Salient Object Detection Based On Backbone Enhanced Network

The CNNs with encoder-decoder architecture has shown powerful ability in semantic segmentation and it has also been applied in saliency detection.We usually used the parameters of the backbone network which have been pre-trained on the Image Net and retrained using the new dataset to adapt the new task.But the retraining will weaken generalization of the backbone and result in over-fitting,especially when the scale of the training data is small.To make a balance between generalization and precision,and to further improve the performance of the encoder-decoder network,we proposed backbone enhanced network.A encoder with structure of dual backbone networks to extract more diverse feature maps.A connection module based on improved Res2 Net to efficiently fuse feature maps from the two backbone networks and a decoder based on weighted multi-scale feedback module to perform synchronous learning.Our approach is extensively evaluated on six public datasets,and experimental results show significant and consistent improvements over the state-of-the-art methods without any additional supervision.In addition,we also explore the saliency instance segmentation which is an extend area of saliency object detection.Combining with the semantic segmentation and object detection network to generate the bounding box in order to output high-quality instance.We consider that the saliency object is class-independent,so we re-designed the network through improved detection and segmentation branches.Our segmentation branch not only pays attention to the detection window,but also the surrounding context.By assigning transition weights,we can better distinguish the salient object from the background and improve accuracy of the network.Our method evaluated on a public benchmark dataset and showed that it outperforms the latest approaches.Finally,we conduct ablation research for each module to analysis each of network components and discuss how to improve performance.