Research On Salient Object Detection Method Based On Deep Learning

In recent years,salient object detection has received extensive attention and research.It has extensive applications in the fields of visual identification,monitoring,tracking and positioning and so on.The deep learning-based method usually learns saliency features by the fully convolutional neural network.However,it is easy to get the poor predict result in scenes with low contrast between foreground and background,since the pixel of salient objects is difficult to capture.The main reason is that the background pixel is similar to the foreground or the background occludes the foreground.Since these hard-capture pixels exist in the saliency boundary region,boundary-aware methods are proposed to tackle this problem.The boundary-aware method generates the saliency boundary by adding a boundary detection branch,and uses the saliency boundary to guide the network to pay more attention to the region of the boundary.In order to guide the saliency prediction better,this paper studies the boundary-aware method based on the complementarity between the saliency object and the saliency boundary.The content of this paper mainly includes:First,this paper proposes the RTGRNet for the problem that the saliency boundary is difficult to accurately predict.RTGRNet is composed of a Prediction Module and a Recurrent Refinement Module.The goal of the prediction module is to extract salient features and salient contour features simultaneously by using the Two-stream Guided Refinement Block.After that,The Refinement Unit in the Recurrent Refinement Module utilizes the guide block to complete the mutual guidance between these two subtasks.More refinement units are stacked to refine the two subtasks at the same time,thereby improving the performance of RTGRNet.Second,this paper proposes the CMGNet for the problem that the introduction of irrelevant information and the ignorance of the difference between two guidance paths.CMGNet first designs a Correlation-based Mutual Guidance Module,which is composed of a Local Saliency-Correlated Module and a Non-Local Hybrid-Correlated Module.The former utilizes the saliency information around the boundary to eliminate the non-saliency boundary information,and the person utilizes the global hybrid information to enhance the attention to the boundary position in the saliency prediction.Finally,CMGNet designs a Dissymmetrical Dual-Stream Prediction Unit and stacks the unit to further improve the prediction performance.Third,Experiments on six public datasets show that the proposed RTGRNet,CMGNet achieves state-of-the-art results in terms of both F-measure,MAE and S-measure evaluation measures without any pre-processing or post-processing.The results show that the RTGRNet is 0.5% and 1.1% higher than the other methods in the F-measure and S-measure scores of the DUT-OMRON dataset.Meanwhile,the CMGNet is 0.8% and 0.7% higher than the other methods in the F-measure and S-measure scores of the DUTS-test dataset.At the same time,the Guide Block,Refinement Unit,Correlation-based Mutual Guidance Module and Dissymmetrical Dual-Stream Prediction Unit proposed in this paper can significantly improve the performance of prediction.