Resaerch And Implementation Of Image Captioning Algorithm With High-level Semantics Based On Deep Learning

With the collaborative development of Internet technology,a great plenty of information stored on the social media and network.Among them,as an important information medium,image often presents the characteristics of large amount of data,more content and wide coverage.In the face of the large volume and continuous visual information in the network,manually annotation is a heavy workload,high cost and many differences,and it is not conducive to improving the efficiency of image classification and retrieval in the era of big data.Image description generated system allows users to input images and automatically return the natural language description matching the content.This mode helps to save time and labor costs,and realize the automatic understanding of image visual information.It can not only solve the efficiency and performance problems of image retrieval,but also maintain the security of network environment.As deep learning advances,image caption has made great progress in last few years.However,there are still a series of "semantic gap" problems in the existing methods,such as the absence of accuracy,detail and diversity,which means that there is still a huge space for improvement.Aiming at the existing problems in automatic description,the following research results are achieved:(1)For the semantic information deviation of generated description,this paper proposes a caption algorithm based on the fusion of residual attention mechanism and ordered memory module.It can extract visual features efficiently,and take advantage of ordered memory to learn the level information in sentences.Experiments show that,this fusion algorithm can enhance the semantic correctness of description details,and generate a more accurate and better structured description in actual.(2)For the lack of high-level semantic details in generated description,this paper proposes a visual relationship description algorithm based on knowledge map.It presents the graph of visual relationship in image vividly,explores the visual relationship between entities deeply to enrich the description details.Experiments show that,this algorithm can enhance the semantic integrity of description,and generate a more comprehensive and natural description.(3)Based on the above research,this paper proposes an image caption algorithm for high-level semantic based on deep learning.Combined with many advanced mechanisms in deep learning,the above two algorithms are integrated to optimize and improve the image description technology.The purpose is to bridge the "semantic gap" and realize automatic generation of high-quality image description.This paper evaluates the proposed algorithm on large datasets,confirms that it can generate a more accurate,comprehensive,hierarchical,fluent language description.