Research On Semi-Supervised Salient Object Detection Based On Deep Learning

Recently,salient object detection algorithms based on deep learning have been widely proposed and have achieved significant performance compared to traditional computer vision algorithms.However,these algorithms also rely on a large amount of pixel-level fine manual annotations which are time-consuming and costly.In order to reduce the dependence on fine labeled data,current researchers have proposed a series of salient object detection algorithms based on unsupervised learning and weakly supervised learning.But there is still a big gap in performance between these methods and the fully supervised learning methods.In the real world,in addition to the vast majority of weak labeled data,there is also some pixel-level labeled data that can be used.Based on this observation,semi-supervised learning has become a better choice to solve the problem of labeling costs in salient object detection.The simplest semi-supervised learning method is to directly merge the strong and weak labeled data sets,and then train the salient object detection network in an end-to-end manner.However,such a method cannot fully take advantage of the weak labeled data,leading to pool performance.Even it is worse than the performance obtained by using a small number of strong labeled data.In this paper,an auxiliary model is first proposed to learn the conversion process between scribble annotations and ground truth maps for obtaining high-quality pseudo labels,which are used as weak labeled data in the next training stage.In order to make better use of weak labeled data,we propose a method that uses a dual-branch network to process strong labeled data and weak labeled data separately,and imposes fully and weakly supervised learning on two branches.In addition,an asymmetric guide fusion module is designed for the network,so that the weak branch pays more attention to the position and contour information of the object and the strong branch focuses on detailed information of the object.When the gradient flows back in the dual-branch network,a shared module obtains the joint features of the strong and weak labeled data,which is used to eliminate the negative impact caused by two different training data.With the designed strong and weak dual-branch network and asymmetric fusion module,great performance improvement is obtained while only a small number of pixel-level ground truth annotations are used for training.The method in this paper has been tested on six publicly available salient object detection datasets.And quantitative and qualitative results compared to other unsupervised,weaklysupervised methods and some state-of-the-art fully-supervised methods have been given.Experiments show that the proposed method is far superior to other methods based on unsupervised learning and weakly supervised learning in terms of performance and visual effects.And while using only 20% of the labeling cost,it can reach around 95% of performance compared to the fully supervised learning methods.Thus,the dependence on pixel-level annotations is effectively reduced.