Impact of facial occlusions in age estimation algorithms for forensic applications

Abstract

[EN] Age estimation is a critical tool for identifying minors and potential offenders in Child Sexual Exploitation Materials (CSEM). However, deep learning methods, which are the current state-of-the-art for age estimation, have low accuracy in predicting the age of minors and older adults due to a lack of data in existing datasets. In addition, facial occlusion, used to hide victims’ identities in some CSEM, can further impact the accuracy of age estimators. This study evaluates the performance of six deep learningbased age estimators on facial images with and without occlusion. To simulate CSEM conditions, we used non-occluded datasets (i.e., FG-Net and APPA-REAL), and also synthetically occluded versions with black masks of these datasets. Results indicate that age estimators are more affected by eye occlusion than by mouth occlusion and facial occlusion affects minors and the elderly more than other age groups. This study could serve as a reference for the estimation of age in forensic applications such as victim profiling on CSEM.