Scholarlink

<p>Currently, human age estimation and gender classification are used in several tasks, such as public safety, video surveillance, and access control. A person's gait is a unique behavioral biometric that cannot be faked. It allows evaluating the age and classifying the gender of a person based on video recordings taken from a long distance and low-resolution images. The main challenges associated with recognizing a person by gait include wearing accessories and variations in clothing. Also, the shape of the silhouette makes it easier to distinguish between a male and a female, but the obtained image may not contain enough information to determine age. In this paper, we propose an approach based on transfer learning that aims to address these issues. The sinogram of the Radon transform from the gait energy image is fed into the Mobilenet and Densenet models. This procedure is applied in the feature extraction stage. They operate in parallel. Local Zernike moments are also extracted and fused with features from deep neural networks. The hML-KNN classifier, which combines the extracted features, is applied to improve the proposed approach's accuracy. The proposed method is evaluated on two datasets: CASIA-B and OU-ISIR OULP-Age. Explainable artificial intelligence with Grad-CAM is used to visualize the proposed model's performance. The experimental results were compared with other well-known models. The approach demonstrated high efficiency, achieving an average accuracy of 99.25% for the CASIA-B dataset, 95.35% for the female class, and 94.93% for the male class across different age groups for the OU-ISIR OULP-Age dataset. The application and development of the proposed approach will improve the functionality of automated information systems in the medical, law enforcement, and banking sectors and help experts in decision-making.</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

Human gender classification and age estimation based on gait images using deep learning

  • 期刊论文

  • 发表日期：

    2025年8月25日

查看全文
  • Rasim Alguliyev,
  • Ramiz Aliguliyev,
  • Lyudmila Sukhostat

摘要

Currently, human age estimation and gender classification are used in several tasks, such as public safety, video surveillance, and access control. A person's gait is a unique behavioral biometric that cannot be faked. It allows evaluating the age and classifying the gender of a person based on video recordings taken from a long distance and low-resolution images. The main challenges associated with recognizing a person by gait include wearing accessories and variations in clothing. Also, the shape of the silhouette makes it easier to distinguish between a male and a female, but the obtained image may not contain enough information to determine age. In this paper, we propose an approach based on transfer learning that aims to address these issues. The sinogram of the Radon transform from the gait energy image is fed into the Mobilenet and Densenet models. This procedure is applied in the feature extraction stage. They operate in parallel. Local Zernike moments are also extracted and fused with features from deep neural networks. The hML-KNN classifier, which combines the extracted features, is applied to improve the proposed approach's accuracy. The proposed method is evaluated on two datasets: CASIA-B and OU-ISIR OULP-Age. Explainable artificial intelligence with Grad-CAM is used to visualize the proposed model's performance. The experimental results were compared with other well-known models. The approach demonstrated high efficiency, achieving an average accuracy of 99.25% for the CASIA-B dataset, 95.35% for the female class, and 94.93% for the male class across different age groups for the OU-ISIR OULP-Age dataset. The application and development of the proposed approach will improve the functionality of automated information systems in the medical, law enforcement, and banking sectors and help experts in decision-making.