Research And Application Of Image Annotation Method Based On Topic Model

Due to the rapid development of big data technology,the unlabeled images in the Internet have shown great commercial value.How to quickly filter and use the unlabeled images is a problem of great research value.Image retrieval can efficiently retrieve the required images,and the effect of retrieval on the unlabeled images depends on the image annotation method.In the field of image annotation,due to the great difference between the visual content and text semantics of images,it is still very challenging work to propose an excellent image annotation model or improve the existing image annotation model.The paper focuses on the improvement of image annotation model and how to combine deep learning with traditional image annotation model.And it has done following work:(1)an image annotation method based on topic fusion and frequent patterns mining is proposed,which is an improvement on the image annotation method based on LDA.The LDA topic model regards data of image visual modality and text modality as mutually independent and generates topic distributions of corresponding modalities.Therefore,topics of different modalities are mutually independent.In order to enhance the relationship between image visual data and text data,the weighted topic fusion of different modalities was carried out during the model training and image annotation.Both the image annotation based on LDA topic model and its improvement model don't consider the influence of the association between image text information on image annotation.So the association of image text information can be used to improve the annotation results of the topic model.The improved LDA image annotation method is used to obtain the initial label set,and then the frequent patterns mining algorithm is used to mine the initial label set and find out the potential label set of the image.The inter-word correlation of all text labels in the image set was calculated,and then the inter-word correlation of potential labels is fused with the probability of labels based on the improved LDA topic model,and the initial label set is adjusted to improve the performance of image annotation.(2)an image annotation method of fusion convolutional neural network and topic model is proposed,which is a combination of deep learning and traditional image annotation model.The LDA topic model is used to model for the text data of the image training set,and generates the potential topic distribution and text topic label distribution of the image training set.The processing of text data of image training set makes up the problem of large dimension and sparse distribution of text data in convolutional neural network classification training set.The high level visual features extracted by convolutional neural network make up for the complexity of traditional image feature extraction and the limited visual information of image transmission.In order to improve the classification recall rate of image low-frequency text topics,smoothing processing is carried out for high-frequency text topics in the classification training set of convolutional neural network.Then convolutional neural network classifiers was constructed based on high-level visual features of images and corresponding potential text topics,and multi-label classification of image text topics was conducted to obtain the distribution of image text topics.This text topic distribution and the text topic label distribution generated by the LDA topic model can obtain the probability of the image labels according to the calculation formula of the probability of image label in the topic model,and to determine the set of image labels.In the image dataset,the improved image annotation method in this article is compared with the traditional image annotation model.This improved traditional image annotation method has a certain improvement in recall rate and precision rate.The image method which is proposed in this paper and combined with deep learning is compared with the traditional image annotation models and the image annotation based on deep learning.The performance of this image annotation method is better than the traditional image annotation models.Compared with the current advanced image annotation method and image annotation based on deep learning,this image annotation method performs slightly worse in precision rate.But it has been improved to some extent in recall rate.