A Study Of Image Generation That Satisfies Compositional Rules

With the popularity of smartphones and professional cameras,users take a lot of photos every day.However,due to various factors such as shooting level and shooting environment,photos taken by ordinary photographers often fail to meet people's aesthetic requirements.Beautifying images involves expertise in aesthetic laws and composition layout,which is very complicated and time-consuming,making it difficult for ordinary users to complete convenient image beautification processing.Among many factors that affect the aesthetics of an image,composition has a great impact on how good an image looks.When a photographer with professional photography knowledge can know what composition to use for the current given scene to capture the best looking image,however,for people without professional photography knowledge,it is more difficult.In order to solve the difficulty of composition optimization for ordinary photographers after shooting,we view the problem of aesthetic composition optimization as an image generation problem based on the original input image,and we propose an automatic image generation algorithm framework that satisfies the composition rules,using composition detection,semantic segmentation,image composition optimization,and image generation techniques,based on our built-in different professional photography The algorithm uses generative adversarial networks to generate images based on different composition rules when users input images,so as to guide them for the next shot.It is demonstrated that our proposed algorithm framework can select the appropriate composition optimization module according to the different composition rules of the input images to generate images that match the corresponding composition rules.In order to further evaluate the optimized semantic images in terms of composition and select the one with the best composition to guide the image generation,we need to do aesthetic composition evaluation of the generated different semantic images.Since the current aesthetic evaluation is to rate the original input images and lacks the corresponding labels for the semantic images as well as the corresponding datasets,this paper converts the composition evaluation problem of semantic images into a comprehensive evaluation problem that combines the original input image and the semantic image.This paper proposes a multi-feature integrated aesthetic evaluation network.It extracts the aesthetic global features and the compositional features of the semantic image on the original image and the semantic image,respectively,and fuses these two features through a feature cascade to perform a comprehensive evaluation.Through experiments,it is demonstrated that the proposed aesthetic evaluation network not only improves the performance of the aesthetic evaluation of the original image,but also can be integrated into the framework of the image generation algorithm satisfying the composition rules in this paper to give different scores combined with the original image and different semantic images.