Research On Video Salient Object Detection Algorithm

Human visual processing capabilities can accurately and quickly distinguish salient targets from other regions in complex scenes,depending on how the eyes react differently to different objects.The saliency detection is based on the attention mechanism of the human eye,and the computer simulates the visual attention mechanism to judge whether the area is a salient area according to different perceptions of different areas by the human eye.Saliency detection is widely used in images,but calculating the spatiotemporal salience of video is an emerging and challenging issue.The purpose of video saliency detection is to detect salient targets in a video frame by frame,which is different from image saliency detection as follows:First,video saliency detection requires more consideration of motion characteristics.Second,the motion cues of the salient objects in the video provide clues to the salient regions.However,the motion of the background brings difficulties to the locating objects.Third,the motion of the object in the video is continuous,and it is necessary to ensure the consistency of the salient object in the global space-time.Video saliency is very valuable in many applications,such as video reconstruction,video object tracking,and video object segmentation.This paper first proposes an bootstrap learning video saliency detection algorithm based on simple frame selection.The algorithm first preprocesses the video frames,clusters them at the pixel level,uses a 138-dimensional feature descriptor to represent each pixel,and then uses the gPb-owt-ucm method to hierarchically segment the video frames to calculate the video frame clustering.Regional similarity.Then this paper proposes a sorting scoring standard for measuring video simple frames and sorting the video frames.Finally,in order to ensure that objects of simple frames are common throughout the video,an energy function is constructed to obtain truly reliable simple frames.After obtaining the simple frame,based on the initial saliency map,the robust foreground features and foreground labels are obtained.After inputting the multi-core bootstrap learning model,the final saliency map is obtained,and the motion features are spread on the entire video set to obtain the final salient map.Then an adaptive space-time structured low rank matrix decomposition video saliency detection algorithm is designed.The algorithm is based on the structured low rank matrix decomposition model,which performs superpixel segmentation on video frames,uses superpixels as matrix rows,extracts 58-dimensional spatiotemporal features of video frames as columns,and decomposes low rank matrix in groups of 20 frames.The super pixel constructs its spatial relationship with spatiotemporal features,and the time interval between the frames is constructed by the SIFT stream and the corresponding neighbors,forming a tree structure regularization term of time and space and a Laplacian regularization term.At the same time,before the decomposition of low rank matrix,an adaptive selection algorithm based on appearance similarity and background static degree is designed.Finally,the ADM algorithm is used to solve the sparse term of the model to obtain the final saliency map.Finally,based on the video significant detection method,this paper proposes a video segmentation model,which sets the video object segmentation into a pixel labeling problem with two labels(foreground and background),and defines an energy function for marking all pixels.Including:salient items,appearance items,position items,smooth items.The graph cut method is used to calculate the optimal binary mark according to the energy function,thereby obtaining the final segmentation result.This paper conducted a comparative experiment on the SegTrackv2 dataset.Through the experimental comparison,the video saliency detection algorithm proposed in this paper obtains higher accuracy in the data set.Compared with the traditional algorithm,the algorithm has obvious advantages in PR curve and F-measure value.At the same time,this paper compares the segmentation result based on the algorithm with other segmentation algorithms,and also obtains higher precision.