Research On Visual Sentiment Analysis Based On Convolutional Neural Network

Since the twenty-first century,the research on artificial intelligence has reached a new climax thanks to the breakthroughs in computer technology and deep neural network technology.In this context,more and more researchers are focusing on emotion as a derivative of intelligence,hoping to restore biological emotional intelligence through the study of affective computing and to endow machines with rich emotions.Visual sentiment analysis is an important part of the field of affective computing,aiming to use computers and specific algorithms to predict the psychological changes that occur when a person sees an image,which can predict and control the emotional changes of readers when browsing and reading to a certain extent,and has important application prospects in the fields of social media analysis,multimedia analysis,and public opinion prediction.In recent years,visual sentiment analysis has attracted some researchers to conduct research and propose various sentiment prediction models with good results,but there are still some problems to be solved,such as the difficulty to extract and synthesize the deep and shallow sentiment features of sentiment images,the difficulty to reflect the local attentional characteristics of sentiment images and the difficulty to process sentiment features.In order to solve these problems,this paper carries out the research of visual emotion analysis based on visual emotion features,which includes:(1)To address the problems that existing visual sentiment analysis methods are difficult to extract and synthesize the deep and shallow emotional features of emotional images,and over-reliance on a single network and hierarchy,this paper designs an visual sentiment feature extraction model based on horizontal and vertical feature connection,and designs and uses horizontal connection modules and vertical connection modules to effectively extract and utilize the deep and shallow features of sentiment images to improve the prediction accuracy.For the input image,firstly,the feature extraction is completed by the pre-trained convolutional neural network;after that,the features of the convolutional neural network at multiple depth levels are obtained by the horizontal feature connection module,and the connection is completed in the width direction to expand the model width;secondly,the vertical connection module is used to connect multiple horizontal feature connection modules vertically in the depth direction to eliminate the dependence of the sentiment feature vector on a single depth,eliminating the dependence of the sentiment vector on a single depth layer.Finally,the sentiment mapping is completed by the classification network to obtain the sentiment labels.(2)Aiming at the relationship between the two characteristics of subjective localization and objective globalization of emotional processes,this paper proposes a visual emotional analysis method based on emotional features.The method includes objective global branch and subjective local branch to complete the imitation of visual emotion process.The objective global branch mainly extracts the low-level image such as color,texture and shape,and the high-level image features automatically learned and extracted by deep convolutional neural network in two parallel parts;the subjective local branch mainly performs image segmentation through subject target recognition,and subsequently determines the most significant subject target in the image through the processes of candidate frame fusion and saliency calculation,and completes feature extraction by deep convolutional neural network to complete feature extraction.The deep neural network used for advanced feature extraction uses either VGG-16 or the visual sentiment feature extraction model based on horizontal and vertical feature connection,and then feature connection networks are designed for each of the two extraction surface models to fuse the features extracted from each of the two branches and complete the emotion mapping through the classification network.Finally,the sentiment mapping is completed by the classification network.Finally,the effectiveness of the method is verified by ablation experiments and comparison experiments.Finally,this paper verifies the validity of the two visual sentiment analysis methods proposed in this paper in terms of design ideas through ablation experiments on two publicly available datasets;and proves the superior performance of the two methods in visual sentiment analysis problems through comparative experiments.