Research On Multi-label Scene Image Classification Based On Deep Learning

Scene classification is a special case of image classification.It assigns corresponding semantic categories to images according to the visual content of images.Compared with general image classification,the data set of scene classification training may be small,and scene classification often has multiple labels.At the same time,there are obvious differences within classes,and there are many similarities between classes.Effective classification of multi-label scene images has many challenges.Deep learning has a very good performance in identifying objects in images.Image classification based on deep learning has long exceeded human averages on specific tasks.This paper uses the advantages of deep learning related neural network algorithms in feature extraction,and establishes classification models by combining multi-instance multi-label learning(MIML)and transfer learning related theories.Given a group of images,scenes or objects contained in the images are identified for multi-label classification.Multi-instance multi-label learning algorithm based on deep neural network,BP neural network and RBF neural network are introduced into the multi-instance multi-label learning framework,and the classification model of MIMLNN algorithm and MIMLRBF algorithm is established respectively,which can use the non-linear function to describe the more complicated label cooccurrence relation.Meanwhile,label correlation is fed back to the input to improve the system performance.Compared with the traditional MIML correlation algorithm,the classification effect is obvious.At the same time,by improving the Hausdorff distance expression in the MIMLRBF algorithm,an improved W-MIMLRBF algorithm is proposed to further improve the multi-label scene image classification effect.Multi-label scene image classification algorithm based on transfer learning consumes a small amount of computational resources and training time to achieve effective scene image classification.It is a common method to use deep learning technology in computer vision tasks and take the trained model as the starting point of the new model.Based on the trained Inception-V3 network model,this paper carries out parameter and knowledge transfer on the basis of it.The Inception-V3 model is reserved for the convolutional layer of image feature extraction,and its fully connected layer is modified to meet the multi-label classification output requirements.The classification model based on transfer learning can effectively extract the essential features of the image,and excellent results have been obtained in the multi-label classification evaluation criteria.