisualRank algorithm with dynam-ically changing parameters and solve this problem to some extent.The descriptive power of image feature is also an important issue in image re-ranking.Deep convolutional neural network(DCNN)has provided breakthrough per-formance in computer vision,such as image classification and retrieval.Features from different layers of DCNN focus on different aspects of visual information.Low-level,such as convolutional(Conv)layer feature mainly focuses on describing texture,shape information of images and usually contains noise information.High-level,such as fully-connected layer(Fc)feature is suitable for classification task because it represents se-mantic meaning while the object category also belongs to the semantical domain.In terms of image retrieval,visual consistency is essential as well as semantical consis-tency.Thus applying Conv feature to image re-ranking can make up the weakness of fully-connected layer feature.In this paper,we include DCNN feature in our re-ranking algorithm.We extract Fc layer and Conv layer feature to construct holistic feature and local feature separately and analyze their respective strengths.Then,we extend our algorithm according to their respective properties and further improve the performance.