A Transfer Learning Approach for the Tattoo Detection Problem

Abstract

Tattoos are still poorly explored as a biometrics factor for human identification, especially in public security, where tattoos can play an important role for identifying criminals and victims. Tattoos are considered a soft biometrics, since they are not permanent and can change along time, differently from hard biometrics traits (fingerprint, iris, DNA, etc). The identification of tattoos are not simple, since they do not have a definite pattern or location. This fact increases the complexity of developing models to address this problem. In addition, the tattoo identification roadmap is very complex, including several steps and, in each step, specific methods need to be developed. Among the several problems identified in this roadmap, we tacked the identification problem, which is defined as: given an image of a person, determine if there is a tattoo or not. We present a deep learning model based on transfer learning for the tattoo detection problem. We also used data augmentation to improve the diversity of the training sets so as to achieve better classification accuracy. Along the work two new datasets for tattoo detection were created. Several comparative experiments were done to evaluate the diversity of images in the datasets, and the accuracy of the proposed model. Results were very promising, achieving an accuracy of 95.1% in the test set, and a F1-score of 0.79 in an external dataset. Overall, results were satisfactory, given the complexity of the problem. Future work will focus on expanding the datasets created and addressing the other problems of the tattoo roadmap.

AutoML e transferência de aprendizado por reforço entre problemas de otimização combinatória

“…A transferência de aprendizado ou Transfer Learning, refere-se a transferência de conhecimentos entre problemas semelhantes. Esta abordagem torna possível que ocorra uma reduc ¸ão do tempo gasto durante o treinamento do problema objetivo, uma melhor aprendizagem e consequentemente, melhores resultados finais [46], [47].…”

Iban plaited mat motif classification with adaptive smoothing

“…The image was sourced from [4] To date, a significant amount of research had been done on the robust detection, description and matching of invariant features related to motif and pattern classification. Features extraction algorithm and classification methods were applied to batik motif and batik making using convolutional neural network (CNN) model architectures [6]- [10], using multiwindow and multiscale extended center symmetric local binary patterns (MU2ECS-LBP) [11] and [12], using scale invariant feature transform (SIFT) [13], [14] using gray level co-occurrence matrices (GLCM) [15], [16], fine arts [17]- [19], for license plate recognition [20], [21], and tattoo recognition [22]- [24]. Even though the reported performance was quite high but these methods still suffer from false positives due to similar features that contain more than one pattern and noisy background.…”