Research On Content-Aware Image/Video Retargeting Technology

With the continuous development of hardware equipment and network transmission,multimedia content have become an important part of our daily lives.People can watch images/videos through various portable display devices anytime,anywhere.Image/video content is usually produced with specific target aspect ratio,and different display devices often do not have a uniform aspect ratio.The image/video retargeting technology is to adapt the image/video to the target screen.However,the naive retargeting methods which are widely used lack the consideration of image/video content characteristics and have their own fundamental flaws.Therefore,content-aware image/video retargeting technology has become a new hot research direction in the field of computer graphics and computer vision.The content-aware image/video redirection technology can protect the salient areas in the original image/video as much as possible during the retargeting process,and bring a better viewing experience to the users.Although traditional content-aware image/video retargeting methods have tried many ways to avoid obvious visual artifacts in the original image/video,the effect is still not ideal.In recent years,deep learning has demonstrated its superior performance in image/video processing.Researchers have proposed neural network-based image retargeting methods that surpass traditional methods in subjective visual effects,but there is still much room for improvement in both effects and efficiency.Based on systematic research and summary of the current image/video retargeting methods at home and abroad,a new image/video retargeting method combined with full convolutional neural networks is proposed.The main work and innovations of this paper include: the existing image retargeting methods usually use the low-level features of the image to predict the importance of the pixel level,and lack the maintenance of the high-level semantics of the image and the overall structure of the image.In order to solve this problem,this paper uses foreground segmentation full convolutional network to extract high-and low-level features of the input image,generate saliency maps suitable for image retargeting,and compare them with non-uniformly grid warping methods which are combined to produce a more visual-pleasing retargeted image.Experiments on public datasets prove that our method has better subjective visual quality than existing image retargeting methods.After implementing the content-aware image retargeting algorithm,this paper extends the algorithm and proposes a video retargeting method.How to ensure the continuity of the retargeted video in the spatial and temporal domain and ensure the processing efficiency is a major difficulty of the video retargeting algorithm.The method in this paper uses perceptual time-domain constraints and progressive retargeting,which effectively improves the temporal consistency of the retargeted video,reduces distortion,and enables more efficient streaming processing of the video.