Research On Image Salient Object Detection Based On Deep Learning

With the wide application of deep learning in the computer field,the vigorous development of artificial intelligence technology has greatly promoted in recent years.How to use deep learning,a mainstream artificial intelligence method,combined with traditional computer vision tasks is a main research topic in the field of computer vision today.Due to the powerful fitting ability of deep learning,great progress has been made in the visual fields such as saliency detection,semantic segmentation,object tracking,object detection,and image captioning.Among these tasks,image salient object detection is a machine vision application for the purpose of detecting the interest region of the visual system,aiming at imitating the intelligent behavior of the human visual attention mechanism.This paper studies the related methods of image salient object detection,focusing on the methods based on deep learning.Deep learning methods build deep feature extraction networks by stacking multiple layers of multi-scale convolution,pooling and other operations.The feature learning ability of deep neural networks relies on a large number of weight parameters,and it is necessary to rationally design its network structure and train through a large amount of data to better handle various complex decision-making tasks.As far as image salient object detection is concerned,how to employ deep learning models to effectively integrate different scale features,suppress vague boundary phenomena,and enhance the robustness of complex scenes is the main challenge in this field.Based on the deep learning architecture,this paper attempts to construct a reasonable feature fusion method to enhance the effectiveness of feature fusion at different stages,maintain integrity of salient object and further refine the boundary of the object by using supervised training approach.The main work of this paper are as follows:In the first part,this paper explains the significance and trend of image salient object detection.We illustrate the research background and meaning combined with its real-life application scenarios.To sort out the development status in the field of image salient object detection at home and abroad,this paper reviews the traditional methods,and emphasizes on the methods based on deep learning.Then this paper summarizes the main development trends in the field of image saliency.In the second part,the introduction of the deep learning theory related to this paper,mainly includes the construction of deep learning models,mainstream backbone networks,and visual attention mechanisms,is mentioned.According to the introduction of the principle of deep neural network,we can deeply understand the basic theory of deep learning model,and get the theoretical support for the construction of the image salient object detection model proposed in this paper.In the third part,this paper proposes a Cross Stage Feature Interleaved Fusion Network(CSIFNet).CSIFNet constructs an effective cross-stage feature fusion architecture.This structure builds cross-stage information transmission through progressive feature fusion,which increases the effectiveness of low-level and highlevel feature aggregation.At the same time,through the stacking of modules in the stage of decoder,multi-scale information and feature responses are gradually aggregated to obtain the finally enhanced saliency prediction.In the fourth part,this paper proposes a Multi-level Feature Shrinking and Refining Network(MFSRNet).Based on the CSIFNet architecture,this new model replaces the original module design,introduces a new refinement module,and designs a new structure-aware hybrid loss function based on the original multilateral loss function,so as to further refine the boundary and strengthen the supervision in the optimization process.The CSIFNet and MFSRNet proposed in this paper are tested on five mainstream public datasets and analyzed on experiments.The experimental results show that these two models have excellent performance,strong robustness in complex backgrounds,and can output more refined boundary information of salient object.