Research On Image Aesthetics Optimization Method Based On Composition Rules

More than 90% of the information that humans receive every day comes from vision.With the continuous improvement of people's quality of life,the aesthetic demand for visual media is gradually increasing."Composition" is an important factor affecting the beauty of an image.It and other elements of photography determine the vividness and narrative ability of the image.Although the current research on image aesthetics implicitly considers simple composition rules,due to the subjectivity and ambiguity of composition,there are no reliable automatic composition classification methods and image optimization methods that explicitly consider composition rules.The two research contents of this article aim to solve the above problems.First of all,for the unreliable problem of the existing composition classification model,we divide the composition into 9 categories and propose a composition classification model based on space invariance convolutional neural network,called RSTN,which uses Resnet-blurpooling as The backbone network,in which the RSTN structure we designed is added,makes the model have translation invariance and rotation invariance,and increases the generalization effect of the model for general snapshots or skewed images.In order to solve the problem of a convolutional neural network limiting the size and aspect ratio of the input image,we use an adaptive pooling layer in the model.In the end,our composition classification prediction model improved the accuracy of Baseline by 6% and the rotation consistency by 16%.Furthermore,based on the composition classification model,we divided the images into three categories according to their sensitivity to editing and designed a composition optimization method for each category of images.Secondly,we propose a real-time image view recommendation model CVPN with clear composition information.It takes an image as input and outputs a sequence of partial views from high to low aesthetic value and the composition category of each partial view.The model is trained by a novel model distillation framework(teacherstudent architecture).In order to train the model,we separately trained two networks,one is the composition prediction model mentioned above,and the other is the aesthetic evaluation model.These two models are used as teacher models to supervise and train CVPN as student models.And a set of extraction process of crop anchor frame based on golden section theory is proposed.At the same time,we also constructed a dataset VCAD for training the view recommendation model.The dataset consists of two image datasets in different fields.The labels are automatically labeled by our two teacher models and candidate frame extraction algorithms.Our real-time view recommendation model has a wide range of applications,such as image cropping,image thumbnail generation,real-time photography guidance,partial image composition analysis,composition retrieval,automatic album management,and video cover generation.