| As one of the important methods of image editing,image cropping technology can remove irrelevant parts in the image and generate pleasing cropped image,which has important research significance and application value in the era of increasingly large number of images.In this thesis we study the inefficiency of the multi-candidate region generation and evaluation paradigm used by current mainstream cropping algorithms.The main work includes:1.Put forward the hypothesis that "there is some connection between the subject of the image and the aesthetic composition".From this hypothesis,the aesthetic area can be directly returned from the image subject without the need to generate and evaluate multiple candidate areas,thereby improving the efficiency of the cropping algorithm.2.Based on the foregoing hypothesis,we design and implement two efficient image cropping networks:subject-based aesthetic regression network and end-to-end cropping network based on probability frame.Both networks achieve the final aesthetic region cropping by modeling the correlation between the subject and the aesthetic region;the difference is that the former uses a two-stage network to achieve image cropping,while the latter models the image cropping task as conditional probability maximization problem.3.Aiming at the problem that existing subject positioning algorithms require iterative search and are not differentiable,and cannot be applied to end-to-end networks,we propose a simple,efficient and differentiable area localization algorithm,and design an interesting object localization layer to solve the problem of locating the main subject in the end-to-end network.The evaluation results on three public datasets show the rationality of the hypothesis proposed in this thesis and the effectiveness of the two proposed cropping networks.The network proposed in this thesis is superior to existing image cropping algorithms in terms of accuracy and speed,and its cropping speed can be as fast as 20 milliseconds per image. |
PDF Full Text Request