Algoritmo Para Reconhecimento De Imagem Via PCA: Análise de Desempenho

Abstract

This work presents the implementation and performance evaluation of a facial image recognition algorithm based on the Principal Component Analysis (PCA) technique — known as EigenFaces — combined with the Support Vector Machine (SVM) classifier. The Cropped Yale Face Database B, composed of grayscale images under different lighting conditions, was used. The methodology was divided into steps: preprocessing and standardization of images, feature extraction with PCA, classification using SVM (with RBF kernel and one-vs-one strategy for multiclass classification), and model validation through stratified cross-validation. Tests defined the optimal number of principal components (75) and classifier parameters. The results demonstrated high performance: average validation accuracy of 95.13%, low standard deviation (0.0094) among folds, and high F1-Scores for most classes, even under adverse lighting. The confusion matrix showed occasional errors, without systematic patterns. It was concluded that combining EigenFaces and SVM is effective and robust for facial recognition in controlled environments, even with lighting variations. For future work, it is recommended to apply the model to more complex datasets and compare it to other techniques, such as Convolutional Neural Networks.