Scholarlink

<p>In present research area, biometric-based authentication has become an active topic because it uses in various applications and has the necessity of accurate personal identification for security. While entering the secure facilities or unlocking the cellphones, the personal authentication model becomes a secure and convenient option. In the research community, periocular recognition attains high popularity in the present times due to its significance of the smaller template size than face recognition. In addition, when compared with the iris biometric, less user cooperation is required. The periocular region generally involves the skin texture, eyelashes, eye socket, eye shape, tear duct, eye fold, etc. Moreover, periocular region-based authentication system generates superior accuracy than other biometric classifications. Only a limited set of features are used by the traditional periocular-based authentication, which resulted to be highly poor performance. Hence, this research initiated deep learning-based method for tackling limitations creating by accurate and generalizable periocular recognition approach. Initially, the required images for biometric recognition such as iris, periocular, and sclera are taken from standard database. Further, collected images are preprocessed by Retinex for enhancing image quality. Preprocessed images were subjected to the feature extraction phase, where the Region Vision Transformer (RViT)-based Visual Geometry Group19 (VGG19) act as the feature extractor to take out the meaningful features from three images. These features are further fused using the cross-attention mechanism. In the end, the recognition task is carried out using the Adaptive Nested Dilated DenseNet (AN-DiDNet). In addition, for enhancing accuracy of recognition, the parameters in AN-DiDNet model are tuned using the Modified Hunting Strategy of Clouded Leopard Optimization (MHSCLO) algorithm. The authentication is performed from this recognition result. An effectiveness and robustness of proposed model are determined by comparison of state-of-the-art model regarding corresponding dataset.</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

A novel authentication framework using multimodal-based periocular biometrics data with feature integration and adaptive deep learning network

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • V. Sivasakthi,
  • S. Sujatha

摘要

In present research area, biometric-based authentication has become an active topic because it uses in various applications and has the necessity of accurate personal identification for security. While entering the secure facilities or unlocking the cellphones, the personal authentication model becomes a secure and convenient option. In the research community, periocular recognition attains high popularity in the present times due to its significance of the smaller template size than face recognition. In addition, when compared with the iris biometric, less user cooperation is required. The periocular region generally involves the skin texture, eyelashes, eye socket, eye shape, tear duct, eye fold, etc. Moreover, periocular region-based authentication system generates superior accuracy than other biometric classifications. Only a limited set of features are used by the traditional periocular-based authentication, which resulted to be highly poor performance. Hence, this research initiated deep learning-based method for tackling limitations creating by accurate and generalizable periocular recognition approach. Initially, the required images for biometric recognition such as iris, periocular, and sclera are taken from standard database. Further, collected images are preprocessed by Retinex for enhancing image quality. Preprocessed images were subjected to the feature extraction phase, where the Region Vision Transformer (RViT)-based Visual Geometry Group19 (VGG19) act as the feature extractor to take out the meaningful features from three images. These features are further fused using the cross-attention mechanism. In the end, the recognition task is carried out using the Adaptive Nested Dilated DenseNet (AN-DiDNet). In addition, for enhancing accuracy of recognition, the parameters in AN-DiDNet model are tuned using the Modified Hunting Strategy of Clouded Leopard Optimization (MHSCLO) algorithm. The authentication is performed from this recognition result. An effectiveness and robustness of proposed model are determined by comparison of state-of-the-art model regarding corresponding dataset.