Saliency Detection Based On Local And Global Information Fusion

Visual saliency detection mainly uses computer technology to imitate human visual attention mechanism,extracts the information of interest in complex image scenes and give priority to processing,which is helpful to improve the efficiency of image processing.It has been widely used in video compression,image recognition,retrieval and other fields.The existing detection algorithms consist of traditional detection algorithm and detection algorithm based on deep learning.Traditional detection algorithms only rely on low-level features,resulting in low confidence of saliency map.Detection algorithms based on deep learning could improve the confidence of salient objects,but many pooling operations will lead to fuzzy contour of salient object.In this paper,a saliency detection model based on local and global information fusion is proposed,which can effectively obtain the salient targets in complex scenes,and improve the robustness of the saliency detection model.The main work of this paper includes:(1)Aiming at the problem of saliency target edge blurring causing by the large number of pooling operations in fully convolutional network,a new detection model that refine salient object edges through combining multi-scale feature maps is studied.This model extracts feature maps on different scales under low-level convolution groups firstly,then according to the complementary relationship between low-level feature information and high-level abstract semantic features,obtains the global saliency maps based on weighted fusion of feature maps on different scales and saliency maps detected by the full convolutional network,finally optimizes the saliency target boundaries.The comparison experimental results on the standard datasets show that the model can enhance the edge of the detected objects.(2)In order to further optimize the boundary of global saliency map,a visual saliency detection model fusing global and local information is studied.Firstly,the local saliency map is obtained based on the multi-scale SLIC super-pixel segmentation algorithm.Then fused with the global saliency map based on the Hadamard product fusion model to obtain the initial saliency map,the conditional random field is used to smooth the boundary contour of salient targets.Experimental results show that the model can effectively suppress the interference of complex scenes in the image,and detect salient targets more accurately and completely.(3)The saliency detection model implemented in the paper is applied to the pedestrian segmentation in complex traffic scenarios.The pedestrian segmentation model fusing saliency map and GrabCut algorithm is studied.The obtained saliency map is used to initialize the mask of Grabcut algorithm,and saliency is consided as the constraint condition of data items in the energy function.The experimental results on the dataset PEDD show that the model makes efficient use of saliency map to determine the location information of pedestrian,overcomes the shortcomings of Grabcut algorithm,and improves the robustness of pedestrian segmentation model.