Salient Object Detection Based On Global And Local Perception

It is a problem that computer vision must solve to efficiently extract the important information of interest from the natural scene.Human vision system deals with this problem by means of visual attention mechanism,which involves the eye perception mechanism such as feature integration,visual search and attention transfer,spatiotemporal information integration,local ? whole / whole ? local integration.The human eye can not only realize the global perception of the environment effectively,but also has the ability to observe local fine features.The integration mechanism of global and local information in human brain is still unclear,which needs to be further explored.In order to learn from the perception ability of human eyes,this paper discusses the detection of salient objects in images from the perspective of global and local perception and their integration.(1)Global perceptual salient object detection is based on multi-level feature fusion.In order to integrate the complementarity of low-level detail features and high-level semantic features for salient object detection,we use the encoding decoding convolutional network as the backbone structure,and propose the global perception model of multi-level feature fusion.The encoder extracts the rough features from the local details to the global by convolution layer by layer,and then the decoder integrates the features to extract the salient objects in the image.Finally,using the overlay fusion operation earns the global fusion perception map,and achieves the global perceptual salient object detection.In the open data,the test results show that the multi-level feature fusion improves the detection accuracy of salient objects.And the model is also better than other classic salient object detection models.The model can also be used for fixation prediction.The test results show that the performance of the model is better than the classic visual focus prediction model in SIM and EMD.(2)Global to local visual search guides salient object detection.In order to imitate the human eye's attention mechanism to objects,this paper proposes a salient object detection model which integrates global and local perception based on visual search guidance.On the basis of the multi-level feature fusion map of the global perceptual,MSER algorithm is used to intercept the local hot sub areas in the early stage of the global perceptual map with the purpose of improving the integrating capacity of the global and local perception,generating the visual guidance search path,and fusing the local fine perceptual map to get the salient integrating object detection results of global and local perceptual through SENet network step by step.This model is similar to the human eye observation process from coarse to fine and iterative observation behavior.It shows excellent performance in the salient object detection experiment.The experimental results show that our model is better than other models in F,and better or close to other classical models in S and MAE.(3)Salient object detection is improved by enhancing the capacity of fine perception of local boundary.In order to get further improvement of the boundary definition of salient object detection,a mixed loss function is made by us,including cross entropy function,structure similarity function and overlapping rate loss function,then using in-depth supervision learning from three levels which are pixel-block-feature graph,which make the boundary of salient object closer to the real boundary of salient object.In the detection results of six salient object detection data sets,MAE has been greatly improved.It is a good guidance for the local boundary processing of the model,and optimize the object edge detection results by using the loss function of structural similarity and the loss function of overlap rate.From the perspective of global and local perception and their integration,the global perceptual salient object detection based on multi-level feature fusion is brought up.On the basis of the introduction of the visual search mechanism,a salient object detection method based on global and local visual search guidance is proposed.On the premise of local fine perception of boundary,the ability of local fine perception of salient object detection model is improved.The above research is an exploratory work for the modeling of human like attention mechanism,which has great reference value for the related research in this field.