Face Datasets Research Articles

A Comprehensive Analysis of Face Recognition Based on Advanced Deep Learning Technology

Abstract. Facial recognition technology is becoming increasingly important in many fields due to its non-invasiveness and rapid improvements in accuracy. This study provides a comprehensive review of face recognition methods, covering both traditional and modern approaches. The key models such as feature face, Fischer face, Visual Geometry Group face (VGG-Face), FaceNet and ArcFace are analyzed, and the complete process from data preprocessing to feature extraction and model training is described in detail. By using the Wild Labeled Face Dataset (LFW) and the Network Face Dataset of the Institute of Automation of the Chinese Academy of Sciences (CASIA-WebFace) for experiments, the research found that although the deep learning model has high accuracy, it also faces challenges such as data dependence and large computing requirements. At the same time, the research highlights the issues of fairness and bias in these models, providing directions for future research and contributing to the sustainable and ethical development of face recognition technology.

Employing the Artificial Intelligence Object Detection Tool YOLOv8 for Real-Time Pain Detection: A Feasibility Study.

Effective pain management is crucial for patient care, impacting comfort, recovery, and overall well-being. Traditional subjective pain assessment methods can be challenging, particularly in specific patient populations. This research explores an alternative approach using computer vision (CV) to detect pain through facial expressions. The study implements the YOLOv8 real-time object detection model to analyze facial expressions indicative of pain. Given four pain datasets, a dataset of pain-expressing faces was compiled, and each image was carefully labeled based on the presence of pain-associated Action Units (AUs). The labeling distinguished between two classes: pain and no pain. The pain category included specific AUs (AU4, AU6, AU7, AU9, AU10, and AU43) following the Prkachin and Solomon Pain Intensity (PSPI) scoring method. Images showing these AUs with a PSPI score above 2 were labeled as expressing pain. The manual labeling process utilized an open-source tool, makesense.ai, to ensure precise annotation. The dataset was then split into training and testing subsets, each containing a mix of pain and no-pain images. The YOLOv8 model underwent iterative training over 10 epochs. The model's performance was validated using precision, recall, and mean Average Precision (mAP) metrics, and F1 score. When considering all classes collectively, our model attained a mAP of 0.893 at a threshold of 0.5. The precision for "pain" and "nopain" detection was 0.868 and 0.919, respectively. F1 scores for the classes "pain", "nopain", and "all classes" reached a peak value of 0.80. Finally, the model was tested on the Delaware dataset and in a real-world scenario. Despite limitations, this study highlights the promise of using real-time computer vision models for pain detection, with potential applications in clinical settings. Future research will focus on evaluating the model's generalizability across diverse clinical scenarios and its integration into clinical workflows to improve patient care.

SSAT: Active Authorization Control and User’s Fingerprint Tracking Framework for DNN IP Protection

As training a high-performance deep neural network (DNN) model requires a large amount of data, powerful computing resources and expert knowledge, protecting well-trained DNN models from intellectual property (IP) infringement has raised serious concerns in recent years. Most existing methods using DNN watermarks to verify the ownership of the models after IP infringement occurs, which is reactive in the sense that they cannot prevent unauthorized users from using the model in the first place. Different from these methods, in this article, we propose an active authorization control and user’s fingerprint tracking method for the IP protection of DNN models by utilizing sample-specific backdoor attack. The proposed method inversely and multiplely exploits sample-specific trigger as the key to implement authorization control for DNN model, in which the generated triggers are imperceptible and sample-specific for clean images. Specifically, a U-Net model is used to generate backdoor instances. Then, the target model is trained on the clean images and backdoor instances, which are inversely labeled as wrong classes and correct classes, respectively. Only authorized users can use the target model normally by pre-processing the clean images through the U-Net model. Moreover, the images processed by the U-Net model will contain unique fingerprint that can be extracted to verify and track the corresponding user’s identity. This article is the first work that utilizes the sample-specific backdoor attack to implement active authorization control and user’s fingerprint management for DNN model under black-box scenarios. Extensive experimental results on ImageNet dataset and YouTube Aligned Face dataset demonstrate that the proposed method is effective in protecting the DNN model from unauthorized usage. Specifically, the protected model has a low inference accuracy (1.00%) for unauthorized users, while maintaining a normal inference accuracy (97.67%) for authorized users. Besides, the proposed method can achieve 100% fingerprint tracking success rates on both the ImageNet and YouTube Aligned Face datasets. Moreover, it is demonstrated that the proposed method is robust against fine-tuning attack, pruning attack, pruning attack with retraining, reverse-engineering attack, adaptive attack, and JPEG compression attack. The code is available at https://github.com/nuaaaisec/SSAT .

