Research On Image Caption Method Based On Deep Learning And Its Application In Content Security

The Internet generates a large amount of image data all the time.How to understand and describe these image data has always been an important problem for researchers.The rapid development of the Internet makes the spread of illegal images more widely,and the problem of image content security in the Internet is becoming more and more serious,and the problem of information security needs to be solved urgently.Image caption is a multi-modal task of converting images into text,which brings new opportunities and means for content security monitoring.At present,the mainstream image caption task usually adopts the encoder-decoder structure based on deep learning.Although some progress has been made in image caption task,some existing methods still have some shortcomings.In this paper,a new image caption model is proposed.The main innovations are as follows:(1)A text generation model combining context is proposed in this paper.Aiming at the problem of information loss in the traditional model,the long and short time memory algorithm can be integrated with the context information,so that the previous text information will not be lost in the decoding process.(2)It is proposed in this paper combined with space-time text generation model of attention mechanism,aiming at different times can find the problem of weak concentration area,can take full advantage of the characteristics of visual features and double-layer text generator to generate a more accurate caption of statements,compared with the traditional attention mechanism,makes a caption of the model performance significantly increased;(3)This paper proposes a long text generation model based on the GPT-2 model.Aiming at the problem that the caption statement of the picture is too short,the GPT-2 model is introduced to extend the caption of the long text of the image content,which can produce a richer understanding of the image content.(4)In this paper,the proposed model is applied to the aspect of content security,and the combination of image caption method and content security monitoring can not only enhance the perception,cognition,prediction and early warning of illegal content in the network environment,but also reduce the cost of content security protection.To sum up,this paper studies the image caption algorithm based on deep learning and improves the existing problems of the traditional algorithm.By extending the short text,the long text caption of the image is realized.The image caption method is applied to the content security monitoring,and the illegal and criminal pictures can be effectively identified.The results of ablation experiment and comparison experiment show that the proposed image caption algorithm can describe images more accurately.