Research On Image Emotion Analysis Based On CNN

The image emotion analysis makes use of the computer with the algorithm to compute the emotion semantics contained in the image.Nowadays,image emotion analysis research based on Deep Neural Networks(DNN)is one of hotspots in the field of emotion computing.This thesis delve into the research on image emotion analysis based on Convolutional Neural Networks(CNN),in the aspects of image emotion analysis model construction,hierarchical image feature extraction,significant emotion feature mining and data imbalance problem in existing image emotion data set.The main work and research results of this thesis are as follows:1.Aiming at data imbalance problem in the image emotion data set and insufficient attention paid to the hard samples by cross entropy loss function,the Focal loss function,which has the characteristics of mining hard example and alleviating the data imbalance problem of training data set,was applied to the image emotion analysis model based on CNN.In order to adapt Focal loss suitable into multi-class emotion classification problem and improve the ability of the model to mine hard samples in training data,the parameters of the Focal loss function were modified.The balance parameter ? depends on class weight value,and the focal parameter ? was modified in different stage of training.Then this modified Focal loss was applied in the training of emotion image analysis model.The experimental results show that the modified Focal loss function can improve the performance of emotion image analysis model,and the accuracy,the macro recall and the macro precision of model are promoted by 0.5%-2.3%,0.4%-3.9%,0.5%-3.3% respectively.2.Since different level of image feature have different impact on emotion in image,a CNN-RNN image emotion analysis model based on attention mechanism was proposed.A CNN was used to extract the hierarchical image features in this model,then these features of different levels were fused by a bidirectional Recurrent Neural Network(RNN)with attention mechanism,according to the impact of image features of different abstraction.Meanwhile,to alleviate negative effects of imbalance problem in training data set,a parameter-modified Focal loss function was introduced into the training phase to improve the performance of the proposed emotion analysis model.The experimental results show that the proposed emotion analysis model can effectively integrate image features of different levels.Compared with the relevant emotional analysis model,the classification accuracy of proposed model is 2.9%?17.5% higher than that of other models.Compared with the cross entropy loss function,the modified Focal loss function can improve the accuracy of proposed model by 1.2%?1.4%.