| Salient object detection(SOD)which is a fundamental task for computer vision aims to detect the most conspicuous objects or regions in an image.SOD often serves as the pre-processing step for many computer vision tasks such as image segmentation,visual tracking,person re-identification,image captioning and so on.In SOD,Multi-scale features fusion plays a critical role in salient object detection.Most of existing methods have achieved remarkable performance by exploiting multi-scale features fusion strategies.However,an elegant fusion framework requires expert knowledge and experience,heavily relying on labor trial and error.How to automate this process has become the purpose of this paper.A multi-scale feature fusion framework based on neural architecture search is proposed,which is mainly divided into two stages: search and verification.The first stage will conduct preliminary search in a small dataset,and the verification training will be carried out on a large dataset after a suitable model is searched.In this framework,firstly we design a novel search cell,which can automatically decide multi-scale features aggregation.Rather than searching one repeatable cell stacked,we allow different searchable cell to flexibly integrate multi-level features.Secondly,we design a new search space,and fully consider that multi-scale features depend on different degrees of spatial attention mechanism and channel attention mechanism in the basic search space.The proposed search space effectively focuses on the most prominent indigenous regions and mitigates incomplete object structures or over-predicted foreground regions caused by progressive fusion.Finally,we propose a progressive polishing loss to further obtain boundaries by penalizing misalignment of salient object boundaries.Moreover,this loss increase the networks’ confidence for the boundary of salient objects.We conduct experiments on 5 datasets and compare the proposed method with 13state-of-the-art SOD methods under 4 evaluation metrics.On the DUTS dataset,the F-measure reaches 0.856,and on the more difficult databases DUT-OMRON and PASCAL-S,it also achieves the best metrics of 0.85 and 0.776,which demonstrates the superiority of the proposed method.Besides,we achieve the best performance only using 3 hours. |
