Research On Saliency Detection Based On Fully Convolutional Neural Network

The image processing technologies corresponding to digital image which is one of the most important carriers conveying thoughts and emotions,have been widely applied to medical engineering,communication,military security and aerospace industry.However,the key point lies in how to effectively process data with limited resources.Visual saliency region detection can simulate human eye system to predict areas that human takes priority over,thus providing effective pre-processing result for subsequent tasks to improve efficiency.Applying neural network for saliency region detection can avoid explicitly features extraction and presentation in early stage,and implicitly learns the features needed for saliency regions during training,and the trained network presents generalization.Firstly,this paper introduces the research background and the development status of salient region detection,then analyzes the basic concept of visual attention mechanism and the two kinds of visual attention model.Secondly,this paper presents a new saliency detection algorithm based on fully convolutional network(FCNN)and Graph.This algorithm obtains an initial saliency map through trained FCNN.Based on the initial saliency result,graph theory is utilized to strengthen the inner relationships among neighboring areas,which helps to optimize the initial result.The AUC value and F-Measure values of this method on SOD and DUT are superior,and capable of capturing salient regions better.Then,this paper presents a saliency detection algorithm based on FCNN and multiple kernels learning(MKL).This algorithm digs out commonness of saliency areas among different images through FCNN saliency detection structure.The balance among different features of saliency areas within an image are further learned through supervised MKL.The AUC value and F-Measure value of this method on SOD and DUT dataset are superior,indicating the proposed algorithm can capture the saliency areas correctly as well as reduce the subjective intervention,at the meantime the proposed method has good generalization ability.Finally,the paper is summarized and future research prospects are made.