[Paper Review] A Simplified Un-Supervised Learning Based Approach for Ink Mismatch Detection in Handwritten Hyper-Spectral Document Images

Muhammad Farhan Humayun, Hassan Waseem Malik+1 more|arXiv (Cornell University)|Jun 11, 2022

Currency Recognition and Detection4 citations

TL;DR

This paper proposes a simplified unsupervised learning approach for detecting the number of visually similar inks in handwritten hyper-spectral document images (HSDIs) using K-means clustering on spectral signatures extracted from text pixels. The method achieves comparable results to state-of-the-art techniques without requiring labeled data or prior knowledge of ink types, identifying five major ink clusters in the iVision HHID dataset with minimal preprocessing and no supervised training.

ABSTRACT

Hyper-spectral imaging has become the latest trend in the field of optical imaging systems. Among various other applications, hyper-spectral imaging has been widely used for analysis of printed and handwritten documents. This paper proposes an efficient technique for estimating the number of different but visibly similar inks present in a Hyper spectral Document Image. Our approach is based on un-supervised learning and does not require any prior knowledge of the dataset. The algorithm was tested on the iVision HHID dataset and has achieved comparable results with the state of the algorithms present in the literature. This work can prove to be effective when employed during the early stages of forgery detection in Hyper-spectral Document Images.

Motivation & Objective

• To develop an automated, unsupervised method for estimating the number of distinct inks in handwritten hyper-spectral document images.
• To eliminate the need for labeled training data or prior knowledge of ink types in forensic document analysis.
• To enable early-stage forgery detection by identifying spectral inconsistencies in ink composition.
• To evaluate the effectiveness of K-means clustering on spectral signatures extracted from HSDI pixels.
• To compare the proposed method’s performance with existing supervised approaches in terms of accuracy and practicality.

Proposed method

• Pre-processing hyper-spectral cubes via cropping to isolate text regions and applying binary thresholding to extract foreground text pixels.
• Generating spectral signatures by computing mean intensity values across 149 spectral bands for each text line.
• Applying Boolean masking to isolate text pixels and suppress background for accurate spectral response analysis.
• Using K-means clustering on the masked spectral data with K=7 to identify distinct ink clusters, including one for background and one for flat printed lines.
• Generating color-segmented RGB images from K-means cluster outputs to visualize ink groupings.
• Selecting Python with Spectral, NumPy, OpenCV, and Matplotlib libraries for implementation due to strong support for spectral image processing.

Experimental results

Research questions

• RQ1 Can an unsupervised learning approach accurately estimate the number of visually similar inks in handwritten HSDIs without labeled data?
• RQ2 How effective is K-means clustering in identifying distinct spectral clusters corresponding to different ink types in hyper-spectral document images?
• RQ3 To what extent does the proposed method outperform or match existing supervised approaches in ink mismatch detection?
• RQ4 What are the limitations introduced by non-text features such as flat printed lines in the clustering output?
• RQ5 Can spectral response analysis alone provide a reliable basis for ink cluster estimation in forensic document analysis?

Key findings

• The proposed method successfully identified five major ink clusters in the iVision HHID dataset using unsupervised K-means clustering on spectral signatures.
• K-means clustering converged after 84 iterations with K=7, producing seven spectral classes, including one for background and one for flat printed lines.
• Visual analysis of spectral response curves confirmed the presence of five distinct ink types, matching the K-means output clusters.
• The method achieved comparable performance to state-of-the-art supervised algorithms despite not requiring any labeled training data.
• The algorithm detected five major ink clusters, though the ground truth contains six ink types, indicating a minor underestimation due to spectral similarity.
• The inclusion of flat printed lines in the data introduced an additional, spurious cluster, which was later identified as a source of noise in the clustering output.