Research On Lightweight Salient Object Detection With Regional Augmentation

The attention mechanism of the human eye helps humans to quickly locate the most salient objects in a scene,and the salient object detection task was developed from this.The salient object detection task aims to use the computer to mimic the human attention mechanism in order to detect the most attention-grabbing targets in an image.This technology plays an important role in production and life as a fundamental task of computer vision.With the advent of mobile era,image resources are becoming more and more abundant,salient object detection algorithms which are based deep learning have made breakthroughs in performance,but there are still many urgent problems to be solved.First,how to further improve the detection quality of the model,such as alleviating the lack of important features,enhancing the consistency of pixels in saliency regions and improving the sharpness of foreground edges are the key issues to improve the accuracy of model detection.Furthermore,facing the needs of current production and life,how to design a lightweight salient object detection model to achieve a better balance between performance and overhead is also a problem worth exploring.This paper is mainly based on the deep convolutional neural network to achieve lightweight salient object detection with regional augmentation.The specific research contents are as follows:(1)To tackle the problem of missing features and poor regional consistency in existing salient object detection algorithms,a salient object detection network with regional augmentation is proposed.First,the context-aware prediction module is used to extract the multi-scale feature information of the image.Second,the feature hybrid enhancement module is used to effectively integrate the feature information generated by the prediction module.In addition,the multi-loss fusion method is used to supervise the network,which combines the binary cross-entropy loss function,the structured similarity loss function and the proposed regional augmentation loss function.The network with the multi-loss fusion method can maintain the integrity of the foreground region and enhance the regional pixel consistency.Experimental results show that the saliency map generated by the model has clear edges,excellent regional consistency and complete foreground structure.(2)To tackle the problem of excessive storage consumption in salient object detection algorithms,a lightweight salient object detection method based on multiple features aggregation is proposed.The multiple features aggregation network mainly consists of neighborhood stage interactive module and top-to-down aggregation module.First,the neighborhood stage interactive aggregation module aggregates the primitive features extracted by adjacent stages and enhances the original features extracted by each stage.Second,the top-to-down aggregation module aggregates the features of each stage step by step.In this process,the resolution of the feature map is continuously improved.The top-to-down aggregation module outputs a saliency map with the same resolution as the input image in the last stage.Experimental results show that the model has fewer parameters and its performance is better than other algorithms.(3)The lightweight salient object detection algorithm has a faster training speed,which is constructed by using the pre-trained backbone network as the feature extraction module.However,the model constructed based on the backbone network,whose feature are redundant,is not sufficiently lightweight.To tackle the above problems,a lightweight salient object detection method based on multi-scale basic blocks is proposed.In the method,multi-scale basic block is constructed by using lightweight modules such as octave convolution,depthwise separable convolution and atrous convolution.The feature extraction module is formed by stacking multi-scale basic blocks.The lightweight feature fusion module effectively fuses the multi-level features,and outputs the saliency map.Experimental results show that the parameters of the model are only 0.2M which the basic detection performance is maintained.Additionally,performance and overhead can be balanced by tuning hyperparameters.