| According to the 2019 Internet Trends Report,The number of images continues to increase.In the face of massive image data,how to quickly and accurately extract valuable information has become a critical problem to be solved in the field of computer vision.The saliency detection technology simulates the human vision system to segment the object or area of interest from the image,and then provides services for other computer vision tasks.Since 1998,many excellent algorithms based on heuristic rules have been developed in the field of saliency detection.In recent years,fully convolutional networks have pushed saliency detection to a new height.This paper studies saliency area detection of complex scenes and proposes two saliency detection algorithms based on fully convolutional networks.The main research contents are as follows:A saliency detection model based on multi-scale cascading attention mechanism is proposed.Our paper takes fully convolutional network as the basic framework,and at the same time learns from the idea of recurrent neural network,a multi-scale cascade network model is proposed.The network sends the output of the previous time step to the network to continue to participate in training and continuously optimize the output feature map.At the same time,for the fully convolutional neural network to treat each feature channel and feature space position without difference,a Channel-wise Attention Mechanism based on foreground background and a Spatial Attention Mechanism based on saliency prediction feature map are proposed.The multi-scale cascaded network model can generate intermediate saliency prediction feature maps at multiple scales,and use these prediction feature maps to divide shallow features into foreground feature maps and background feature maps,and calculate the channel weights of features based on the foreground and background feature distributions.Using the predicted feature map to calculate the spatial position weight of the feature map makes the network pay more attention to the saliency-related areas and suppress the interference of the background area to the saliency area.The six benchmark data sets in the field of saliency detection are compared with other nine algorithms in qualitative and quantitative aspects,which proves the superiority of the saliency detection model based on the multi-scale cascading attention mechanism.In the early days,traditional saliency detection algorithms relied on heuristic rules to capture context information,while saliency detection algorithms based on deep learning relied on powerful learning capabilities to capture scene semantic information.In order to capture both contextual and semantic information,a fully convolutional neural network based on heuristic rules is proposed.The network structure is divided into multi-scale cascade network flow and heuristic rule prior flow.The heuristic rule priori flow first uses Simple Linear Iterative Clustering algorithm for superpixel segmentation,and then uses the superpixel as the basic unit to calculate the color contrast and color distribution priori saliency feature maps.The priori saliency feature maps are integrated into the multi-scale cascaded network flow by dynamic channel weighting.The multi-scale cascade network flow draws on the ideas of fully convolutional neural networks and recurrent neural networks,and performs supervised learning in the process of gradual upsampling of feature maps while iterating the predicted saliency feature maps from the previous scale to the next scale,and finally the predicted saliency map with the same scale as the input image is obtained.In order to further optimize the salient object detection effect,this paper directly integrates Conditional Random Field into the network stream to realize the integration of detection and optimization.Experimental comparison and analysis with other algorithms proves that the fully convolutional network based on heuristic rules has better detection performance in various scenarios. |
PDF Full Text Request