Multi-feature Fusion For Salient Object Detection

Salient object detection aims to automatically identify the most attractive regions in images like human visual systems.Salient objects are of various sizes,therefore multi-scale feature fusion has been proved to be a key role in achieving remarkable results.With the development of 3D sensors,multi-modal data is gradually employed to improve the performance of salient object detection.With these multi-scale and multi-modal features,how to find an effective way to fuse them becomes a question worth further study.Considering the RGB data,we first explore the fusion strategy for multi-scale features.We propose an RGB salient object detection method via hierarchical feature fusion.This method uti-lizes multi-channel feature maps for fusion to fully maintain the representation ability of deep con-volutional neural network.Three fusion stages are designed to enrich the fusion process.In addi-tion,a global prior module is introduced to extract the global feature,whose receptive field covers the whole image.It helps the model to highlight the salient objects more accurately and completely.Then,we introduce the depth information to solve the situation when objects share similar appearances with their surroundings and propose an RGB-D salient object detection method via adaptive fusion.To learn the complementary relationship between multi-modal data,this method employs an adaptive fusion module,which applies an adaptive weight map to selectively fuse the effective regions of RGB and depth information.Meanwhile,an edge-preserving loss function is design to shape the salient objects more coherently.The experiments on publicly available RGB and RGB-D datasets illustrate the effectiveness of our multi-scale and multi-modal fusion strategies.The method via hierarchical feature fusion achieves state-of-the-art performance on RGB salient object detection and the method via adaptive fusion outperforms other approaches on RGB-D salient object detection.