Research On Image Automatic Annotation Model Based On Improved Convolution Neural Network

With the advent of the big data era,hundreds of millions of unlabeled images are uploaded to the Internet every day.The purpose of "image annotation technology" is to effectively manage these unlabeled image information.Automatic image annotation based on in-depth learning has become the mainstream technology for automatic image annotation because it can independently learn the relationship between low-level features and high-level semantics.However,with the deepening of research,The existing problems are also gradually emerging:first,most of the image annotation methods based on deep learning ignore the research on small-scale annotation objects.Second,most of the deep learning methods do not conduct in-depth research on the relationship between labels in the image,resulting in the difference between the image annotation effect and the real content.To solve the above problems,this paper proposes an improved image automatic annotation model based on the existing research.The specific research contents mainly include:I.In order to solve the problem that small-scale objects are difficult to recognize,a multi-core pool image annotation method is proposed.Firstly,multi-core pool is used to conduct experimental comparison of different fusion layers to select the feature layer to be fused.Secondly,asymmetric convolution is used to conduct experimental comparison for different levels to select different asymmetric levels.After experimental comparison and analysis,the final structure of the model is determined,Finally,in order to verify the overall performance of the model,the iaprtc-12 data set is compared with the traditional method and the method based on deep learning in recent years.The results show that the evaluation parameters of this model are higher than those of other models.II.In order to mine the semantic relevance between image tags and improve the annotation performance of the model,an improved method based on the cyclic neural network model is proposed.Firstly,the original convolution layer and the pooled convolution layer are compared experimentally,and the convolution features are determined throgh the experimental results.Then the experimental comparison of the tag training sequence is carried out according to the relationship between tags,Select the label order of the model through the detailed data of the training process and training results The bi-directional loss function for model training is set,and then several groups of experiments are designed to verify the advantages of the model by comparing it with the traditional image annotation methods and the advanced image annotation methods in recent years on the ESP game data set.Finally,the actual annotation effects of the model for different types of images are compared on the Corel 5K data set,The experimental results show that the prop osed image annotation model is valuable from the evaluation parameters and the actual annotation effect,and can substantially solve some problems in the field of image annotation.