.png){kind=icon}
.png){kind=icon}
Return
Deep Learning-Based Source Printer Identification from Microscopic Ink Dots Using Morphology-Based Augmentation
.png){kind=icon}
PRE
AI
DOI：10.1007/978-3-032-09321-9_31.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Precise tracking of printed documents is essential in forensic investigations; however, acquiring a sufficient volume of real microscopic ink dot images for training deep learning models is both resource-intensive and costly. To address this challenge, this study introduces a morphology-based data augmentation strategy that simulates realistic variations, including rotations, translations, scaling, and local deformations, to enhance dataset diversity while preserving the structural integrity of ink dot morphology. This approach not only alleviates data scarcity but also improves model efficiency and generalization. A comprehensive evaluation is conducted on five state-of-the-art deep learning architectures-ConvNeXt-Tiny, CoAtNet-0, ResNet-50, TinyViT-21M, and MobileViT-XS-assessing both classification performance and deployment feasibility across diverse hardware platforms, ranging from high-performance desktops to resource-constrained embedded devices. Experimental results demonstrate that morphology-based augmentation substantially enhances model generalization, with MobileViT-XS achieving the most significant improvement (2.71%) and ConvNeXt-Tiny attaining the highest validation accuracy (97.29%). These findings highlight the efficacy of modern CNN-Transformer hybrid architectures for scalable, cost-effective, and high-precision source printer identification, reinforcing their applicability in real-world forensic investigations.
Keywords:
Source Printer Identification
Morphology-based Augmentation
Deep Learning
CNN-Transformer hybrid
Journal
C
IF:
0
Papers:
34
Citations:
0