Research On Salient Object Detection Algorithms With Fully Convolutional Models

A large number of detailed experimental studies by cognitive psychologists have shown that human has the ability to quickly capture objects that are salient or interesting in a visual scene.In order to make modern computers have a similar ability of information processing,more and more researchers have carried out relevant research on salient object detection.In recent years,salient object detection has become a hot research topic in the field of computer vision and im-age processing.The aim of salient object detection is to identify and segment salient objects or regions in the image that attract human visual attention.As an image preprocessing method,it can not only promote the understanding of the image scene,but also generally improve the per-formance of the downstream related visual tasks,such as image recognition,retrieval,segmen-tation,compression,restoration,etc.At the same time,with the rapid development of hardware technology,salient object detection is also playing a more and more important role in the field of cutting-edge applications,such as automatic driving,human-computer interaction,industrial robots,etc.Therefore,the research of salient object detection has a broad application prospect and profound scientific significance.Although great achievements have been made,the accura-cy and efficiency of salient object detection methods are still not satisfactory in clutter scenes,low-contrast imaging environments or when objects are not obvious.To handle the aforementioned issues,this thesis will study salient object detection algo-rithms by using effective deep learning technologies,especially full convolutional neural net-work.The main contents and research innovations of this thesis are as follows:(1)Based on the characteristics of different layers of deep convolutional networks,we pro-pose the aggregating multi-level convolutional feature framework for salient object detection.Specifically,in order to better fuse multi-level features,a resolution-based feature combination method is proposed.In this method,deep features at different resolutions are firstly shrunk or expanded appropriately to maintain the same resolution,and then features at different lev-els are weighted and adaptive aggregated.This method eliminates the resolution mismatch of multi-level deep features and enhances the representation of deep features.In order to promote the information interaction between multi-level features,this paper proposes a deep multi-level supervised learning method to train the network.In this training method,the supervision infor-mation is firstly applied to the side-outputs of the network,and then multiple loss functions are jointly optimized to complete the multi-level salient object detection.This method can signifi-cantly improve the model's ability to detect salient objects.The proposed algorithm has strong generality and can be transferred to other effective deep convolutional neural networks.Com-pared with other algorithms in the same period,the effectiveness of the proposed algorithm on the internationally public datasets is significantly improved.(2)In view of the fact that the model robustness is often ignored by existing salient object detection methods,this paper proposes an uncertain convolutional feature learning framework for salient object detection,from the perspective of convolutional feature ensemble.Firstly,the algorithm modifies Dropout based on feature elements,and the new representation can be con-sidered as the probabilistic ensemble of feature elements.By applying it to the convolutional feature maps,it can extract multi-level convolutional feature with uncertainties.Based on the convolutional encoder-decoder structure,the proposed model can be used to detect salient object in an end-to-end manner.At the same time,in order to effectively reduce the "checkerboard arti-fact" in the process of deconvolution up-sampling,a hybrid up-sampling method is constructed in this paper,which inherits the advantages of deconvolution and interpolation operations,thus improving the smoothness of the prediction.Compared with the previous algorithms,this al-gorithm has a significant improvement in model robustness,and achieves superior results in detection accuracy and generalization.(3)Most existing deep models lack of the use of complementarity and structured super-vision information,this paper proposes a salient object detection algorithm based on lossless feature reflection and structured loss function.The salient object detection task mainly focuses on the separability of foreground and background.Image intrinsic reflection and mirror trans-form can well decompose the information of the image.In view of this,the algorithm proposed in this paper first transforms the original image into a content-preserving mirror image through specular reflection.Then,a symmetric fully convolutional network is used to extract comple-mentary visual features,Then a multi-layer feature fusion strategy is used for feature interaction to promote feature fusion.Finally,accurate saliency maps are obtained through the supervised learning with a structured loss function.Through a large number of experimental analysis,the algorithm can achieve excellent performance on public datasets and greatly improve the bound-ary accuracy of salient objects.Finally,this thesis presents the problems and challenges that the current algorithms still face and the future research trends for salient object detection.