Peering Through the Past: Forensic Analysis of Impression Via Deep Learning

Abstract

Indented writing is a crucial part of forensic science because it can reveal hidden informa- tion and evidence that is invisible to the eye. This is achieved by identifying impressions of the original writing that are present beneath the surface of the writing page. Yet, traditional methods for deciphering indented writing, are time-consuming and can pose environmental and health risks. Manual and chemical approaches also carry the risk of evidence distortion. In this study, we propose an innovative approach to deciphering indented writing using semantic segmentation and deep learning. Our objective is to provide a more efficient and non-destructive solution for deciphering indented writing to benefit the forensic domain. To achieve this goal, we carefully created a comprehensive dataset. In our study, we employed the U-Net deep learning model to precisely catego- rize each pixel as either belonging to the indented writing region or not. We assessed our model's performance by employing two distinct image channels: grayscale and color (RGB). When trained with grayscale images, the model had a remarkable accuracy of 98.95% and a Mean Intersection over Union (mIoU) of 70.88%. But when color images were used, the model's performance was greatly improved, with accuracy and mloU of 98.86% and 84.14%, respectively. These findings indicate that accurate indented writing recognition by the model is enhanced by RGB imagery.