ADYOLOv5-Face: An Enhanced YOLO-Based Face Detector for Small Target Faces

Benefiting from advancements in generic object detectors, significant progress has been achieved in the field of face detection. Among these algorithms, the You Only Look Once (YOLO) series plays an important role due to its low training computation cost. However, we have observed that face detectors based on lightweight YOLO models struggle with accurately detecting small faces. This is because they preserve more semantic information for large faces while compromising the detailed information for small faces. To address this issue, this study makes two contributions to enhance detection performance, particularly for small faces: (1) modifying the neck part of the architecture by integrating a Gather-and-Distribute mechanism instead of the traditional Feature Pyramid Network to tackle the information fusion challenges inherent in YOLO-based models; and (2) incorporating an additional detection head specifically designed for detecting small faces. To evaluate the performance of the proposed face detector, we introduce a new dataset named XD-Face for the face detection task. In the experimental section, the proposed model is trained using the Wider Face dataset and evaluated on both Wider Face and XD-face datasets. Experimental results demonstrate that the proposed face detector outperforms other excellent face detectors across all datasets involving small faces and achieved improvements of 1.1%, 1.09%, and 1.35% in the AP50 metric on the WiderFace validation dataset compared to the baseline YOLOv5s-based face detector.

Fisher regularized discriminative broad learning system for visual classification

The Broad Learning System (BLS) is an innovative learning paradigm with significant success in image classification. However, the 0–1 labeling matrix employed in BLS struggles to align with the true data distribution, limiting the flexibility of the regression objective. The ℓ2-norm-based Broad Learning System (L2DBLS) introduces the ɛ-dragging technique to enhance labeling diversity, but the randomness inherent in ɛ-dragging weakens the label correlation within the same category. This paper proposes the Fisher Regularized Discriminative Broad Learning System (FRBLS) to tackle these issues and aims to achieve the following objectives: Firstly, the generated label matrix offers sufficient flexibility to maintain intra-class compactness and inter-class separability. Secondly, the constraints of Fisher Regularization on the same class of features ensure better alignment between samples and labels. Finally, the overall solution process is optimized using an ADAM-based alternating multiplier method, which ensures closed-form solutions at each iteration. Experimental results demonstrate that FRBLS achieves up to 98% accuracy on various face and object datasets, offering superior time efficiency and a 1% improvement in classification performance compared to recent state-of-the-art methods.

Evaluating the Impact of BoNT-A Injections on Facial Expressions: A Deep Learning Analysis.

Botulinum Toxin Type A (BoNT-A) injections are widely used for facial rejuvenation, but their effects on facial expressions remain unclear. This study aims to objectively measure the impact of BoNT-A injections on facial expressions using deep learning techniques. 180 patients aged 25-60 years who underwent BoNT-A application to the upper face were included. Patients were photographed with neutral, happy, surprised, and angry expressions before and 14 days after the procedure. A Convolutional Neural Network (CNN)-based Facial Emotion Recognition (FER) system analyzed 1440 photographs using a hybrid dataset of clinical images and the Karolinska Directed Emotional Faces (KDEF) dataset. The CNN model accurately predicted 90.15% of the test images. Significant decreases in the recognition of angry and surprised expressions were observed post-injection (p<0.05), with no significant changes in happy and neutral expressions (p>0.05). Angry expressions were often misclassified as neutral or happy (p<0.05), and surprised expressions were more likely to be perceived as neutral (p<0.05). Deep learning can effectively assess the impact of BoNT-A injections on facial expressions, providing more standardized data than traditional surveys. BoNT-A may reduce the expression of anger and surprise, potentially leading to a more positive facial appearance and emotional state. Further studies are needed to understand the broader implications of these changes.

