Endmember Analysis of Overlapping Handwritten Text in Hyperspectral Document Images

Abstract

Hyperspectral Imaging (HSI) uses large portions of the electromagnetic spectrum to obtain information from images that would be very difficult to get otherwise. An important task in forensic analysis of documents is to extract signatures for authentication. Signatures in documents often overlap with other parts of a document such as typed text or stamps and hence it is difficult to retrieve them. In this work we present a novel algorithm for recovering signatures from hyperspectral images of documents where signatures overlap typed text, seals, stamps, or other images. We used pure pixel index approaches to analyze the hyper-pixels corresponding to regions where overlaps occur and spectral classification methods to analyze the extracted channels’ information and to separate the overlapping signatures. We used the structural similarity index to validate the extracted signatures. Our experimental results highlight the capability of the proposed algorithm to recover signatures with great precision in HSI document images.

Keywords

• Hyperspectral imaging
• endmembers
• spectrum
• abundance estimation
• spectral analysis