Robust 3D face recognition in unconstrained environment using distance based ternary search siamese network

• Published in: Multimedia tools and applications 2024-05, Vol.83 (17), p.51925-51953
• Main Authors: Bhavani, Siriki Atchuta, Karthikeyan, C.
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Computer Communication Networks; Computer Science; Computer vision; Data Structures and Information Theory; Face recognition; Image contrast; Image enhancement; Image quality; Machine learning; Multimedia Information Systems; Neural networks; Special Purpose and Application-Based Systems; Three dimensional models
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-17545-6

Face recognition has recently become most important and popular in biometrics and computer vision applications. Nowadays, 3D face recognition fascinates researchers with the presence of 2D information and 3D face depth information. Many existing studies reported 2D face recognition, but its performance is reduced due to the influence of poses, illumination, expression, etc. Hence, this article introduces a novel deep learning (DL) model for effective 3D face recognition using point clouds. Initially, image denoising using enhanced trilateral filtering (En-TF) technique, contrast enhancement using fuzzy boosted histogram equalization (Fuzzy-BHE) technique and point cloud extraction are performed in the pre-processing stage to enhance the visual quality of the face image. Then, 3D face recognition is performed using a DL based distance ternary search Siamese neural network (DTS_SiaNet) technique. Within the SiaNet, a dual twofold channel attention convolutional neural network (TCAtt-CNN) technique is introduced to extract essential features from the face image data. In addition, an enhanced snake optimizer (En-SOp) technique is proposed for hyperparameter tuning of the network model. The proposed method is implemented in the Python platform, and the recognition is processed using the MIT-CBCL face recognition and Texas 3D face databases. For the MIT-CBCL database, the proposed method obtains an accuracy of 99.4%, kappa of 97.4%, true positive rate (TPR) of 94.6%, false acceptance ratio (FAR) of 0.0028 and time complexity of 58.3 s. For the Texas 3D face database, the proposed method obtains an accuracy of 99.31%, sensitivity of 97.5%, and F-measure of 97.2%.