Research On Image Annotation Based On Graph Learning And Generative Adversarial Networks

With the continuous development of the Internet and the popularity of mobile intelligent devices,the Internet is full of a large number of unlabeled images,and how to effectively use and manage these images has become an urgent problem.It is time-consuming and labor-intensive to manually label large amounts of images.However,the automatic image annotation technology can predict the labels of the image,which has attracted widespread attention from researchers.However,in the process of image annotation,only the association relationship between label words is considered,and the semantic repetition problem of labeling results is not considered.And the distribution of label classes in the general data set is unbalanced,thus affecting the recall ability of image annotation.In response to these problems,this paper proposes two automatic image annotation algorithm,which can effectively realize the function of image annotation prediction.Firstly,aiming at the problem of imbalance label classification,this paper proposes a novel image annotation method,namely reconstitution graph learning model(RGLM).In RGLM,the algorithm transfers the labels of similar images through graph learning to complete the image annotation.The algorithm first proposes an improved nearest neighbor method,constructs a nearest neighbor image set,then achieves label balance.At the same time,the initial graph of the graph learning algorithm is reconstructed to ensure the accuracy of the algorithm.In addition,the co-occurrence imbalance between labels is used to improve the recall ability of weak labels,thereby improving annotation performance.Secondly,this paper proposes a novel image annotation algorithm based on semantic structure and conditional generative adversarial nets(SS-CGAN),which solves the problem of semantic duplication in labeled results.In the process of generating labels by the generator against the discriminator,the conditional determinant point process is introduced to constrain the generation process,which makes the diversity of label generation.Then the similarity relationship between images is used to correct the annotation results.Finally,image annotation is realized by weighted semantic path constraint to predict label selection.In this paper,the proposed two algorithms are experimented on three benchmark datasets,and the comparison algorithm selects the classic algorithms in the field of image annotation for comparison and evaluation.By analyzing the experimental results,the image labeling algorithm based on reconstructed graph learning can effectively solve the imbalance problem of label classification and improve the recall ability of weak labels.Meanwhile,the image annotation algorithm based on semantic structure and conditional generation adversarial network has shown significant performance in the experiment,proving the importance of semantic structure.Based on the comparison and analysis of two experiments,the algorithm proposed in this paper has a good performance in overall performance,and achieves accurate and efficient label prediction in the image annotation task.