Forensic Document Analysis Using Hyperspectral Imaging and Deep Convolutional Spectral Clustering

An unsupervised classification that reveals the hidden information from documents under investigation is proposed for hyperspectral imaging using deep spectral clustering. Usually, hyperspectral document analysis requires a domain expert to analyze the data manually to trace any evidence, which is a time-consuming and cumbersome procedure due to the high dimensionality of the hyperspectral data. In this work, we proposed a deep spectral clustering method that simultaneously learns the constraints of an autoencoder and the clustering parameters. The encoder part of the autoencoder maps the input spectral data into a lower-dimensional space, which is used to learn the clustering parameters iteratively. The algorithm tested with known data for clustering performance evaluation, and the proposed method outperformed the standard methods by a significant margin on important clustering benchmarking indexes.