Design And Implementation On Lightweight Network For Salient Object Detection In Complex Scenes

Salient object detection is an essential task in computer vision,which focuses on simulating the human visual perception system to detect the most salient objects or regions in an image.Thanks to the rapid development of deep learning,the currently proposed methods have achieved near-perfect accuracy for simple scenes(e.g.,simple target structure,single target,and high target-to-background contrast).However,accurate detection of complete salient targets in complex scenes(e.g.,multiple targets of different scales,low foreground-to-background contrast,and complex background)remains a challenge.In addition,current research focuses on detection accuracy while ignoring memory burden and computational cost,resulting in too costly in practical applications.Therefore,how to balance the performance and efficiency of the model is also a very worthy research problem,and this paper conducts a lightweight research of salient object detection model for this purpose as follows:1)A saliency object detection network with progressive aggregation of global information is proposed to address the problems of insufficient global information and the influence of feature differences in existing saliency object detection methods.In this network,firstly,a global pooling aggregation module is constructed to help the network obtain richer global contextual information.Secondly,a feature aggregation enhancement module is constructed to mitigate the feature differences of each layer in the feature fusion phase and refine its fused feature map.In addition,by combing the binary cross-entropy loss function,the intersection-over-Union loss function,and the progressive self-guided loss function into one,the network model is guided to learn from three levels of pixels,overall structure,and spatial correlation to obtain a more complete and significant object.Experimental results show that the model can obtain more accurate prediction result maps in complex scenes and outperforms current state-of-the-art saliency target detection methods in terms of performance.2)A lightweight saliency object detection network based on deep feature aggregation and boundary optimization is proposed to address the problem that existing models cannot balance performance and efficiency.The network designs the decoding stage into three parts:semantic branch,spatial branch,and boundary branch.In the semantic branch,a Residual Scale-aware Enhanced Module and Neighbor Dense Decoder are constructed to deepen the network’s ability to extract multi-scale feature information and efficiently aggregate multilayer advanced semantic features to enhance the expression of advanced features.In the spatial branch,spatial information is effectively aggregated with semantic information using the feature aggregation enhancement module proposed in the previous network.In the boundary branch,the shallow features are first refined through spatial attention,and then edge contour supervision is performed to obtain boundary information after fusion with the semantic branch.Finally,an edge refinement module is constructed to fuse the spatial branch and boundary information to obtain the final prediction result.Experimental results show that the model reduces the number of parameters by 92% compared to the first model,and can maintain comparable or even better performance than current state-of-the-art saliency object detection algorithms.3)The network proposed in this paper is applied to camouflage object detection with specific complex scenarios.The test results show that the model in this paper has good applicability and a graphical interface system for camouflage object detection is designed based on the model in this paper,which can better detect camouflage objects and show the application value of the model in this paper.