The Research On Image Retrieval Based On Weakly-supervised Deep Learning

Content-based image retrieval has always been a hot topic in academia.The core question is how to make computer understand the meaning of the image as accurately as human beings.In recent years,the deep learning technology in the application of the image has made remarkable achievements.A number of content-based image retrieval(CBIR)system has emerged,such as Baidu distinguish image system and so on."Convolution Neural Network + Hash Learning" has become one of the most effective methods to solve image retrieval problem.However,the existing system(such as Baidu distinguish image system)use supervised learning to training neural network,with expensive labor cost in exchange for the excellent performance of neural networks.In addition,the traditional unsupervised learning technique can not completely reflect the high-level semantic information of the image,so it less used in the image retrieval model.With the rapid development of mobile Internet,if you can use large-scale user-generated and weakly-annotated image data to achieve image retrieval,you will save a lot of labor cost to annotate labels.However,although the weakly-annotated data contains the semantic information of the image,still have non-visual semantic labels and label ambiguity(synonyms,word polysemy)and other issues.This paper proposes a set of weakly-supervisied learning framework to use weakly-annotated information which produced by users to train neural network.The specific steps of this paper are as follows.Firstly,we use the word-of-bag(BoW)model to express the image as a vector form,and calculate the TF-IDF cohesive distance and separation distance of the label of the corresponding image.The non-visual semantic label is removed by comparing the centroid distance between the labels.Secondly,The similarity and dissimilarity pairs are formed by expressing the labels as semantic word vectors to compute the semantic similarity between images.Thirdly,Training Convolution Neural Network(CNN)according to semantic similarity relation.The output layer of CNN model is the hash code of the image.The training objectives of this paper is:Let the Hamming distance of a pair of images which have semantic similarity as close as possible,while let others away from each other.The objective function is updated and iterated by the momentum gradient descent algorithm.After mapping the image as a hash code,fast image retrieval is achieved by comparing the hash code between the images.This paper uses the MAP,Precision and other methods to do experiment on NUS-WIDE test set to verify the performance of the image retrieval method which proposed in this paper.Compared with the advanced algorithms such as BRE-CNN,LSH,ITQ and DSCH,the results show that the Weakly-Supervised Hash CNN(WSH-CNN)method we proposed in this paper is better than the above method in the aspect of accuracy.