Journal
EXPERT SYSTEMS WITH APPLICATIONS
Volume 233, Issue -, Pages -Publisher
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.eswa.2023.120960
Keywords
Referring personal image privacy protection; Pixel imbalance problem; Balanced BCE loss; Referring image segmentation; Overprotection problem
Ask authors/readers for more resources
Traditional personal image privacy protection often suffers from overprotection, resulting in unnecessary information loss. This paper introduces a new task called "Referring Personal Image Privacy Protection" (RP-IPP) which aims to protect a designated person in an image based on user's text or voice input. A lightweight yet effective personal protection network, Balanced Referring Personal PrivacyNet (BRPPNet), is proposed, which includes a Multi-scale Feature Fusion Module (MFFM) and a Balanced-BCE loss to accurately localize the referred person. Experimental results demonstrate the superiority of BRPPNet over existing approaches for RP-IPP.
Traditional personal image privacy protection usually suffers from the overprotection problem, where one or more undesired persons in an image may be inevitably shielded, yielding unnecessary information loss. Motivated by this, this paper explores a novel task ''Referring Personal Image Privacy Protection'' (RP-IPP) to protect the designated person in an image according to a user's text or voice input. We propose a lightweight yet effective personal protection network ''Balanced Referring Personal PrivacyNet'' (BRPPNet), which introduces a Multi-scale Feature Fusion Module (MFFM) with a proper ''Balanced-BCE loss'' to effectively localize the referring person. Technically, MFFM adopts a lightweight CNN backbone to filter noise information as well as complement visual features for high-quality mask generation. What is more, we have theoretically proven the insufficiency of binary cross-entropy (BCE) loss and its variants for RP-IPP, which suffers from the serious imbalance problem during gradient propagation, and thus formulate ''Balanced-BCE loss'' to alleviate the gradient propagation bias caused by unequal positive and negative pixels. To verify the effectiveness of BRPPNet, we manually construct a dataset ''Referring Personal COCO'' (RPCOCO). The experimental results demonstrate that BRPPNet outperforms the advanced approaches for RP-IPP, and the proposed ''Balanced-BCE loss'' embedded into several existing approaches consistently boosts performance, yielding remarkable improvements on all the metrics.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available