Chronological differentiation of printed or handwritten text and stamps based on hyperspectral imaging and...

To address the technical challenges associated with determining the chronological order of overlapping stamps and textual content in forensic document examination, this study proposes a novel non-destructive method that integrates hyperspectral imaging (HSI) with convolutional neural networks (CNNs). A multi-type cross-sequence dataset was constructed, comprising 60 samples of handwriting-stamp sequences and 20 samples of printed text-stamp sequences, all subjected to six months of natural aging. Spectral responses were collected across the 400-1000 nm range in the overlapping regions. To suppress spectral noise, multiplicative scatter correction (MSC) was employed as a preprocessing step. The proposed dual-layer CNN architecture consists of an initial convolutional layer with 64 3 × 3 kernels followed by 2 × 2 max pooling, a second convolutional layer with 128 3 × 3 kernels and another 2 × 2 pooling layer, followed by a fully connected layer with 256 neurons that integrates spatial-spectral features, culminating in a four-class classification using Softmax. The model was trained over 150 epochs using the Adam optimizer (learning rate = 0.001) and L2 regularization (λ = 0.001). The approach accurately distinguished between the chronological order of laser-printed toner, gel pen ink, and traditional/photo-sensitive stamp inks. Experimental results demonstrate a classification accuracy of 97.62% (AUC = 0.9965) on the printed text dataset and 96.67% (AUC = 0.9921) on the handwriting dataset, outperforming both extreme learning machine (ELM) (90.42%) and long short-term memory (LSTM) (96.43%) baselines. All pure (non-overlapping) samples were correctly classified with 100% accuracy. Feature analysis confirms the CNN's ability to extract highly discriminative spatial features, effectively overcoming the subjectivity and material-damaging limitations of traditional microscopic techniques.