Effect of Validation-Guided Weighted-Sum PCA-LDA Feature Fusion on ABC-Optimized CNN Performance for Multimodal Biometric Identification

Authors

  • AKANDE, Olubunmi Victoria FEDERAL POLYTECHNIC AYEDE OYO STATE
  • Afolabi A. O. Department of Cyber Security, Ladoke Akintola University of Technology, Ogbomoso, Nigeria
  • Adeosun O. O., Department of Computer Science, Faculty of Computing and Informatics, Ladoke Akintola University of Technology, Ogbomoso, Nigeria
  • Salahudeen S. A. Federal Polytechnic Ayede, Oyo State
  • Oyewale S. O Federal Polytechnic Ayede, Oyo State
  • Araba T. A. Federal Polytechnic Ayede, Oyo State

Keywords:

ABC Optimization, Convolutional Neural Network, Face-Palm-Iris Recognition, Feature-Level Fusion, Linear Discriminant Analysis, Multimodal Biometrics, Principal Component Analysis, Weighted-Sum Fusion

Abstract

Feature-level fusion can combine complementary biometric information without increasing classifier complexity, but the benefit of a weighted fusion rule depends on how its component representations are normalized, weighted, and evaluated. This study investigated validation-guided weighted-sum Principal Component Analysis (PCA)-Linear Discriminant Analysis (LDA) fusion for face-palm-iris multimodal biometric identification. The MULB dataset provided 174 matched identities, each with 20 face, 20 right-hand/palm and 20 right-iris images (10,440 images in total). Source images were partitioned before constructing identity-consistent triplets, yielding 2,227 training, 557 validation, and 696 independent test triplets. PCA, LDA, and separate standardization transforms were fitted on training data only; both feature representations were restricted to 150 dimensions. Nine PCA:LDA weight pairs from 0.1:0.9 to 0.9:0.1 were evaluated on validation data. The selected 0.1:0.9 weighting achieved 96.04 ± 0.62% validation accuracy and 95.91 ± 0.63% macro F1, compared with 94.33 ± 0.77% and 94.13 ± 0.80% for equal weighting. With the final ABC-optimized CNN frozen, ten matched-seed runs produced independent-test accuracies of 97.19 ± 0.39% for PCA, 97.78 ± 0.29% for LDA, and 97.66 ± 0.23% for the selected fusion; corresponding macro F1 scores were 97.12 ± 0.42%, 97.71 ± 0.30%, and 97.61 ± 0.23%. Friedman tests found no significant overall difference in test accuracy (χ²=5.389, p=0.0676) or macro F1 (χ²=3.800, p=0.1496). The selected weighted representation therefore improved validation performance over equal weighting and PCA, while final testing showed competitive, low-variance performance rather than statistically significant superiority over LDA.

 DOI: https://doi.org/10.5281/zenodo.23415861

 

Downloads

Published

2026-10-16