Image Aesthetic Quality Evaluation And Cropping Based On Heuristic Image Region Relationship And Deep Neural Network

Computational aesthetics is to construct a model of human aesthetic perception through computer methods to quantify the degree of beauty of real objects.With the increasingly strong pursuit of beauty and the potential practical value of computational aesthetics in the field of daily life,computational aesthetics has gradually become a hot research topic for researchers.The research of computational aesthetics is helpful to guide students’ painting study,help designers judge the aesthetic degree of poster elements,assist photographers to evaluate and cut the current scene in the field of education and social business activities,and provide better visual enjoyment for people watching images.Therefore,it is of commercial and humanistic value to study computational aesthetics and implement image aesthetics model with excellent performance and landing application.Image aesthetic quality assessment(IAQA)is a challenging computer vision technology.Its research content is to let the computer simulate human visual senses to appreciate all kinds of images(such as scenery,portrait,animal,etc.)in reality.In order to solve some limitations of the current image aesthetic quality evaluation methods,such as homogenization evaluation,not introducing the position association between multiple objects.This paper proposes a new image aesthetic quality evaluation framework based on graph convolution neural network,named G-AANet.The G-AANet designed in this paper includes three modules:convolution feature extraction network of depth image,convolution network of graph and distribution matching of aesthetic quality score.Convolution feature extraction network of depth image is composed of multi-level and multi style convolution layers,which is used to extract local and global features of image.The graph convolution network calculates the information of each region of the image and outputs the distribution of aesthetic quality fraction of the image,taking into account the correlation of each region in the image.The aesthetic quality score distribution matching module is used to solve the problem of homogenization evaluation.The G-AANet designed in this paper is trained and tested on the large image aesthetic quality assessment data set AVA,and achieves excellent experimental results.Image automatic cropping is a common image editing method,which aims to improve the image aesthetic quality by gradually cutting from the edge of the image,removing the redundant part and recomposing.Most of the previous methods produce and calculate a large number of useless candidate boxes in the clipping process,which not only consumes a lot of computing power and time,but also easily leads to the problem that the best clipping box is not in the candidate box.In order to solve these two problems,inspired by the image crop process in human real life,the image crop process is regarded as a Markov sequence decision problem.In this paper,we propose a simple lightweight framework named LA2C based on deep reinforcement learning algorithm(named advantage actor critical(A2C)).LA2C includes agent,image crop environment and rollout storage.LA2C method simulates the process of human crop image,and according to the judgment of the next crop action of the current local image,crops images several times to achieve real-time,efficient and reasonable automatic image crop.The LA2C method is evaluated and tested on the open Flickr crop data set(FCD).The IoU index results show that compared with the previous automatic crop tools,our method achieves excellent performance with fewer clipping steps and time.The two works of this paper are in the field of computational aesthetics,which can be applied in aesthetics teaching,photography,design and other fields.They promote the development of image aesthetics quality evaluation and automatic image cropping in computational aesthetics,and have practical significance and social value.