The Research On Thumbnail Algorithm Based On Saliency Matting

In recent years,due to the emergence of a variety of mobile smart devices,the size adaptation of images on different smart devices has become a serious problem.The scaling of the image size should not only adapt to the screen size,but also consider the image content,and keep important follow-ups.Notable objects.The generation of thumbnails is an important research content in the field of image processing.Thumbnails can be seen everywhere in daily life.For example,major e-commerce platforms and portals need to deliver as much information as possible to users within a limited space.To reduce transmission time and improve browsing efficiency,the concept of thumbnails is proposed.The generation of thumbnails must not only consider the image size,but also pay attention to several important issues such as image global information,structural smoothness,and prominent expression of important objects.In some traditional methods,the pixels of important targets in the foreground are usually deleted in a small amount,while the non-important areas of the background are deleted in a large amount.This will cause irregular distortion of the reduced image,which will lose the representative content of the original image.At present,deep learning methods are becoming more mature and widely used in thumbnail work.By using a large number of data sets to train neural network models,the quality of thumbnail generation is getting higher and higher.The current mainstream deep learning algorithm relies on neural networks to detect important areas and crop the original image.Although it can effectively retain the important information of the image,it cannot be applied to images of multiple scenes and has certain limitations.This paper takes into account the characteristics of thumbnails and image scenes,combines traditional methods with deep learning methods,and proposes a new thumbnail generation idea-"separationfusion".That is to separate the important objects in the original image from the background area for different processing,and finally merge to generate thumbnails.The main research results are as follows:(1)For the important areas of the image foreground,a saliency target matting network with the attention mechanism is proposed to extract the important areas of the image.The network model mainly includes a target detection module and a saliency matting module.First,Faster R-CNN is used to detect all important targets in the image,and then the improved saliency matting network module is used for saliency target detection.Finally,the important target and the salient target are merged and matted to obtain the important area of the image.(2)Combining image texture structure extraction and depth-aware image scaling methods,input RGB color image and its corresponding depth map,and consider the depth information of the image and the smallest perceivable error.Firstly,texture removal is completed based on the novel local mutation measurement,and then the energy optimization method based on multi-scale graph cutting is performed to adjust the image scaling,which can effectively remove the pixel seams of the image and complete the stitching well.(3)For background images,an extended generative confrontation network is proposed,which repairs the background after matting and then performs content-aware scaling.Finally,through the corresponding relationship between the target and the background of the structure map,the important area is merged with the background to generate the thumbnail.The effects of the proposed algorithm are evaluated through qualitative and quantitative experiments.The experimental results show that the algorithm proposed in this paper significantly improves the effect of thumbnail generation compared with the current more advanced thumbnail algorithms.