Solving Double-Sided Puzzles: Automated Assembly of Torn-up Banknotes Evidence

Area of Science:

• Computer Vision
• Forensic Science
• Digital Archiving

Background:

• 2D object reconstruction is crucial in forensics, archiving, and banking.
• Existing literature often treats reconstruction as a one-sided puzzle problem.

Purpose of the Study:

• To address torn banknote reconstruction as a novel double-sided puzzle problem.
• To develop a new dataset (ToB) for this specific reconstruction task.
• To propose new metrics for evaluating reconstruction success and loss ratios.

Main Methods:

• A Borda count-based selection approach was employed to identify the optimal keypoint-based method.
• The Accelerated-KAZE (AKAZE) algorithm was selected as the most effective keypoint-based method.
• A new dataset, ToB, was created for training and evaluating the reconstruction system.

Main Results:

• The Accelerated-KAZE (AKAZE) method demonstrated superior performance in the double-sided puzzle reconstruction.
• The proposed reconstruction system achieved an average success rate of 95.55% on the ToB dataset.
• New metrics were successfully developed to quantify the success and loss ratios of reconstructed banknotes.

Conclusions:

• Torn banknote reconstruction can be effectively modeled as a double-sided puzzle problem.
• The AKAZE-based system offers a highly accurate solution for torn banknote reconstruction.
• The developed metrics provide a reliable way to assess the quality of banknote reconstruction.