A Deep Embedding Features Based Method For Salient Object Detection

Salient Object Detection,as a task in computer vision,could simulate human visual mechanisms to automatically find areas of human focus in an image.Thereby,Salient Object Detection can be used as a pre-processing process for many other computer vision tasks,such as Visual Object Tracking,Person Re-identification,Semantic Segmentation and Visual to Caption.Traditional Salient Object Detection mainly relies on the use of some low-level visual features and clues such as color,light intensity,contrast,and other prior knowledge.These features are based on some prior knowledge of human visual attention,thereby could hardly capture some advanced semantic information.When dealing with tasks in complex scenes,these traditional methods are difficult to distinguish significant targets from the background.Recently,Deep learning based methods,especially those based on Convolutional Neural Networks have achieved good results in many computer vision tasks.It could extract multi-level features from images through an extremely deep network and use task-related features to achieve desired results.However,apart from producing features,convolutional neural networks can bring about a large amount of redundancy which will result in noise and interference.It remains a key issue how to design a reasonable network to effectively use useful features and filter out redundant features.To solve the afore-mentioned problems,we propose an embedding feature network based on attention mechanism to filter features that are helpful for saliency detection tasks.Considering that the current methods fail to solve the problem of feature redundancy.In this paper we propose to use preliminary results obtained with a lightweight subnetwork as input to the Feature Embedding Module and then add weights on the features of other subnetwork to filter out the interference in the foreground of the image and the redundant information in the background.Besides,the Convolutional and Pooling layers of the Convolutional Neural Network reduce the resolution of the original image,which will result in the lack of many details.Although the resolution of feature maps can be restored by upsampling operations,including linear interpolation or deconvolution,it will result in the boundary blur between the junction area of foreground and the background.Therefore,we present a deep-to-shallow Recursive Feature Integration Network to gradually improve the details of saliency prediction results.Finally,a Guided Filter Refinement Network is proposed to further improve the smoothness and consistency of the foreground part at the boundary areas.We compare our algorithm with fourteen state-of-the-art methods saliency detection methods on five large scale saliency detection datasets.The experimental results show that our algorithm can better deal with the interference of complex background and detect the whole foreground object.At the same time,in some quantitative experiments(Precision-Recall curves,mean F-measure values,MAE values),our method achieves state-of-the-art result.