Research On Image Salient Object Detection Algorithm Based On Deep Learning

As an important carrier of recording and expressing information,images play an increasingly important role in the society with deepening informatization.Human visual system can quickly extract valuable information from images,but with the continuous popularization of various terminal mobile devices,only using human visual system to process massive images has been far from meeting the needs of society.So researching and developing computer vision system is an urgent solution for social development.Image Salient Object Detection is one of the key points in computer vision.Its task is to locate and detect the most obvious and noticeable objects or regions in the image.The deep salient object detection methods use the rich multi-level semantic features extracted from convolutional network to make the model performance surpass that of traditional detection methods.However,the detection of deep models still has some problems such as uneven prediction of salient pixels and inaccurate prediction of edges.From this perspective,this thesis proposes corresponding solutions.The main contributions of this thesis are as follows:In order to fully mine comprehensive and accurate saliency information of images,a deep salient object detection model based on multi-attention features is proposed in this thesis.The model includes progressive channel compression module,residual decoding module,position attention module,aggregation and mixed attention module.Progressive channel compression makes full use of convolutional operations to extract channel interaction information and unify the number of channels encoding features.The residual decoding module transmits the high-level semantic guidance information to the low level layer by layer,and fuses and refines the potential details of the multi-level convolutional feature map.The location attention module further mines the important spatial information of each level to obtain more comprehensive salient pixels.The aggregation module enriches the context between hierarchical features.The mixed attention module filters background interference information to improve detection accuracy.Experimental results show that the model has excellent detection performance.Aiming at the two challenges of salient object detection in complex scenes:missing local details of salient object and misidentification of non-salient region,this thesis proposes a deep salient object detection method based on average and max pool.The model consists of four parts:original feature extraction,average and max pool module,feature fusion and deep supervision mechanism.First,the model extracts multi-level original features.The average and max pool module then extracts multi-level complementary contextual features in spatial and channel dimensions.Then the feature fusion module fully integrates global spatial information and local detail information.Next,the model uses the top-level semantic guidance information of two top-down feedback paths to improve the accuracy of detection.Finally,the deep supervision mechanism improves the generalization performance of the network on different datasets.Experimental results show that this method has excellent performance of salient detection.In order to solve the problems of uneven detection and unclear edge details,a twostage dual U-shaped network is proposed in this thesis.This method uses a dual U-shaped network to model salient and edge convolution features globally in stages.In the first stage,the coarse salient and edge features of five levels are extracted based on the single-encoding and dual-decoding U-shaped network to locate the salient object.In the second stage,the coarse features are refined layer by layer from high to low based on the independent salient and edge U-shaped network,and the hierarchical correlation of the two features is mined to make full use of the semantic features of different levels and enrich the details of feature maps.Finally,the feature fusion module fuses the two types of features layer by layer to generate the final saliency map.The algorithm performs well in challenging natural scene images such as multi-target,low contrast and complex background.