Multi-scale Features Fusion Network For Salient Object Detection

Salient object detection aims to identify the most visual y distinctive objects from a single input image,playing an important role in various computer vision applications.However,detecting the salient object accurately is a chal enging research task since the definition of the salient object in the input image is determined by many mutual y affected factors,including image structures,object semantic meaning,and context information.Despite promising results from recent works based on deep learning,existing works are still not sufficient to incorporate momentous global context to detect salient objects due to the limitation of the receptive field.To address the above problems,we present two multi-scale features fusion networks to enlarge the receptive field of features and enhance the global context of features,which could effectively improve the performance of detecting salient objects.Our contributions can be summarized as follows:1.We present a novel deep network to aggregate the attentional dilated features(AADFNet)for salient object detection.There are two contributions to AADF-Net.First,we leverage the attention mechanism to build an attentional Dense ASPP(AD-ASPP).The AS-ASPP first uses multiple dilated convolutional layers to extract multi-scale features and enlarge the receptive field of features,then it exploits the complementary information between different dilated features by attention mechanism to enhance the global context.Second,we develop an aggregation network to integrate the refined features to predict finer results by formulating two consecutive chains of residual learning based modules.2.We present a structure-aware dual pyramid network(SDPNet)for salient object detection.There are two contributions to SDPNet.First,we develop a structure-aware dual pyramid module by formulating self-attention mechanism into the sub-regions based contexts in the spatial and channel-wise dimension,respectively.The dual pyramid module could extract multi-scale features and enhance the global context and structural context of features without sacrificing the spatial resolution of features.Second,with the structure loss composed of edge loss and adversarial loss,the edge structure context and global structure context are introduced to refine the structural coherence of the predicted results.