Research On Saliency Detention Based On Fully Convolutional Networks

Saliency detection refers to the detection and segmentation of the image consistent with human visual perception and can express the main content of the image objects with precise contours.Saliency detection is widely used in many fields,such as image classification,image retrieval,image compression,etc.The traditional saliency detection methods have some problems such as insufficient feature learning and poor robustness detection effect.With the development of deep learning technology,relevant researchers began to apply convolutional neural network to saliency detection of images,and achieved good results.However,there are still some problems such as low accuracy,incomplete object region and blur boundary.To solve these problems,this paper proposes two kinds of network models for optimization.To solve the problem of insufficient features in the saliency detection model.A saliency detection algorithm of multi-scale feature fusion based on fully convolutional neural network was proposed.It only relied on the deep features to generate the final feature map through upsampling or deconvolution,which lost a lot of detail information.The overall result of the image was blur,so shallow features were added to make up for this shortcoming.In addition,two different methods are adopted in the multi-scale feature extraction,one uses a common convolution operation to extract features,while the other uses a dense connect method to extract features.To solve the problems of incomplete region and blur boundary of objects detected in complex scenes,a network model based on feature guidance mechanism was proposed,and the model was based on FPN(Feature Pyramid Networks)structure.However,FPN has the problem of deep semantic information being diluted in the process of transmission from top to bottom.In order to alleviate this problem effectively,our model repeatedly uses the deep semantic features to direct guide the shallow features,which can help the shallow layer to locate the object position and reduce the loss of information due to dilution.Experimental results show that the proposed algorithm achieves good detection performance on 6 widely used public datasets in both qualitative and quantitative analysis.