Deep Learning Research On Scene Classification With Background Knowledge

With the rapid development of multimedia technology and the Internet,and the continuous decrease in the cost of digital storage devices,the number of images has grown at an exponential rate every day.Facing such huge image data,traditional methods of scene classification have become ineffective.In recent years,the rapid development of artificial intelligence has continuously improved the quality of people's lives.In particular,deep learning methods built by neural networks have achieved many breakthroughs in computer vision,natural language processing,and speech recognition.Due to the complex and changeable characteristics of scene images,the scene classification technology based on deep learning is more versatile than traditional techniques,and has been proved to have excellent performance in large-scale scene image recognition,which will be the main trend of future scene classification technology development.As an important research branch in the field of computer vision and pattern recognition,scene classification technology plays an important role in many fields.The overall layout of the scene image,the position of each target object,and the positional relationship between the objects make the scene image more complicated and diverse.Based on the research of the existing scene classification technology,this paper proposes a scene classification model using both convolutional neural network and external knowledge base information.The method firstly uses convolutional neural network to obtain the global features of the scene images.The global features can reflect the structure features of the entire environment.Then the detection results are used as labels,using SPARQL technology to search DBpedia knowledge base,using the trained Doc2 Vec model to encode the results returned by the search as semantic features,combined with the global features of the scene,and combined with classifiers for scene classification.Through experimental verification on the two data sets of MIT indoor and Scene-15,it is shown that the method of combining the CNN features of the scene and introducing external knowledge base knowledge methods can make the trained model have better generalization performance and higher classification accuracy.