A Comparative Study of Transfer Learning Models for Tattoo Detection

This study explores the use of transfer learning (TL) with computational neural networks (CNNs) to address the challenge of forensic tattoo identification. Tattoos are distinctive markers that play a crucial role in personal identification in forensic investigations. Traditional manual methods for tattoo identification are often time consuming and prone to errors, underscoring the need for automated and reliable alternatives. Pre-trained CNN models, using TL, have shown potential in similar image classification tasks. To evaluate their applicability in tattoo detection, eight pre-trained models were tested: MobileNetV2, Xception, NASNetMobile, InceptionV3, VGG16, DenseNet121, ResNet50 and EfficientNetB0 on a balanced dataset of 1,000 tattoo images, enhanced with data augmentation techniques. The models were trained using a fixed 70%-15%-15% split for training, validation, and testing. The data set was randomly shuffled before splitting to minimize bias in the training process. Although cross-validation is a common approach in machine learning, we opted for a fixed split to maintain consistency in model evaluation. The results indicate that MobileNetV2 achieved the highest accuracy at 99.9%, followed by DenseNet121 (99.7%) and Xception (99.4%). NASNetMobile also performed well, with an accuracy of 99%. InceptionV3 and VGG16 demonstrated moderate precision levels (97.5% and 95.4%, respectively), while ResNet50 and EfficientNetB0 achieved lower precision levels of 83.1% and 70.3%, respectively. Based on these results, MobileNetV2, DenseNet121, and Xception emerged as the most effective models were evaluated based on precision, precision, recall, F1 score and computational efficiency. This evaluation provides a comparative analysis of TL-based CNN models, offering insights into their performance and resource requirements for forensic tattoo identification.