Abstract
The identification of perpetrators, present in Child Sexual Abuse Imagery (CSAI), is a significant challenge due to the use of anonymisation techniques that mask their identities. Consequently, researchers have investigated the use of uncommon biometric identifiers such as knuckle patterns, palmprints and the dorsal side of the hand. This research proposes a Convolutional Neural Network (CNN) based, fully automated approach to biometric identification using dorsal hand images. The identification performance of three different CNN architectures, AlexNet, ResNet50 and ResNet152, is experimentally determined against two similar datasets, the 11k Hands and IITD dorsal hand databases. A transfer learning approach is used and the final output layers of the CNNs are modified to match the number of classes present in the datasets. The results showed that ResNet CNNs achieved identification accuracies greater than 99.9% on both datasets, whereas the AlexNet CNN achieved between 80.1% and 93.7%. These results demonstrate that it is feasible to use deep, off-the-shelf CNNs, such as ResNets, for automated biometric identification using dorsal hand images. This highlights the potential of using dorsal hand images to identify perpetrators of child sexual abuse from CSAI.
