| With the continuous upgrade of high-performance computing equipment and the era of big data,the field of artificial intelligence has made great progress.Deep learning technology has become a research hot-spot and has been widely used in the field of computer vision.In many sub-domains of computers version,such as image classification,image detection,and image segmentation,Deep neural networks(DNNs)achieves state-of-the-art performance.Convolutional neural networks(CNNs)are undoubtedly the most compelling type of deep neural network structure.Most of the current image-related tasks apply this structure.The convolutional neural network structure is a typical data-driven.The algorithm needs to use large-scale data into training,and data labeling needs to consume a lot of expensive costs,which undoubtedly limits the generalization of the algorithm in other fields.Salient object detection method can effectively reduce the cost in the process of data annotation.In the study of related issues,we propose a new Salient Object Detection structure.Although the existing structure based on CNN has achieved high detection accuracy,but most of them can not effectively use multi-scale features and global semantic information.We can further tap the potential of the algorithm and propose a new Salient object detection structure,Gating for Double Pyramid Network(GDPNets).It consists of two pyramid structures:Feature Pyramid Network(FPN)and Pyramid Pooling Module(PPM).FPN has the capability of capturing the inherent muli-scale and pyramid hierarchy,while PPM can exploit the global context information by different-region-based context aggregation.It is known that the irrelevant information corresponding to non-salient objects or background may deteriorate the performance of the model.We introduce two gating strategies,i.e.,Cross-Gating for FPN and Single-Gating for PPM,to suppress the incurred irrelevant information in hidden features.Our proposed approach achieves state-of-the-art performance on five benchmark datasets,which demonstrates the effectiveness and robustfeature extraction capability of the proposed GDPNet.The redundancy of structure parameters and huge computation costs seriously restrict the application of the method.This paper analyzes the redundancy of parameters of network structure,studies on the current general network lightweight technology,and designs a networks light-weight strategies which is combined methods.According to the network structure we proposed,pruning the network trained.On the premise of not damaging the network performance,we reduce the number of network parameters significantly,and the binarization or temization parameter quantification is performed on based.The advantage of this method is that it can reduce the network size and speed up the network operation without changing the network structure.The application value of the method is greatly enhanced. |
