Enhancing Face Recognition Performance: A Comparative Evaluation of ArcFace and FaceNet using Image Preprocessing and ROC–EER Metrics

الملخص

Deep learning has pushed face recognition to the point where standard benchmarks are no longer much of a challenge accuracy at near-human levels has become almost routine. What has received far less attention is a quieter but consequential question: how does image preprocessing affect the verification performance of models that have already been trained? This paper takes that question seriously, presenting a systematic comparison of two leading face recognition models FaceNet and ArcFace as they are subjected to a progressive preprocessing framework applied to the Labeled Faces in the Wild (LFW) dataset.The evaluation used 1,000 face pairs, split evenly between 500 matched and 500 mismatched, and ran each pair through five preprocessing stages. Stage 0 used raw images with no modification. By Stage 4, a full pipeline was in place, incorporating face detection, CLAHE histogram equalization, Gaussian blur, and pixel normalization. Performance at each stage was measured using the ROC curve, the Equal Error Rate (EER), classification accuracy, and error rate metrics chosen because they do not depend on selecting a single decision threshold and are well suited to binary verification tasks.At baseline, ArcFace was the stronger model across every metric, reaching 97.30% accuracy, an AUC of 0.9837, and an EER of 0.0340, against FaceNet's 95.60%, 0.9812, and 0.0480 respectively. What happened next ran counter to what most practitioners would expect. Rather than improving as preprocessing became more involved, both models declined consistently, across every stage, with no preprocessing step producing a measurable gain over the raw baseline. ArcFace felt the impact more sharply, losing 6.40% in accuracy from Stage 0 to Stage 4, while FaceNet dropped 3.20% over the same range. From Stage 2 onward, the ranking between the two models flipped: FaceNet posted lower EER values than ArcFace at every subsequent stage.The picture that emerges from these results is fairly clear. ArcFace is the better choice when imaging conditions can be controlled and images arrive at the model largely intact. FaceNet is more resilient when the pipeline involves external processing that the model has no control over. More broadly, the findings push back against the common assumption that image enhancement is a safe or neutral intervention applied without validation, it can quietly degrade the very systems it was meant to help