Multi-attention Guided Feature Fusion Network For Salient Object Detection

Saliency detection simulates the primitive visual attention mechanism where the first step to perceive a scene is to filter out areas of interest and then move them into the center of the retina,thereby reducing the complexity of the scene.There are two branches in saliency detection.One is eye-fixation detection whose purpose is to detect the position of fixation points.The other is salient object segmentation whose purpose is to separate the significant objects in an image from the background completely.The latter is used in a wider range for practical applications.It is often used as a preprocessing step for other visual tasks and has attracted rising attention of scholars,which is also the subject of this article.Early conventional saliency detection methods were often based on hand-crafted features,such as color,intensity and contrast,which are not representative enough and have poor generalization.In recent years,numerous saliency detection networks have been emerging unceasingly due to the great success of convolutional neural networks in other visual tasks,which have improved the accuracy of the saliency task greatly and become the popular methods in saliency detection.Although the CNN-based methods break the bottleneck of the development of saliency detection,there are still certain problems.The pooling operation in CNNs reduces the resolution of the feature maps and discards spatial detailed features,which are very important to saliency detection.Therefore,in order to learn a more powerful feature representation,scholars usually use feature fusion mechanism to relieve the space scale problem.However,not all features are effective for saliency detection.Some cluttered background features may even cause interference.In order to solve this issue,the paper designs a saliency detection network based on the multi-attention mechanism.First,it utilizes a novel channelwise attention feature fusion module.Second,we combine spatial attention and self-attention to generate a positional attention model.Spatial attention can highlight salient features and suppress background features and self-attention can capture sufficient contextual information and long-distance pixel dependencies.We compared the entire network with the other 13 algorithms on five public benchmark datasets.According to the PR curve,F-measure and MAE,we prove the effectiveness and advancedness of the proposed algorithm.