Research On Image Scene Semantic Recognition Based On Probability Topic Model

With the rapid development of photo-taking devices and social media,hundreds of millions of pictures appear on social networks every day.In order to manage such a large amount of image data,it is of great significance for scientific research and industrial application to find a way to correctly analyze and understand these scene content.However,the complexity of image scene content,such as intra-class inconsistency,inter-class consistency,illumination variation,and the need to know the relevant domain expertise to extract the key features,all of which greatly hinder the recognition of the image scene.How to extract image scene feature and mine the potential topic of image automatically and availably has become a rapidly developing research topic.The deep neural network model and the Latent Dirichlet Allocation model(LDA)show their respective advantages when dealing with problems in the field of vision.In order to improve the limitation of traditional image scene semantic recognition,this paper attempts to combine the advantages of these two models and proposes a method of image scene semantic recognition based on probabilistic topic model.The original feature space is projected into the topic space,then the image is represented by a topic distribution.The main contents of this paper are as follows:First,aiming at the problem that the traditional method of scene recognition has insufficient ability to express features and requires high professional knowledge,this paper makes use of the excellent automatic learning ability of deep neural network.The weight trained on the ImageNet dataset is transferred as the initial weight of the network,then fine-tune the convolutional neural network to extract the visual features of the image scene automatically and hierarchically.Second,in order to obtain the input that conforms to the topic model,we first need to encode the extracted features,and use the feature coding scheme to aggregate the extracted features into the form of global vector representation.In order to overcome the shortcomings of traditional feature coding schemes,in this paper,an improved hard allocation coding scheme is proposed to cluster the extracted features to produce a suitable visual dictionary.Third,aiming at the situation that the traditional methods fail to make full use of the latent semantics of the image,this paper combines the idea of the Latent Dirichlet Allocation model,by learning image modeling and probability derivation,abstracting the potential relevance of visual words to obtain a higher level of topic semantic,then the topic distribution of image scene that is more close to the human visual cognition is obtained to improve the accuracy of image scene semantic recognition.Through experimental analysis and verification,the method of image scene semantic recognition based on probabilistic topic model proposed in this paper can effectively combine the powerful feature learning ability of convolution neural network and the topic derivation ability of the Latent Dirichlet Allocation model,which provides a new idea for semantic recognition of image scene.