Introducing a novel dataset for facial emotion recognition and demonstrating significant enhancements in deep learning performance through pre-processing techniques

Facial expression recognition (FER) plays a pivotal role in various applications, ranging from human-computer interaction to psychoanalysis. To improve the accuracy of facial emotion recognition (FER) models, this study focuses on enhancing and augmenting FER datasets. It comprehensively analyzes the Facial Emotion Recognition dataset (FER13) to identify defects and correct misclassifications. The FER13 dataset represents a crucial resource for researchers developing Deep Learning (DL) models aimed at recognizing emotions based on facial features. Subsequently, this article develops a new facial dataset by expanding upon the original FER13 dataset. Similar to the FER + dataset, the expanded dataset incorporates a wider range of emotions while maintaining data accuracy. To further improve the dataset, it will be integrated with the extended Cohn-Kanade (CK+) dataset.This paper investigates the application of modern DL models to enhance emotion recognition in human faces. By training a new dataset, the study demonstrates significant performance gains compared with its counterparts. Furthermore, the article examines recent advances in FER technology and identifies critical requirements for DL models to overcome the inherent challenges of this task effectively. The study explores several DL architectures for emotion recognition in facial image datasets, with a particular focus on convolutional neural networks (CNNs). Our findings indicate that complex architecture, such as EfficientNetB7, outperforms other DL architectures, achieving a test accuracy of 78.9 %. Notably, the model surpassed the EfficientNet-XGBoost model, especially when used with the new dataset. Our approach leverages EfficientNetB7 as a backbone to build a model capable of efficiently recognizing emotions from facial images. Our proposed model, EfficientNetB7-CNN, achieved a peak accuracy of 81 % on the test set despite facing challenges such as GPU memory limitations. This demonstrates the model's robustness in handling complex facial expressions. Furthermore, to enhance feature extraction and attention mechanisms, we propose a new hybrid model, CBAM-4CNN, which integrates the convolutional block attention module (CBAM) with a custom 4-layer CNN architecture. The results showed that the CBAM-4CNN model outperformed existing models, achieving higher accuracy, precision, and recall metrics across multiple emotion classes. The results highlight the critical role of comprehensive and diverse data in enhancing model performance for facial emotion recognition.

Diagnosis and Screening of Velocardiofacial Syndrome by Evaluating Facial Photographs Using a Deep Learning-Based Algorithm.

Early detection of rare genetic diseases, including velocardiofacial syndrome (VCFS), is essential for patient well-being. However, their rarity and limited clinical experience of physicians make diagnosis challenging. Deep learning algorithms have emerged as promising tools for efficient and accurate diagnosis. This study investigates the use of a deep learning algorithm to develop a face recognition model for diagnosing VCFS. The study employed publicly available labeled face datasets to train the multitask cascaded convolutional neural networks (MTCNN) model. Subsequently, we examined the binary classification performance for diagnosing VCFS using the most efficient face recognition model. A total of 98 VCFS patients (920 facial photographs) and 91 non-VCFS controls (463 facial photographs) were randomly divided into training and test sets. Additionally, we analyzed whether the classification results matched the known facial phenotype of VCFS. The face recognition model demonstrated high accuracy, ranging from 94% to 99%, depending on the training dataset. The accuracy of the binary classification diagnostic model varied from 81% to 88% when evaluating with photographs taken at various angles, but reached 95% evaluating with frontal photographs only. Gradient-weighted class activation mapping heatmap revealed the high importance level of perinasal and periorbital areas, exhibiting consistency with the conventional facial phenotypes of VCFS. This study shows the feasibility and effectiveness of MTCNN-based model for detecting VCFS solely from facial photographs. The high accuracy underscores the potential of deep learning in aiding early diagnosis of rare genetic diseases, facilitating timely interventions for patient care.

