Research On Image Quality Assessment And Improvement Algorithm Based On Aesthetics

The love for beauty is native to humans.Since ancient times,people have never stopped pursuing"beauty".Aesthetics is an innate ability of people.It is a meaningful research to use artificial intelligence technology to let comput-ers perceive,discover and generate"beauty".With the rapid development and widespread popularity of photo-capable devices such as cameras,camcorders and smart phones,people have become accustomed to using photos to record important moments in life.However,limited by personal time and energy,peo-ple cannot quickly select high-quality photos.Limited by personal aesthetic awareness and aesthetic ability,people cannot improve the aesthetic quality of images like professional photographers.Faced with such problems,the image aesthetics quality assessment and improvement technology with artificial intel-ligence as the core provides an economical and feasible solution for people to obtain high-quality photos.At the same time,the technology promotes the de-velopment of artificial intelligence technology to simulate human aesthetic and thinking.The main work and innovations of this paper are as follows:1.An image aesthetic quality assessment algorithm based on the saliency symbiosis network is proposed.Considering that in the process of aesthetic assessment of images,peo-ple not only pay attention to the overall aesthetic effect of the image,but also pay special attention to the main area of the image.This paper proposes an image aesthetic quality assessment algorithm based on the saliency symbio-sis network.Firstly,deep features are extracted from the original image and salient image.Secondly,a new combined pooling strategy is proposed to solve the problem that the input image of the network is of fixed size.Finally,an adaptive loss function is proposed to improve the performance of the model.Experimental results show that the classification accuracy of the proposed al-gorithm on the three datasets is increased by 2.9%,1.8%and 5.4%,respectively,and the calculation speed is increased by 9 times.2.An aesthetic attribute prediction algorithm based on multi-label is proposed.Aiming at the problem that fusion of global deep features and aesthetic at-tribute features will lead to high complexity and feature space explosion,this pa-per proposes a multi-label-based aesthetic attribute prediction algorithm.First of all,five aesthetic features are designed and extracted,which are brightness,color harmony,depth of field,motion blur and rule of thirds.Secondly,in or-der to solve the problem of high complexity in the process of feature fusion,a novel aesthetic feature fusion algorithm is proposed.Finally,in order to solve the problem of feature space explosion,the proposed algorithm mines the corre-lation between aesthetic attributes,and uses only three aesthetic features to re-alize the prediction of multiple aesthetic attributes.Experimental results show that the proposed algorithm can accurately evaluate multiple aesthetic attributes of images,especially in predicting the rule of thirds scores and image clarity scores,and achieves very good results,with classification accuracy reaching 97.8%and 99.4%,respectively.3.An image aesthetics cropping algorithm based on reinforcement learn-ing is proposed.Aiming at the problem that the automatic image cropping algorithm ig-nores the image composition,this paper proposes an image aesthetics cropping algorithm based on reinforcement learning.First of all,in order to obtain image composition information,the rule of thirds features are designed and extracted.Secondly,in order to solve the problem of the huge number of cropping candi-date windows,the image cropping task is modeled as a Markov decision process based on reinforcement learning,and the feature extraction module,agent and action space of the algorithm are designed.Finally,in order to improve the aes-thetic and composition quality of the cropped image,the aesthetic composition reward function and the aesthetic global score reward function are added in the design of the reward function.Experimental results show that the proposed al-gorithm increases the IoU(Intersection over Union)value on the public dataset by 7.18%,and reduces the number of cropping steps by 2 steps.4.An image aesthetics deblurring algorithm based on the generative adversarial network is proposed.Aiming at the problem that the image definition enhancement algorithm only evaluates the image from objective indicators,but does not evaluate and improve the image quality from the aesthetic point of view,this paper proposes an image aesthetics deblurring algorithm based on the generative adversarial network.Firstly,in order to improve the aesthetic quality of the deblurred im-age,the aesthetic assessment network is used to calculate the loss of aesthetic clarity and the overall loss of aesthetics.Secondly,in order to improve the edge detail expression ability of the algorithm,the deep feature perception loss is improved.Then,to reduce the difference between the generated image and the clear image,the pixel loss is added.Finally,an aesthetic deblurring dataset is constructed.Experimental results show that the proposed algorithm improves the aesthetic score and clarity score of the deblurred image,and exhibits a good subjective visual effect.