Research On Image Automatic Annotation Algorithm Based On Deep Learning

Automatic image annotation is a research hotspot in the field of computer vision.The rapid development of the Internet has led to a surge in the number of network pictures,and traditional image retrieval methods have been unable to provide people with convenient and fast service.In recent years,a semantic-based image Retrieval method has emerged,and automatic image annotation plays an important role in this method.This article makes a deep research on how to auto-annotate images more effectively.Firstly,there exist two shortcomings of the current automatic image annotation algorithms:(1)the model is not transferable;(2)the feature fusion ability is not strong enough.This paper proposes an automatic image annotation algorithm based on weighted multi-view non-negative matrix factorization,the key idea of our approach is to treat labels as another view in addition to visual features,and find a joint factorization of all views into basis and coefficient matrices such that the coefficients of each training image are similar across views.This forces each basis vector to capture same latent concept in each view as well.Eventually we get a generation model based on the query image,the basic matrix of the model and the visual characteristics of the query image are used to transfer the labels.Secondly,there are some problems in automatic image annotation,such as the redundant labels and the lack of sufficient information.Aiming for the problems,we propose an automatic image annotation algorithm based on Cascade Network and Semantic Hierarchy(CNSH)in this paper.Firstly,from the input images and label list of dataset,we extract image features through a cascaded VGG network.The condition determinantal point process(DPP)model is trained to compute the quality score of labels that is used to determine the list of candidate labels.Secondly,we obtain the semantic hierarchy and synonyms via a label set,Word Net to build a weighted semantic path.Finally,this paper samples the candidate label set by using the DPP algorithm to obtain the final annotation results.Finally,aiming at the top-k evaluation method used in many current image automatic labeling algorithms does not conform to the labeling habits of humans,and it can easily cause the problem of mislabeled information.Therefore we propose an image automatic annotation algorithm based on LSTM recurrent neural network that jointly uses Convolutional Neural Networks and Recurrent Neural Networks(RNN).This paper transforms image annotation task into a sequence generation problem so that the model can natively predict the proper length of tags according to image contents.