Application of Visual Transformer in Low-resolution Thermal Infrared Image Recognition

Abstract Addressing the challenges of inadequate accuracy and limited robustness exhibited by current lightweight object detection networks specifically tailored for low-resolution thermal infrared face detection scenarios, this paper delves into developing an ultra-lightweight thermal infrared face detection algorithm that leverages visual attention mechanisms. To ascertain the optimal neural network complexity, a series of comparative experiments are meticulously conducted. With Yolo-FastestDet serving as the benchmark, this study endeavors to compress the backbone network, striking a delicate balance between network depth and detection speed. Additionally, to bolster the network’s capacity for profound feature extraction and precise discrimination of target edges and small objects, a Transformer-Encoder-based visual attention module is seamlessly integrated. Consequently, a lightweight face detection algorithm, enriched with attention mechanisms, is formulated. Furthermore, to mitigate the scarcity of low-resolution infrared face image data, a self-constructed visible-infrared face dataset is employed for training and evaluation purposes. The experimental outcomes reveal that the proposed algorithm attains an impressive mAP@0.5 score of 0.953 on the test dataset while satisfying the stringent real-time detection criterion of 30 frames per second (FPS) when deployed on an embedded Raspberry Pi CPU.

Pain assessment from facial expression images utilizing Statistical Frei-Chen Mask (SFCM)-based features and DenseNet

Estimating pain levels is crucial for patients with serious illnesses, those recovering from brain surgery, and those receiving intensive care etc. An automatic pain intensity estimator is proposed in this study that gathers information about pain and intensity from the user’s expressions. The faces in the database are first cropped using a ‘Chehra’ face detector, which performs well even in wildly uncontrolled environments with a wide range of lighting and position fluctuations. The suggested technique extracts the beneficial and distinct patterns from facial expressions using novel Statistical Frei-Chen Mask (SFCM)-based features and DenseNet-based features. As it offers quick as well as accurate pain identification and pain intensity estimation, the Radial Basis Function Based Extreme Learning Machine (RBF-ELM) is employed for pain recognition and pain intensity level estimation using the characteristics. All the data is kept, updated and protected in the cloud because availability and high-performance decision-making are so important for informing physicians and auxiliary IoT nodes (such as wearable sensors). In addition, cloud computing reduces the time complexity of the training phase of Machine Learning algorithms in situations where it is possible to build a complete cloud/edge architecture by allocating additional computational resources and memory in use. The facial expression images from the UNBC-McMaster Shoulder Pain Expression Archive and 2D face dataset are used to test the proposed method. The measurement of pain intensity uses four stages. When compared to the results from the literature, the proposed work attains enhanced performance.

Perbandingan Kinerja Model Pre-Trained CNN (VGG16, RESNET, dan INCEPTIONV3) untuk Aplikasi Pengenalan Wajah pada Sistem Absensi Karyawan

Face recognition has become a key technology in improving the efficiency and security of modern employee attendance systems. This study compares the performance of three pre-trained Convolutional Neural Network (CNN) models - VGG16, ResNet50, and InceptionV3 - in the context of face recognition for employee attendance systems. The study evaluated the accuracy, consistency, and generalization of the models on a dataset of employee faces, using prediction accuracy, confusion matrix, and classification report measurement methods. Results showed InceptionV3 performed best overall, with high consistency and confidence, achieving up to 99% accuracy on the test data. ResNet50 showed consistent performance in some cases but required further fine-tuning, while VGG16 showed the worst performance. These findings have significant practical implications for the industry, recommending the use of InceptionV3 for the implementation of reliable face recognition-based attendance systems, with consideration of the use of confidence thresholds to optimize accuracy. This research also highlights the importance of further optimization, including hyperparameter fine-tuning and more sophisticated data augmentation strategies, to improve system performance under various work environment conditions.

