Image Analysis And Annotation Based On Deep Learning

With the rapid development of media technology and Internet technology,media platforms are becoming popular and present a variety of images.These image data was shared with various media platforms and the Internet by netizens.More and more people were attracted by images with plenty of colors.Since then we entered the time of big data in images.How to search the image data that meet user‘s needs quickly and efficiently in such a large scale image,is a very important and valuable study project.Because the current popular retrieval method is text-based retrieval,automatic image annotation has become the core research content of text retrieval.to get the image quickly,it is necessary to analyze the content of the image and annotate the image comprehensively.Based on the analysis and optimization of the current popular target detection algorithm,this paper proposes a feature re-calibration algorithm based on deep learning,which reduces the model volume,reduces the calculation amount and does not lose accuracy.And integrate the Show and Tell model to automatically generate image description according to the image,and store the basic information,keywords and image description into Elasticsearch to facilitate multi-dimensional image retrieval.And the main work of this paper is as follows:(1)The basic network optimization method based on Refine Det was designed.Because the vgg-16 network has the disadvantages of large model size and large computation,an optimized version of the Refine Det method was designed to greatly reduce computation and resource footprint.In this paper,the Mnase Net model based on reinforcement learning is used to replace vgg-16 in image feature extraction.The accuracy is stable but slightly decreased.However,this method reduces the computation and resource occupation.(2)This paper proposes a method to introduce SENet into the ARM and ODM modules of Refine Det,to improve the accuracy without introducing computation.SENet considers the relationships between feature channels to explicitly model the interdependencies between feature channels.The importance of each feature channel is acquired automatically through learning,and then the useful features are promoted according to this importance,and the features that are not useful to the current task are suppressed,and feature recalibration is carried out.Through the further optimization of the network,the amount of calculation is slightly increased,but the increase of the relative accuracy is negligible.(3)The image management platform based on image annotation is designed and implemented.the Image Management Platform is implemented,which can manage a large number of images and realize the rapid retrieval of images.The Image Management Platform relies on image content for classification,including Exif information of the image,image category information automatically annotated by deep learning algorithm model,and other image information manually filled in by users.Meanwhile,to facilitate image retrieval,the above picture information and image description information automatically generated by the Show and Tell model can be stored into Elasticsearch.Based on Elasticsearch to search the picture efficiently,fastly.