Research On Weakly Supervised Salient Object Detection Via Pseudo Label Optimization

As an essential task in the computer vision,salient object detection(SOD)plays an important role in high-level semantic understanding,it provides necessary prior knowledges for various image processing and computer vision tasks.As a pixel-level image segmentation task,SOD relies heavily on high-quality,large-scale pixel-level annotations as data support.However,the collection of such kind of annotations is time-consuming and laborious.Therefore,weakly supervised learning based on efficient image-level annotations has received much attention in recent years.Most of the existing works focus on how to generate high-quality pseudo labels,while ignoring the negative impact of its noise information on saliency networks,leading to sub-optimal performance.To this end,this thesis proposes two countermeasures from different perspectives,i.e.,the multiple information optimization and the pseudo label self-calibration optimization,to perform effective optimization on original noisy pseudo labels,and then help to establish high-performance saliency networks.Therefore,the main research works and contributions of this thesis are as follows.(1)This thesis establishes the first saliency-based category dataset,i.e.,DUTS-Cls,on weakly supervised salient object detection.It contains 44 categories and 5959 image samples,making sure that all the category labels accurately match the salient objects.(2)To address the prejudiced problem in single pseudo label,this thesis proposes a WSOD framework based on multiple information integration.First,it generates two groups of pseudo labels with different styles based on two existing image refinement algorithms,and uses them as the supervisions of the proposed framework.Then,a novel multi-filter directive network including two directive filters and a saliency decoder is proposed.During the training process,the directive filters learns the multiple saliency cues from its corresponding supervisions.With the constraints of a self-supervision loss and an image refinement algorithm,the learned multiple saliency cues are then integrated and refined.Finally,the purified and comprehensive saliency cues are propagated to the saliency network in a supervision form.Extensive experiments support that the proposed method outperforms all the state-of-the-arts on multiple test datasets over multiple metrics,and,moreover,shows its good ability of generalization and expandability.(3)To solve negative noise problem in pseudo labels,this thesis proposes a simple but effective pseudo label self-calibrated optimization method,i.e.,the self-calibrated training strategy.Based on the network’s relative robustness to the noise at the early training process,this thesis constructs a mutual calibration loop between pseudo labels and saliency networks.It gradually expands the accurate saliency cues in the pseudo labels,and meanwhile,eliminates the negative influence of the noisy information on the saliency network.An image refinement algorithm based on pixel affinity is adopted to add detailed information.Comprehensive experiments on multiple test datasets show that our method outperforms all the existing methods.Moreover,our proposed self-calibrated strategy can also be extended to other methods and improve their performance without increasing their model complexity.