Image Saliency Detection Based On Prior Information

Computer vision means using computer to simulate the mechanism of human visual system,so that it can automatically understand and analyze natural scenes as human beings do.As one of the hot research topics in the research field of computer vision,image saliency detection has gradually been paid attention and researched by scholars.Saliency detection is usually used as a preprocessing step of image to apply to many computer vision tasks,such as image segmentation,image retrieval and image compression.The main purpose of the image saliency detection is to imitate the human visual attention mechanism to extract the visually salient regions,which are the most attractive regions in an image.After detecting the visual saliency of an image,a small amount of important information can be selected from a large amount of visual information for further processing and analysis,thus improving the efficiency of image processing.Therefore,the research of image saliency detection is of great significance.Due to the diversity and complexity of the image,it is difficult and challenging to accurately detect the salient regions.Because of the lack of high-level information,it is difficult to obtain satisfactory results from the bottom-up saliency detection method if only low-level feature information is used.Therefore,in this paper,we propose two different methods of image saliency detection based on the high-level prior information.Firstly,a method of image saliency detection based on background prior and foreground prior information is proposed.In order to improve computational efficiency,an image is divided into multiple superpixels.Then,using the background prior knowledge,the superpixels located at the four borders of the image are taken as the background regions,and a set of the prior background seeds is obtained.Therefore,a background-based saliency map is calculated by the difference between the other superpixels and the background seeds set.Next,a coarse saliency region can be obtained by constructing a convex hull that surrounds saliency points,and this convex hull can be considered as a prior knowledge of the foreground regions.Combining the background-based saliency map and the convex hull region,a set of the prior foreground seeds is obtained by threshold segmentation.Therefore,a foreground-based saliency map is calculated by the similarity between the other superpixels and the foreground seeds set.Finally,the two saliency maps are integrated and further optimized to generate a smoother saliency map,so that the detection result can not only highlight the salient regions,but also effectively suppress the background noise.Secondly,an image saliency detection method combining foreground prior information and manifold ranking algorithm is proposed.The traditional saliency detection methods based on manifold ranking algorithm have the problem of over-reliance on priori background seeds,so the detection result of complex images are not good.To deal with this problem,we use the priori foreground seeds to replace the prior background seeds,and two different feature measures are used to represent the differences between regions for saliency detection.First,convex hull of an image is calculated to get the rough location of the salient object,and the priori foreground seeds are selected by calculating the dissimilarity between the superpixels inside and outside the convex hull.Then,two closed-loop undirected graph models are constructed by using different regional feature descriptors.The priori foreground seeds obtained from the previous step are taken as query nodes,and saliency detection based on graph-based manifold ranking algorithm is performed to generate two salience maps.Finally,the two saliency maps are fused and optimized to generate a final saliency map.The experimental results in three public image databases show that the two methods proposed in this paper can achieve good detection results,and the detection performance are greatly improved compared with some state-of-the-art methods.Finally,in order to demonstrate the practicability of image saliency detection technology,the two detection algorithms proposed in this paper are applied to image processing tasks,such as image retargeting and background blurring.For image retargeting,the important target can be more focused by using saliency detection technology.For background blurring,only the non-salient regions are need to be blurred,and the detected salient regions are not blurred,which can play a prominent role in highlighting the image theme.