Research On Image Retargeting Algorithm Based On Saliency Detection

Image Retargeting is mainly used to preserve important information and visual effects of an image while resizing it.The technology is designed to address the potential for image distortion and loss of important content due to differences in aspect ratio and resolution when images are displayed on different display devices,such as smartphones,tablets,TVS,etc.At present,the research difficulties of image retargeting mainly include quality loss and practicability.Simple image retargeting methods,such as zooming and cropping,can lead to distorted and blurry images;The image processing method based on deep learning requires a large amount of training data and computing resources,and may have poor generalization ability,so its practicability is slightly low.In order to improve the image quality and practicability of image retargeting,this paper explores an image retargeting algorithm based on saliency detection.Firstly,a significance detection algorithm is proposed to generate a significance graph with accurate content and clear boundary.Secondly,the obtained saliency map was fused with the gradient map and Canny edge detection map to generate a new significance map,which guided the joint carving to remove the low energy joints in the image for image retargeting.Finally,a threshold value is calculated based on the importance graph,which determines the switching point for applying the joint carving and scaling methods.The main research work of this paper is as follows:First,the existing cascaded partial decoder network(CPD)is improved.On the one hand,considering the high spatial resolution of low level features,low contribution to performance,but high computational cost,the network constructs a partial decoder,discarding the shallower features with high resolution can achieve acceleration,and integrating the deeper features can obtain relatively accurate salient maps.On the other hand,hybrid loss guides the network to learn at pixel level,image block level and image level through the fusion of binary cross entropy(BCE),structural similarity(SSIM)and cross association(Io U)losses.The use of hybrid loss helps to reduce the pseudo errors caused by the cross propagation of information learned at the boundary and information learned at other regions of the image.The improved cascaded partial decoder network can effectively segment significant target regions and accurately detect fine structures with clear boundaries.Experiments on six benchmark data sets show that the proposed significance detection algorithm not only runs fast,but also performs well in accuracy and clear boundary acquisition.Secondly,an image retargeting algorithm based on saliency detection is proposed.First,the saliency map is fused with the gradient map and the Canny edge detection map to generate a new significance map,which guides the seam carving to remove the low energy seams in the image for image retargeting.Secondly,to prevent excessive distortion when seam carving reduces the image size by more than one point,a threshold is calculated from the importance graph which determines the switching point at which the seam carving and scaling methods are applied.Finally,when the set threshold is reached,the retargeting map carved through the seam still does not reach the target size,then the zoom method is used to adjust the image size.This paper analyzes the image retargeting method through subjective evaluation and objective evaluation.Because the method in this paper integrates the salient image in the joint carving,the importance of the main part of the image is improved,so that the low energy part of the image is mainly removed and the high energy part is retained in the image retargeting,the image main deformation is reduced,and the quality of the image retargeting is improved.The experimental results show that,compared with prev Io Us algorithms,the proposed method can produce high-quality and stable retargeted images in a short time when the image aspect ratio is changed to a high degree,and has good effect in reducing image subject deformation,image distortion and visual artifacts,and has high real-time performance,stability and practicability.