Font Size: a A A

Research Of Automatic Image Cropping Algorithm Based On Reinforcement Learning

Posted on:2021-05-14Degree:MasterType:Thesis
Country:ChinaCandidate:Y Q ZhangFull Text:PDF
GTID:2518306308968339Subject:digital media technology
Abstract/Summary:PDF Full Text Request
Automatic image cropping is a common image processing task.It aims at changing the composition of images to improve its aesthetic quality.Excellent automatic image cropping methods can provide professional advice for image editors and save time.Most of the existing methods are based on specific features.They adopt sliding window mechanism to generate numerous cropping candidates,and then select the final results based on these specific features.It is very time-consuming,and has high hardware requirement.What's more,it can only produce cropping results of a limited aspect ratios.In this condition,the optimal results may not be obtained.In the face of these situations,we propose a new automatic image cropping method based on reinforcement learning.The method improves the quality and efficiency of automatic cropping.On this basis,the proposed algorithm is tested for mobile application scenarios.The main work and innovations of this paper are as follows:In this paper,image cropping is regarded as the process of serialization decision making,and an automatic image cropping method based on reinforcement learning is proposed.It not only guarantees the cropping effect but also improves the cropping efficiency.Experiments on the Flickr Cropping Dataset and the CUHK Image Cropping Dataset show that reinforcement learning can improve cropping efficiency.The cropping can be done in less than three steps.It crops step by step,thus overcomes the defect of sliding window mechanism.And in theory it can produce results with any aspect ratio.This paper also proposed an automatic image cropping method combined with visual significance prediction and aesthetic evaluation,which improves the cropping quality.Experimental results show that our method achieves excellent cropping effect on the open datasets.The method combines the basic feature with saliency feature of the images as the state representation for reinforcement learning.Inspired by human decision making,the cropping history is also embedded in the state.During the cropping process,the aesthetic evaluation of the current results and the cropping precision are integrated to guide the cropping process.After completing the automatic image cropping model,the model is tested and evaluated on the mobile phone.The model has the advantage of fast operation speed and great cropping effect,which has practical application value.
Keywords/Search Tags:automatic image cropping, reinforcement learning, visual saliency detection
PDF Full Text Request
Related items