Research On Image Annotation Algonthm Based On Convolutional Neural Network

With the popularity of smart phones,home computers and other digital devices and the development of communication technology,visual data such as images can be seen everywhere on the Internet sharing platform.In order to manage and use them effectively,researchers put forward image retrieval technology.Due to technical limitations and user habits,search engines provide keyword based image retrieval.This retrieval method needs to use keywords to annotate the image in advance,but it is difficult to imagine the time cost and labor cost only relying on manual annotation,so the automatic image annotation technology has developed rapidly.Due to the complexity of the model,poor generalization performance and low accuracy of the traditional image automatic annotation algorithm,this paper proposes two image automatic annotation algorithms based on convolution neural network.The main work is as follows:(1)Aiming at the problems of low accuracy and imbalanced categories of small and medium scale object annotation,an image annotation method based on multi-scale features and cost sensitive learning is proposed.This method adjusts the network structure of vgg16 and adds feature fusion module.The feature fusion module is divided into multi-scale feature extraction module and fusion feature module.The multi-scale feature extraction module extracts multi-scale features from convolution features,and the fusion feature module adaptively fuses features in the process of network learning.Based on the multi label loss function,a cost sensitive multi label loss function is proposed.Experimental results show that the proposed algorithm can improve the labeling performance of low-frequency tags while ensuring the labeling performance of high-frequency tags.(2)In order to solve the problems of insufficient training samples and unbalanced annotation categories in image annotation dataset,an image annotation method based on double convolution neural network is designed.Firstly,an image expansion method based on generative countermeasure network is proposed,which is combined with the traditional image expansion method to solve the problem of insufficient training samples.Secondly,the convolution neural network structure is improved,and deformable convolution and filter pooling are introduced to enhance the ability of labeling objects of different scales.Finally,the dataset is divided into all datasets and low-frequency label datasets,respectively two convolutional neural network models are trained independently,and a labeling result fusion module is designed to fuse the labeling results of the two models.The model trained from low-frequency dataset is more suitable for labeling low-frequency labels,which reduces the impact of class imbalance on low-frequency labels.Experiments show that the image annotation algorithm based on double convolution neural network model can improve the accuracy of image annotation.