Spatial-temporal Salient Object Detection Based On Dynamic Attention Center

With the rapid development of Internet technology and the popularity of 4K and other ultra-high resolution videos and images,the massive visual information has brought serious challenges to the real-time analysis of videos and images.How to optimize the allocation of computing resources efficiently and allocate the limited computing resources to areas of interest attract the attention of researchers in computer vision and other related fields.The salient object detection technology can greatly reduce the amount of visual information to be processed by imitating the human visual system's ability to select information for its own needs.Although there have been many salient object detection models in video have great detection results,but there are still some issues in restraining the disturbance of background and the temporal consistency between the salient objects in continuous multi-frame.Inspired by the Center-Prior hypothesis in image processing,we proposed a new Dynamic Attention Center-Prior hypothesis for video sequences.That is,the region locates far away from the dynamic attention center,tends to be the background area of the image,the region locates near the dynamic attention center,tends to be the salient object of the image.By calculating the distance between the image region and the dynamic attention center,the background disturbance is restrained.The dynamic attention center of each frame is calculated by the salient object of the previous frame,which can enhance the consistency of the salient object in the time domain.Due to the limitation of the motion division between the moving object and the background,the detection results of the salient object in time domain are not accurate sometimes.By analyzing the overall intensity of motion changes in the image,we presented a dynamic fusion method to combine the spatial detection result and the temporal detection result,leading to higher spatiotemporal consistencyThe experiments on Freiburg-Berkeley Motion Segmentation Dataset show that our method outperforms several state-of-art methods on subjective visual perception and objective measurements.