Scholarlink

<p>Facial recognition is essential for public and private security, widely used to identify individuals and manage access with precision. This study evaluates the impact of image compression on the accuracy of facial recognition models, considering three types of codecs: HEVC, H.264, and JPEG2000, which exhibit distinct compression characteristics. Three facial recognition models were analyzed: Local Binary Pattern (LBP), Extreme Learning Machine (ELM), and the proposed YOLO-M model. The YOLO-M, a hybrid architecture developed in this study, combines YOLOv8 for rapid face detection with ResNet-50, which extracts detailed and robust features. The extracted information is integrated and processed through convolutional and dense layers, culminating in a softmax classifier, ensuring efficiency in challenging scenarios. The results showed that LBP has limited performance, achieving 82.41% accuracy on the original dataset but degrading significantly under severe compression due to its reliance on local texture patterns. ELM, on the other hand, demonstrated greater robustness, reaching 96.9% accuracy on the original dataset and maintaining good performance under moderate compression, though with a decline under extreme compression. YOLO-M was the most resilient and accurate among the methods, achieving 98.99% accuracy on the original dataset and maintaining superior performance even under severe compression, such as JPEG2000 with high quantization levels. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparing the efficiency of YOLO-M for face recognition in images and videos degraded by compression artifacts

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Fernando Rodrigues Trindade Ferreira,
  • Loena Marins do Couto,
  • Guilherme de Melo Baptista Domingues

摘要

Facial recognition is essential for public and private security, widely used to identify individuals and manage access with precision. This study evaluates the impact of image compression on the accuracy of facial recognition models, considering three types of codecs: HEVC, H.264, and JPEG2000, which exhibit distinct compression characteristics. Three facial recognition models were analyzed: Local Binary Pattern (LBP), Extreme Learning Machine (ELM), and the proposed YOLO-M model. The YOLO-M, a hybrid architecture developed in this study, combines YOLOv8 for rapid face detection with ResNet-50, which extracts detailed and robust features. The extracted information is integrated and processed through convolutional and dense layers, culminating in a softmax classifier, ensuring efficiency in challenging scenarios. The results showed that LBP has limited performance, achieving 82.41% accuracy on the original dataset but degrading significantly under severe compression due to its reliance on local texture patterns. ELM, on the other hand, demonstrated greater robustness, reaching 96.9% accuracy on the original dataset and maintaining good performance under moderate compression, though with a decline under extreme compression. YOLO-M was the most resilient and accurate among the methods, achieving 98.99% accuracy on the original dataset and maintaining superior performance even under severe compression, such as JPEG2000 with high quantization levels.