Research On Saliency Detection Algorithm Based On Joint Multi-level Features And Its Application In Image Segmentation

The purpose of visual saliency detection is to simulate the human visual attention mechanism through algorithms and locate key areas of images.As a basic step of image understanding,it models around the region of interest in the scene and finally endows the computer with visual perception ability.Therefore,the research on visual saliency detection is of great significance.Saliency detection mainly analyzes the salient region of the image by feature extraction and calculating the feature contrast.However,there are some problems in the detection results,such as inaccurate positioning of salient areas,unclear edges of salient areas,inconsistent salient values within salient areas and so on.This reduces the accuracy of saliency detection to a certain extent.Thus,the effect of subsequent image processing steps such as image segmentation is affected.This paper starts with the traditional saliency detection algorithm and the saliency detection model based on deep learning.The object candidates,low rank matrix recovery and multi-layer detection are taken as the research objects.It provides new ideas to improve the detection accuracy and robustness.The specific research work of this paper is as follows:(1)Existing prior information is insufficient in suppressing background regions and highlighting object regions.A saliency detection algorithm based on Edge Boxes and improved low-rank matrix recovery is proposed.First,in the Edge Boxes algorithm,the boundary connection information and prior information are introduced to re-correct the score of each candidate box.According to the score,a fixed number of candidate boxes are selected to frame all the positions in the image that may become salient objects.The non-maximum suppression is used to further filtered candidate regions.According to the center point of each candidate frame,the DBSCAN clustering algorithm is used to obtain the center of the salient object in the image.By extracting the minimum circumscribed matrix of the target,the location of the salient region of the image is determined.The above steps solve the problem of salient object localization.Finally,in the selected salient region,the final saliency map is calculated according to the improved saliency detection algorithm based on low rank matrix recovery.Compared with traditional methods,this method is not only more accurate in salient objects positioning,but also can reduce the amount of calculation to a certain extent.(2)The problem of saliency object positioning is solved.The algorithm still has the problems of unclear edge of the salient region and uneven salient value in the salient region.By introducing the hierarchical segmentation algorithm,region saliency detection algorithm based on improved hierarchical segmentation algorithm is proposed.Firstly,the architecture of hierarchical segmentation algorithm is introduced.The method of calculating the boundary probability in the original architecture is replaced by structured edge detection operator.At the same time,using the concept of multi-scale,the image contour signal is obtained by the weighted sum of multiple edge images.The object priori is applied to the contour region merging calculation.These measures ensure the efficiency and rationality of the hierarchical segmentation algorithm.Then,in the segmented region,the segmented region is optimized according to the salient value frequency histogram in each region.Compared with the results of other methods,the edge of the salient region is clearer and the consistency of the salient values within the region is higher.(3)Traditional saliency detection can not reflect the visual saliency driven by high-level semantic objects.A dense U-net saliency detection model based on adaptive Inception structure and attention mechanism is proposed.The algorithm is based on the U-net network and fuses Inception structure,attention mechanism and salient information based on low-level features.In the coding part of U-net network,different Inception structures are set according to different levels.It overcomes the shortcomings of single scale features and retains different levels of features from the perspective of multi-scale.The saliency map based on low-level features is resized to be the same size as the output of the corresponding decoding layer,and feature fusion is performed.In the skip connection,a global attention gating module with enhanced features is added.The feature map obtained by the Inception module of the encoding part is controlled by the high-level features in the decoding layer.It can realize the purpose of weight control of low-level feature map.The detection model fuses low-level features with high-level features,and it is able to recover the lost details of the image.The obtained results are closer to the ground-truth map in terms of visualization and evaluation metrics.(4)Traditional image segmentation methods can not accurately segment the salient areas of the image.It does not conform to human visual characteristics.A GrabCut image salient object segmentation algorithm combining visual saliency is proposed.The saliency map obtained by the saliency detection algorithm is used as the mask of the graph cutting method.It does not need human interaction,and automatically completes the distinction between image object and background.The algorithm changes the operation unit from pixel to super pixel,which improves the operation efficiency.At the same time,adaptive parameters are added to the smoothing term to further improve the spatial consistency of the segmentation results.Finally,the salient object segmentation image is obtained through iterative calculation.For the pictures collected in the real scene,the proposed algorithm can also clearly segment the image object,and it proves the universality of the algorithm.