| As a current research hotspot in the field of computer vision,salient object detection aims to accurately locate and segment the most eye-catching regions in images by simulating the human perceptual system,and is now widely used in various image areas of computer vision tasks,such as image classification,object recognition,image retrieval,scenario analysis,etc.In recent years,with the rapid advances in computer technology and artificial intelligence,traditional methods based on manual feature extraction have gradually lost their advantages,while salient object detection algorithms based on deep learning are more competitive in terms of arithmetic power and performance.When faced with complex scenes,it can still be improved the generalization ability of the model through its own learning and training.Therefore,the contents of this paper revolve around the study of salient object detection algorithms based on deep learning and related applications,with the main work as follows:Under interference such as shallow noise or complex backgrounds,most existing salient object detection methods suffer from blurred information on the edges of salient objects and low localization accuracy.To address the above problems,we proposes a fused hybrid self-attentive picked feature aggregation saliency detection network,which is introduced with a mask convolution module to capture the spatial information in the feature map and improve the grasp of relative location information.To refine the boundary information of salient objects,we adopt a selected feature aggregation module in this paper,which utilizes the mutual guidance between deep semantic features and shallow features to adaptively fuse input features containing complementary components between adjacent convolution layers,thus to suppress redundant information such as background noise.Finally,the channel attention mechanism is incorporated in parallel on the basis of the self-attention mechanism to further obtain the importance of inter-level features in the channel and the potential correlation among samples to accurately predict the location of salient regions.The experimental results show that the model has better detection performance compared with 15 mainstream methods on all five standard datasets.In particular,the F-measure reaches 0.915 in the ECSSD dataset and the MAE decreases to0.043.When faced with challenging scenes such as small objects and low contrast between foreground and object,the detection of existing SOD algorithms still has some problems.In this regard,we present a multi-scale fusion of skip-layer features interactive saliency detection network.To accurately identify and segment multiple salient objects at different scales in images,we add a multi-scale feature module based on the Res Net-50 architecture,which gradually achieves complementary fusion by extracting features individually through four independent branches and accurately captures the location information of important regions using an improved attention mechanism.Secondly,a skip-layer feature interaction module is added in the decoding part,and the deepest layer of the global information of the network is used as a guide to effectively fuse the feature information obtained by up-sampling from the decoding part layer by layer and the features extracted from the shallower layers to gain more accurate and explicit feature information.Finally we use a binary cross-entropy loss function to dynamically supervise the training and optimizing of the network.From the prediction results can intuitively reflect that this paper has a greater improvement for the detection of complex scenes such as multiple objects and small objects,and the edges of salient objects are clearer and more accurately located.From the quantitative evaluation analysis,the model can achieve better results,mainly on the DUT-OMRON dataset,which is more difficult to detect,where MAE decreases from0.062 to 0.054 and S-measure improves from 0.804 to 0.815. |
PDF Full Text Request