Research On Image Emotion Classification Based On Deep Learning

With the development of science and technology and the progress of network technology,the methods of contact and communication among people have developed a variety of trends.Social networks are increasingly developed into a place where people show real feelings,and to a certain extent,the images under the messages also reflects people's emotional state.How to find the relationship between images and emotions and put the research into use become a hot topic in the field of social networking,which has a certain theoretical significance and practical value.As an important research method in image classification and image recognition,deep learning has become one of the most crucial research methods at home and abroad,which has more obvious advantages than traditional methods.The emotional image is a kind of relatively complex and detailed natural structure,so the algorithms of Deep Belief Network(DBN)and Convolutional Neural Network(CNN)are applied in this paper.According to specific issues of the image emotion classification,the corresponding solutions are put forward,the research content is as follows:(1)Based on the DBN algorithm,aiming at the problems of small gradient and slow convergence in DBN network,the MI-DBN algorithm which was combined with multi-information theory was proposed to improve the DBN algorithm.The feasibility of the MI-DBN algorithm was verified by the MNIST data set,and the handwritten numerals in the social network image were tested in the trained network,which laid the foundation for the later images classification in this paper.Finally,the Caltech 101 dataset was used to test the feasibility of this algorithm.(2)CNN network,characterized by weight sharing,is often used to conduct image classification research,which can effectively reduce the training parameters.In the CNN algorithm,the network layer is deep,the gradient of the input layer is changing slowly when the parameters are modified when using the BP algorithm counterpropagation,and the effect of training is converged to the local minimum value instead of the global minimum value when the gradient decreases,increasing the convergence rate of the gradient.In this paper,the MI-CNN based on multi-interest theory was proposed,which made it possible to modify the parameters by using the new information of the previous cycle in the reverse propagation process.Finally,the CNN algorithm and MI-CNN algorithm were used to classify the image data(amusement,anger,awe,contentment,disgust,excitement,fear,sad)set to verify the feasibility of MI-CNN algorithm.In summary,the research method of this paper solves the problem of convergence speed and precision of BP algorithm in image emotion classification,and proves the validity and correctness of the proposed methods through a large number of tests.