Saliency Detection Algorithm Based On Global Optimization And Local Enhancement

It is particularly important to extract valuable information from images,which provides the necessary premise and technical foundation for visual information processing.The saliency detection is one of the effective methods to extract useful information.Inspired by the human attention mechanism,saliency detection is desired to detect the areas of human visual interest.As a preprocessing step,saliency detection is able to reduce the difficulty in subsequent complex image processing tasks.Therefore,the study on visual saliency detection is of great significance and value in both sense of research and application.The saliency detection algorithm based on deep learning can obtain finer detection results compared to traditional methods.However,the current detection methods are incapable to obtain either accurate position or detailed information of salient target.Our research focuses on two aspects as follows:(1)Research to solve the problem of large changes in scale and inaccurate positioning of significant targets in complex scenes with existing methods.This paper proposes a saliency detection algorithm based on a multi-level global information propagation model.Firstly,we propose a multi-scale global feature aggregation module at a higher level in VGG;secondly,the global information extracted from multiple levels is fused;thirdly,the high-level global semantic information is merged with lower-level features in a feature propagation manner.These operations can extract adequate multi-scale context information to enhance the global expression,while avoid the loss of this information when it is passed back to the lower layers level by level.The global context information is of great importance to strengthen the positioning of salient targets,and restrain irrelevant background interference.Compared with other methods on four benchmark datasets,the results show that the algorithm is superior to several classical image saliency detection methods in accuracy,recall,F-measure value and MAE.(2)Research and solve the problems of incomplete detection area and blurred boundary of salient target obtained by existing algorithms.To this end,this paper proposes a saliency detection algorithm based on the pixel-level feature information enhancement model.Motivated by the idea of hierarchical processing,we integrate feature information in different levels as well.Firstly,the pixel-level attention module and the progressive boundary refinement structure are used to strengthen the spatial local details and target boundary contours.Secondly,the feature pyramid attention module is used to extract more discriminative global information.While these operations optimize the global guidance information,they also fully enhance the local detail features,and play a good role in improving the ambiguous spatial local details and the target boundary.Experimental results illustrate that the algorithm can still highlight targets in complex scenes and achieve more accurate salient target positioning.Compared with traditional and deep saliency detection algorithms,it performs better on four data sets including ECSSD and PASCAL-S.In summary,from the perspective of optimizing global and local enhancement,this paper constructs a multi-level global information transfer model and a pixel-level feature information enhancement model for saliency detection.The experimental results show that the proposed method has better detection performance,and its F-measure values are as high as 0.935 and 0.814 on the ECSSD and DUT-OMRON data sets,respectively.This provides new research ideas for saliency detection in terms of retaining detailed information and enhancing the ability to extract global context information.