Research On Image Aesthetics Review Based On Deep Learning

The purpose of image aesthetic quality assessment is to enable the computer to simulate human thinking and aesthetic judgment on the aesthetic value of a picture,and then output the score or text description.And human vision,and language are closely connected,see pictures can always express some aesthetic ideas in the form of natural language,the language includes descriptions of image aesthetic aspects such as composition,light,color,etc.,so the language description of image aesthetics has very important significance,but only in the study of aesthetics,the richness and smoothness of the description is not perfect.In order to solve the existing problem of imperfect aesthetic description,we first proposed a model called Deep Image Aesthetic Reviewer(DIAReviewer),which is composed of CNN,Aesthetic semantic addition layer and D-Attention.This network structure can make the resulting Aesthetic description more fluent.The aesthetic semantic addition layer proposed in this paper is to integrate the image features extracted from CNN and the aesthetic description features extracted from the aesthetic description into a hybrid aesthetic feature input to the d-attention part.The d-attention section will replace the traditional RNN as the final text output of the model.To verify the proposed model,we also constructed a new aesthetic subtitle data set(ARD).Through experiments,the results show that our method has a certain performance improvement in producing smoother aesthetic descriptions.Secondly,in order to solve the problem of the loss of aesthetic details in the process of convolution,we also proposed the improved model of DIAReviewer,that is,the DIAReviewer-II model.Based on the DIAReviewer model,the model is optimized for feature extraction.That is,on the basis of vgg-19 model,the residual learning idea is introduced,and the residual layer of attention mechanism is formed by combining spatial attention mechanism and channel attention mechanism.In the model,the image features extracted by the first convolutional layer are processed by spatial attention mechanism and channel attention mechanism to carry out residual learning with the last convolutional layer,so as to reduce the loss of image features and make the features extracted by CNN more and more abundant,so as to enrich the aesthetic description of the output.Experimental results show that the language output of our improved model is improved in fluency and richness.