Deteksi dan Rekognisi Wajah Forensik pada Citra Kualitas Rendah Menggunakan Hybrid PCA-LBP dan SVM

Abstract

Face recognition and detection is a biometric technique widely used in digital forensics to identify perpetrators or victims from CCTV or mobile device image evidence, yet real-world evidence images are often low-quality due to blur, noise, small resolution, and poor lighting. The novelty of this study lies in a hybrid feature-extraction scheme combining Principal Component Analysis (PCA/Eigenface) and Local Binary Pattern (LBP), fused through a weighted sum rule whose weights are optimized per degradation type, coupled with three non-learning image enhancement techniques (Non-Local Means denoising, unsharp masking, and CLAHE), together with direct empirical evidence that naive feature fusion without weighting can degrade system performance. This study builds and directly tests the pipeline using 100 images from the official Labeled Faces in the Wild (LFW) subset bundled with the scikit-image library (50 face images and 50 non-face images, 25×25 pixels), evaluated using accuracy, precision, recall, F1-score, and ROC-AUC for the face-vs-non-face classification task, and rank-1 accuracy for the identification task on a gallery of 50 identities. The best results show a hybrid PCA+LBP-SVM classification accuracy of 95.0–97.5% depending on image quality conditions, while for the rank-1 identification task, the PCA feature excels under blur degradation (96%) but drops sharply under dark lighting (2%), the LBP feature shows the opposite pattern, and the weighted score fusion scheme succeeds in matching or exceeding the best single method under every condition. It should be noted that image enhancement does not always improve identification performance; under certain degradation conditions, enhancement actually sharply reduces rank-1 identification accuracy, indicating that the relationship between visual quality and recognition performance is not always linear. This study provides fully reproducible empirical evidence, based on a public dataset, on the contribution of each stage of the digital forensic pipeline to face recognition performance on low-quality image evidence.