Researches On Visual Salient Object Detection Algorithms Based On Feature Relationships

Recently,with the development of information technology and the advent of the era of big data,the amount of image shows an explosive growth trend.It has become an important research topic to quickly and effectively search useful information from the vast array of images and further improve the efficiency of information processing.When handling the complex real scenes,the human vision system can quickly search and locate the objects of interest.This visual attention mechanism enables humans to deal with the complex information quickly and efficiently.Researchers introduce the visual attention mechanism into the computer vision task,and further develop it as visual saliency detection.The saliency detection algorithm,devotes to realize this selection mechanism of human vision,can quickly screen out the important content and eliminate interference from irrelevant content,thereby greatly improving the efficiency of processing image content.Saliency detection,which can focus on the essential image content,has become as a popular computer vision research task.Moreover,as an important pre-processing method,it has been widely used in various computer vision tasks,such as object tracking,object segmentation,image editing and so on.Visual saliency detection can be divided into two types: the task-driven top-down approach and data-driven bottom-up approach.Most existing visual saliency detection tasks focus on the data-driven bottom-up approach.According to the purpose of task,saliency detection can be divided into fixation prediction and salient object detection.This paper mainly researches the bottom-up salient object detection.Considering three intrinsic relationships of features in salient object detection,three salient object detection algorithms are constructed,respectively.The main innovative work and results of this paper can be summarized as follows:1).This paper explores the spatial adjacency relationship of hand-designed features and designs a saliency detection algorithm based on the graph propagation model with non-local connections.Currently,most graph-based propagation models only consider the local neighbor relationship of nodes.However,when the features of the internal area of a salient object vary greatly,the construction of a graph model only based on the local relationship is often not accurate enough,and it is easy to cause inconsistency in the salient area,even erroneously suppress the salient area.To overcome these problems,this paper introduces non-local relations into the graph model and constructs a graph propagation model with non-local connection relations.Specifically,the model first calculates the more accurate initial saliency maps based on the optimized compactness and contrast feature calculation definitions,and then uses a saliency-bias Gaussian model to further refine them.On the refined saliency maps,a two-layer sparse graph model with non-local connections is constructed,and the saliency results are propagated according to the rules of random walk.Finally,the saliency maps after propagation are linearly fused,and then a self-update algorithm is designed to further optimize them and obtain the final saliency maps.The introduction of non-local connection relations overcomes the shortcomings of the graph model constructed only with local relations,so the algorithm can obtain the result with good consistency.A series of quantitative and qualitative evaluation results proved the effectiveness of the proposed algorithm.2).This paper explores the feature priority relationship in saliency detection task and sums up the origin of this priority relationship as feature generation relationship.According to the feature generation relationship,we design a cascaded hourglass neural network for saliency detection.Currently,most saliency detection algorithms based on the deep learning consider that different levels of features play an equally important role.However,in the task of saliency detection,high-level semantic information is closer to the definition of saliency,corresponding to the most internal regions of salient objects,thereby playing a major role and in a natural dominant position.This paper explores the origin of this priority in neural networks and summarizes it as the feature generation relationship.Inspired by the feature conversion capability of pooling operations,this paper proposes a cascaded hourglass feature fusing neural network.The network equips with three sub-modules to integrate multi-scale features gradually.Specifically,the algorithm first designs a cascaded multi-scale context feature extraction module to extract useful context information.Then,the algorithm uses a series of cascaded pooling operations to form the hourglass feature fusing modules,and further connects these modules hierarchically to screen and fuse features at different scales.Finally,the algorithm designs a cascaded decoder to obtain the final saliency result.Comprehensively and sufficiently qualitative and quantitative evaluation results have proved the advance and effectiveness of the algorithm.Simultaneously,the designed algorithm is relatively lightweight and can run efficiently at the real-time speed during forward inference.3).This paper further studies the priority of features at different scales,based on the feature generation relationship,we further develop the priority as a feature principal and subordinate relationship.Moreover,in terms of the relationship,we design a “leader-follower” neural network for saliency detection.Feature subordinate relationship regards the high-level and low-level feature sets in the neural network as two decision variables with superior and subordinate relationships: high-level features are used as superior leaders;low-level features are used as subordinate followers.In this paper,the internal relationship is further represented as a Stackelberg game model.Based on the model,we design a saliency detection algorithm based on the “leaderfollower” feature fusing network.The algorithm includes two sub-modules,namely,a“leader-follower” multi-scale contextual information attention selection module and a feature fusing module.Extensively qualitative and quantitative evaluation results have proved the effectiveness of the proposed module and algorithm.While the algorithm has achieved advanced performance and excellent efficiency,it can also better cope with the task of video saliency detection,showing the superior portability.Finally,we deploy the algorithm on the mobile terminal,and the test results in the real scene prove the practicability of the algorithm.