Acquiring Original Data from Shredded Documents Using Neural Network

Abstract

Shredding is a common way to eliminate documents containing sensitive and confidential data for legal and ethical purposes in different areas such as military applications, scientific inventions, businesses and government organizational applications. Hence acquiring original data from shredded documents is an important part of investigative sciences. Performing recovering process of the shredded documents manually can result in frequent human errors. Therefore, developing a computer-based recovery algorithm is of vital importance in obtaining original data efficiently and accurately. This study investigated an approach of using Neural Network (NN) in reconstructing shredded documents, which takes a scanned image consisting of document strips and outputs the fully or partially recovered original document as another image. The model was developed on single-sided, black and white typed text documents having maximum up to 20 vertical strips from each page. The process consisted of two main steps, namely pre-processing and reconstruction. In pre-processing, paper strips were scanned into an image, and then several image processing techniques were performed to gain shreds into different images without backgrounds. In Reconstruction, a correlation vector was obtained among each shred pairs by comparing the average neighbourhood intensities for each pixel along vertical edges of shreds. This correlation vector was prepared according to the data format suitable for NN. The proposed feed forward back propagation NN model was trained and tested to automatically find the best matching adjacent shred by feeding the correlation vectors as the inputs. This model increases the ability of addressing the vagueness of the algorithm. The algorithm was validated by reconstructing a considerable number of single documents and was proved to be effective with an average resembling frequency of ~78%. The main advantages of the proposed method over the existing approaches are the simplicity, fastness, less human intervention and increased predictive and intelligent characteristics.