Research And Application On Salient Object Detection Based On Multi-feature Fusion

Salient object detection(SOD)is the method that computer simulates human visual attention to detect the most visually distinctive objects or regions in an image.Earlier salient object detection methods were based on hand-crafted features,for example color,contrast or the position.The research on deep Convolutional Neural Networks(CNN)has witnessed substantial progress recently.This end-to-end CNN can extract multi-level and multi-scale features.Benefiting from strong extraction ability based on CNN,the saliency map gets better in complex and low contrast cases.To overcome the drawback of the saliency map can not be highlighted and the boundary is blur,we do the following works:(1)First,the principle of visual perception and the human visual Attention are introduced,including contrast,center-surround,central prior and learning-based Attention.And then we introduce VGG and Resnet,analyze the multi-scale model and compare the different ways of feature integration between deep and shallow feature map to discuss the best way in SOD task.Finally we introduce our multi-scale network.(2)To overcome the drawback of the multi-scale network is redundant and can not utilize the feature effectively.We propose a multi-scale method based on adjacent feature fusion.Firstly,We use the Resnet-50 to extract multi-level features,adjacent levels are input into the Feature Aggregation Module(FAM)to employ the multi-level information.In addition,we combine cross entropy loss with consistency loss to generate different resolution saliency maps.The information is further integrated by the Fusion layer.The model is compared with 7 methods on 5 datasets.For visual comparison,Our proposed method not only highlights the salient regions but also has clear boundary.More importantly,it can effectively suppress the background noise.Experimental results demonstrate that our proposed method against 5 state-of-the-art learning based methods.The FAM can utilize the features more effectively.The combined loss function can handle the pixel imbalance between fore-and back-ground region caused by various scales of object.(3)To solve the problem of earlier methods,such as long positioning time and large amount of calculation,We propose a multi-feature fusion network based on Attention.On the one hand,the candidate salient object based on human visual attention as the input to simplify the neural network.Specifically,the improved COV method is used to automatically locate the candidate regions of optic disc by integrating background prior.We compare the accuracy of localization and visual performance with different methods.On the other hand,the optic disk is further refined on the side output of multi-scale by way of integrating attention module.The feature of side output is enhanced by the attention to effectively learn the features of optic disc,suppress the background noises.Our proposed method can generate clear boundary,laying a foundation for the subsequent diagnosis.