Scholarlink

<p>Given the vulnerability of facial recognition systems to various types of presentation attacks, they need effective Face Anti-Spoofing (FAS) solutions to remain secure. These solutions typically operate by analyzing the luminance information or the color texture (chroma component) of the images. This paper presents a Presentation Attack Detection (PAD) method based on the analysis of color textures. This method extracts seven color channels from the input RGB image, then fuses these channels diagonally to generate a single-layer image. Practical tests determine the channel composition—specifically, identifying which of the numerous possible channel compositions yields the best outcome. An independent convolutional neural network (CNN) completely automatically performs model training and classification after determining the optimal color channel composition. Nine public datasets, including CASIA-FASD, Replay-Attack, MSU-MFSD, WMCA, Replay-Mobile, 3DMAD, SiW, SWAN-Idiap, and OULU, served as testing grounds for the proposed method. The suggested method did very well in tests that went within and across databases. It did better than the best methods in the replay attack, CASIA, and MSU datasets, which had equal error rates of 0.001, 0.03, and 0.005. It also achieved competitive results in most cross-database tests, with a half-total error rate of 33.8% and 28.5% between the replay attack and MSU-CASIA datasets, respectively. The results showed the generalizability of the method to different datasets with different facial expressions and different spoof media.</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

Fusing different color-space channels in a single-layer image for presentation attack detection

  • 期刊论文

  • 发表日期：

    2025年8月5日

查看全文
  • Qasim Majeed,
  • Abdolhossein Fathi

摘要

Given the vulnerability of facial recognition systems to various types of presentation attacks, they need effective Face Anti-Spoofing (FAS) solutions to remain secure. These solutions typically operate by analyzing the luminance information or the color texture (chroma component) of the images. This paper presents a Presentation Attack Detection (PAD) method based on the analysis of color textures. This method extracts seven color channels from the input RGB image, then fuses these channels diagonally to generate a single-layer image. Practical tests determine the channel composition—specifically, identifying which of the numerous possible channel compositions yields the best outcome. An independent convolutional neural network (CNN) completely automatically performs model training and classification after determining the optimal color channel composition. Nine public datasets, including CASIA-FASD, Replay-Attack, MSU-MFSD, WMCA, Replay-Mobile, 3DMAD, SiW, SWAN-Idiap, and OULU, served as testing grounds for the proposed method. The suggested method did very well in tests that went within and across databases. It did better than the best methods in the replay attack, CASIA, and MSU datasets, which had equal error rates of 0.001, 0.03, and 0.005. It also achieved competitive results in most cross-database tests, with a half-total error rate of 33.8% and 28.5% between the replay attack and MSU-CASIA datasets, respectively. The results showed the generalizability of the method to different datasets with different facial expressions and different spoof media.