Salient Object Segmentation Based On Deep Metric Expression

Salient object segmentation is aimed to simulate the Human Visual System(HVS)and extract the most attractive and distinctive objects in images.Since it's beneficial to extract helpful information from massive data quickly,it has been widely used in many related fields.To better address the problem of salient object segmentation,in this paper,an end-to-end generic salient object segmentation model based on deep learning and metric learning is proposed.The main contributions of the paper are as follows:Firstly,since salient object segmentation is easy to be affected by the context of the image,the same object may belong to a saliency or a background according to different contexts.Saliency segmentation usually requires global information about the image,and thus multi-scale information is beneficial for more precise image segmentation.Therefore,we propose a deep learning network that combines low,middle,and high level feature to learn comprehensive feature information which makes the output boundaries of salient object fine-grained.Secondly,in this paper,a deep metric expression is employed to construct the loss function,that is,the network function approximation is used to redraw the semantic topological metric space.If the salient object segmentation have similar performance backgrounds based on low-level features(color,structure),the distances between salient object and background is closed,and it is difficult to extract the salient object.On the contrary,the proposed method adopts deep learning to extract high-level semantic features,and construct a non-linear topological metric space.In this space,the expression belongs to semantic distance,it is easy to extract salient object.Thirdly,multi-scale features and metric loss are integrated into this structure,which make the proposed algorithm is more robust to distorted images.The experiments prove that the performance of proposed algorithm is superior to other CNN-based methods on distorted images.It's noted that the proposed method does not train on distorted images is the same as previous works.During testing,the trained models are employed to test distorted images directly.Furthermore,the proposed model is trained from scratch and does not require pre/post-processing.Fourthly,in this work,we proposed a modified saliency detection model based on contour detection characteristics with guide filter as preprocessing layer.Our model has achieved satisfactory performance,which also can be well generalized on various applications.In comparative experiments,it's clear that the proposed method is very competitive to best performance methods,in terms of objective indexes including Recall,Precision and F-measure.In this paper,all the operations are based on pixel-level.In visual comparison,the boundaries of salient object are accurate and closer to Human Visual System.For the distorted images,all objective indexes of proposed method are better than state-of-the-art methods,the visual performance is better.