Image Understanding Methods Based On Corr-LDA Model And Convolutional Neural Networks

Image annotation and Image classification are main research contents of image understanding,and have been important research points in the fields of machine learning and pattern recognition.In the task of image annotation,image annotation methods based on probabilistic topic model have been pai d much attention by researchers in recent years.These methods combine the advantages of topic model and probabilistic graphical model,and improve the ?semantic gap? problem effectively.In the task of image classification,convolutional neural networks h ave been widely used in the field of image classification due to their remarkable classification performance.In this thesis,we do the tasks of image annotation and image classification based on probabilistic topic model and convolutional neural networks,respectively.The results on two real image datasets demonstrate the validity and rationality of our works.The main innovations in the thesis include two aspects:1.An image annotation method based on Corr-LDA model is proposed.The method mainly considers that category is a kind of valuable information for image annotation,and the category of an image and objects in it are closely related.In general,different categories contain different objects.If the category is known,the scope of image annotation words should be narrowed.Based on this,this thesis builds an image annotation model for each category,and gives new model training and testing process.Experimental results on the Label Me dataset and the UIUC-Sport dataset show that the Corr-LDA-C method is effective.In addition,even though the method is based on Corr-LDA model,it is also applicable to other probabilistic topic models used for image annotation.2.An image classification method of improving the cross entropy loss function is proposed.The method mainly considers that in neural network based on softmax cross entropy loss,the output probability is mainly based on linear computation of parameter vectors of each class in the last layer and hidden features in the layer of sample points.Therefore,the final output of neural network is effected by of the L2-norm of parameter vector of each class.Taking binary-class as an example,if the parameter vector of a class has a large L2-norm,decision boundary is close to another class with smaller L2-norm,so that sample points will be easily assigned to the class with large L2-norm.Based on it,this paper proposes a new softmax cross entropy loss(called SCE-UD Loss),which adjusts the position of decision boundary so that it is not biased to any class.Experimental results on the Label Me dataset and the UIUC-Sports dataset show that the proposed loss is superior to softmax cross entropy loss.