Image Retargeting Quality Evaluation Method Based On Visual Saliency

With the rapid development of mobile Internet and intelligent mobile terminals,there have been display equipment with a variety of resolution.There is a need for an image retargeting method to adjust the image to a variety of resolution and aspect ratio.As the result of traditional scaling methods,such as direct scaling and cropping,distortion or loss of important information often occurs,a lot of image retargeting methods based on image content have been proposed.Nowadays,image retargeting methods have their own applicable image types.Therefore,an accurate method of image retargeting quality evaluation is urgently needed in order to automatically select the retargeting result with best quality.In this paper,the result quality evaluation of image retargeting method is studied,and a new method of image saliency detection and image retargeting quality evaluation based on visual saliency is proposed.Our method has the following innovations:Based on the neural network in image classification field,our paper designs an image saliency detection method fusing features from the deep neural network and classical image contrast features.During implementing of our method,the mid-output from deep neural network will combine classical image contrast features,which are considered as one kind of complementary information.Moreover,the salient edge will be further tuned by using a super-pixel optimization.Final experimental results in SOD,HKU-IS and iCoSeg dataset indicate that our method has achieved greatly improvement in the accuracy,recall,F-measure and MAE.Our paper presents a novel method for image retargeting quality evaluation.Based on the result of the image saliency detection,the overall deformation,local deformation and the detail loss of the image obtained by the image retargeting method are measured,and the quality evaluation of the retargeting image is obtained.The experimental results in RetargetMe dataset and CUHK dataset show that our method outperforms other existing methods in the mean and the standard deviation of Kendall coefficient.