CLASSIFICATION OF DIGITAL ALTERATIONS IN BIOMETRIC IMAGES USING A TWO-BRANCH CONVOLUTIONAL NEURAL NETWORK

Abstract

This study investigates multiclass classification of digitally altered fingerprint images from SOCOFing using convolutional neural networks. The objective is to compare a baseline CNN with a two-branch model that combines image features and local descriptors of fingerprint ridge structure. The structural branch uses Sobel gradients and a local structure tensor. Its contribution to the fused representation is controlled by a trainable spatial gate conditioned on both feature streams and a coherence map. Global average and maximum pooling replace spatial flattening in the classification head. Four classes are considered: Real, Obliteration, Central_rotation, and Z_cut. Before splitting the data, source-image families associated with conflicting labels or exact duplicates across subjects were excluded, leaving 54,978 images. Training, validation, and test subsets were formed by subject so that derivatives of the same fingerprint remained in a single subset. Both models used 96 × 96 grayscale inputs, identical optimization settings, random seed 44, and a maximum budget of 100 epochs. Early stopping was triggered after seven epochs without an improvement in validation macro-F1. On the shared test subset of 8,172 images, the baseline achieved 94.20% classification accuracy and 92.74% macro-F1, while the two-branch model achieved 95.06% and 93.67%, respectively. The number of incorrect predictions decreased from 474 to 404, and the parameter count decreased from 4,738,052 to 43,174. Class-specific F1 improved for all four classes, although precision and recall did not improve uniformly. Confusion matrices and learning curves are reported to show both gains and error trade-offs. Training stopped after 45 epochs for the baseline and 29 epochs for the modified model; checkpoints were selected at epochs 38 and 22 using validation macro-F1. The results describe one run per model on a fixed split and do not establish statistical significance or robustness across random seeds. Because the feature streams, fusion mechanism, and pooling head were changed together, their individual contributions cannot be determined without ablation experiments. The internal SOCOFing test split had been evaluated previously and does not constitute a newly reserved external evaluation set. The task concerns digital alteration categories in SOCOFing rather than physical presentation-attack detection or fingerprint identity recognition. The findings support further evaluation of compact structural feature fusion under repeated runs and independent test conditions.