Research On Image Automatic Annotation Method Based On Semantic Scene Classification And Multi-view Learning

Nowadays,the number of digital images is exploding,and the management of massive data requires an effective browsing,sorting and search system.Automatic image labeling technology is used to assign labels to images for more accurate retrieval,classification results.At the same time,there are usually a variety of image representations,which can provide better image representation ability and improve the effect of image annotation and classification algorithm.This paper is to study the automatic image annotation algorithm and multi-view feature learning.The existing image automatic annotation algorithms can be divided into two categories: database search based methods and model learning based methods.The former methods,in which the candidate tag sequences are directly provided by the labeled images in the database,are simple and effective and have been widely used in recent years.However,on the one hand this method ignores the co-occurrence relationship between labels,resulting in a lower accuracy rate.On the other hand,the KNN algorithm is inefficient in large data sets.In the latter methods,the automatic image labeling problem can be regarded as a multi-class classification problem or a classification problem for each label.Because most of these methods do not take into account the potential information between the tags,huge number of labels which means a huge classification of output space,makes such methods no longer suitable for image annotation.In this paper,a semantic scene partitioning method based on nonnegative matrix decomposition is proposed to realize the mapping between labels and semantic scenes.And then we find the semantically relevant scene using the scene classification,and finally complete the image annotation in the semantic context of the sample using the KNN-based algorithm.Experiments show that the proposed algorithm not only improves the efficiency of the algorithm,but also improves the labeling performance.Because different features have different characterization capabilities for different semantic concepts,image classification and annotation algorithms are usually based on a variety of different low-level features.The introduction of multi-view features improves the performance of the algorithm and also increases the feature dimension of the algorithm,which affects the efficiency of the algorithm and reduces the usability of the algorithm.The existing multi-feature fusion,the reduced dimension algorithm usually belongs to the unsupervised learning mode,and does not use the existing label information in the data set,so the new feature can not contain the semantic relation between the samples effectively.In this paper,we propose a semi-supervised image annotation algorithm based on multi-view feature and graph embedding in view of the above-mentioned problems and the characteristics of samples with multiple labels in image annotation.Firstly,we merge multi-view features into a low-dimension feature vector by a new algorithm model based on the multi-view NMF and graph embedding.Then we leverage KNN-based algorithm based on the new feature to realize image annotation.Experiments show that the algorithm can improve the efficiency of the algorithm while ensuring the effect of annotation.