Research On Multi-label Image Semantic Annotation Method Based On Deep Learning

With the development of Internet technology and urban informatization,image recognition and labeling has become an important research topic.The explosive growth of the number of images brings great challenges to the organization management and retrieval of images.Automatic image annotation is an effective way to solve the above problems.There are two problems in image semantic automatic annotation:1.The existing data sets are mostly unbalanced data sets,and the images with a small number of data sets are not fully trained in the model training stage,resulting in the model for the images.The included object recognition ability is poor,and the corresponding annotation word has a low accuracy rate;2.The traditional convolutional neural network model can not extract multi-scale image features,and since the pixel matrix of the image is non-negative,the negative features of the image cannot be obtained in the feature extraction process.For the above two issues,this paper has done the following work:1.Aiming at the problem that the training samples in the image semantic annotation are not balanced,resulting in low accuracy of low-frequency labeling,a two-channel convolutional neural network model is proposed.The model has two channels,one of which is used to train low frequency training samples in the data set to increase the proportion of low frequency samples in the entire model,and the other is used to train all training sets.In the labeling process,the outputs of the two channels are fused to make a decision on the labeling words of the required labels.Experiments were carried out in the Pascal voc2012 standard dataset.Experiments show that the DCCNN model has significantly improved the accuracy and efficiency of labeling of low-frequency annotated words compared with convolutional neural networks.2.Aiming at the problem that the traditional convolutional neural network feature extraction scale is single and can not obtain the negative value of retained image and the memory resource consumption of dual-channel convolutional neural network is large,a CM-Supplement convolutional neural network model is proposed.The network model combines the advantages of three structures: hole convolution,Inception module,and Supplement network.The CM-Supplement network model uses hole convolution to replace the ordinary convolution,so that the model can increase the receptive field without increasing the parameter amount;the Inception module is integrated into the model,so that the model can extract more images based on reducing the memory consumption.Scale feature information;applying the Supplement network structure to the CM-Supplement network model,enabling the model to obtain negative feature information of the image.Experiments show that the CM-Supplement model has significantly improved labeling accuracy and efficiency compared to the traditional convolutional neural network,and the memory consumption is reduced compared with the two-channel convolutional neural network.