Research On Saliency Object Detection Algorithm Based On Feature Fusion And Attention Mechanism

Salient object detection aims to extract the most attractive targets or regions in an image which is often used in large computer vision tasks such as object detection,semantic segmentation,and object recognition.Salient object detection has become one of the hot topics in computer vision.Traditional methods commonly use a large amount of saliency prior information and hand-crafted features,which is incapable of detecting salient object in complex scenes or structures and has generalization capabilities.With the rapid development of(Deep Neural Networks,DNNs)in recent years,saliency object detection algorithms have seen a tremendous improvement in performance.However,feature extraction and fusion using the nested DNNs to improve detection accuracy and effective is still a challenging issue.To address this issue,we focus on saliency object detection algorithms based on multiple feature fusion strategies,and the main research work and results are as follows:(1)An improved Enhanced Feature Progressive Polishing Network(EPFPN)is proposed to address the common problem of missing spatial position information of targets fused with edge details and background in current saliency object detection algorithms.EPFPN corrects the chaotic feature information and refines the prediction with a recursive manner from deep layers to shallow layers.The network is mainly composed of five parts,Backbone,Feature Transmission Module,Feature Refinement Module,Dynamic Convolution module and Feature Fusion Module.First,features are extracted through the backbone network,and then are changed into the same dimension through the transmission module.Then,the feature information is gradually optimized through the nested feature refinement module of dynamic convolution,and the high-level semantic features are directly integrated into the low-level semantic features,thus reducing information loss.Finally,the features are upsampled to the same size and concatenated through a feature fusion module to locate the object information in the image,resulting in the final saliency prediction map.We conducted sufficient experiments on 5 benchmark datasets,and the experimental results proved that the model can effectively capture the localization information of significant targets and improve the problem of missing details at the target edges.(2)For solving the problem of salient targets being overwhelmed by the background in complex scenes,This paper propose a new Attention-based Boundary-aware Pyramid Pooling Network(ABAPNet).ABAPNet consists of Feature Aggregation Module,Pyramid Pooling Module,Boundary-Aware Module and Cascaded Dual-Attentive Module.By introducing a combination of channel attention and spatial attention,higher-level information is used to assist shallow-level features,and more important semantic features are obtained from multiple channels to deepen the key semantic information in the feature images.In addition,a hybrid loss function combining binary cross entropy(BCE),structural similarity(SSIM)and joint intersection(IOU)is used to reduce the gradient of the network output features,which not only preserves the important semantic features but also takes into account the network’s attention to the quality of salient target edges,thus guiding the network to better segment the target contours from the complex background.Extensive experiments show that ABAPNet outperforms the advanced competitors.