Uncovering labeler bias in machine learning annotation tasks

AbstractAs artificial intelligence becomes increasingly pervasive, it is essential that we understand the implications of bias in machine learning. Many developers rely on crowd workers to generate and annotate datasets for machine learning applications. However, this step risks embedding training data with labeler bias, leading to biased decision-making in systems trained on these datasets. To characterize labeler bias, we created a face dataset and conducted two studies where labelers of different ethnicity and sex completed annotation tasks. In the first study, labelers annotated subjective characteristics of faces. In the second, they annotated images using bounding boxes. Our results demonstrate that labeler demographics significantly impact both subjective and accuracy-based annotations, indicating that collecting a diverse set of labelers may not be enough to solve the problem. We discuss the consequences of these findings for current machine learning practices to create fair and unbiased systems.

Teaching AI to Summarize Like a Human: AReinforcement Learning Experiment

Text summarization is a crucial task in natural language processing (NLP), aiming to distill extensive information into concise and coherent summaries. Traditional summarization methods, including both extractive and abstractive techniques, face challenges in generating summaries that balance brevity and informativeness. This paper explores the application of Reinforce- ment Learning with Human Feedback (RLHF) to address these challenges and enhance the quality of text summarization.We introduce an RLHF-based approach using the FLAN-T5-small model, which integrates human feedback into the reinforcement learning framework to refine summary generation. Our method leverages a dataset from the Hugging Face datasets library, consisting of diverse document-summary pairs. The model is pre-trained on a large corpus and fine-tuned using human feedback, which serves as a reward signal to guide the model towards generating more relevant and coherent summaries.Our experimental results demonstrate that the RLHF-enhanced model significantly outperforms traditional summarization methods. Quantitative evaluations using ROUGE and BLEU metrics reveal substantial improvements in summary quality, with increases of up to 12.5% in ROUGE- 1 and 9.8% in BLEU scores over baseline methods. Qualitative assessments by human evaluators further confirm that the RLHF-based model produces summaries that are more aligned with human expectations in terms of coherence and relevance.This study highlights the potential of RLHF to overcome the limitations of conventional summarization tech- niques, offering a robust framework for generating high-quality summaries across various domains. Future work will explore the scalability of this approach to more complex summarization tasks and the integration of additional feedback mechanisms to further enhance performance.

Beyond the visible: thermal data for facial soft biometric estimation

In recent years, the estimation of biometric parameters from facial visuals, including images and videos, has emerged as a prominent area of research. However, the robustness of deep learning-based models is challenged, particularly in the presence of changing illumination conditions. To overcome these limitations and unlock new opportunities, thermal imagery has arisen as a viable alternative. Nevertheless, the limited availability of datasets containing thermal data and the small amount of annotations on them limits the exploration of this spectrum. Motivated by this gap, this paper introduces the Label-EURECOM Visible and Thermal (LVT) Face Dataset for face biometrics. This pioneering dataset includes paired visible and thermal images and videos from 52 subjects along with metadata of 22 soft biometrics and health parameters. Due to the reduced number of existing datasets in this domain, the LVT Face Dataset aims to facilitate further research and advancements in the utilization of thermal imagery for diverse eHealth applications and soft biometric estimation. Moreover, we present the first comparative study between visible and thermal spectra as input images for soft biometric estimation, namely gender age and weight, from face images on our collected dataset.

Face shape transfer via semantic warping

Face reshaping aims to adjust the shape of a face in a portrait image to make the face aesthetically beautiful, which has many potential applications. Existing methods 1) operate on the pre-defined facial landmarks, leading to artifacts and distortions due to the limited number of landmarks, 2) synthesize new faces based on segmentation masks or sketches, causing generated faces to look dissatisfied due to the losses of skin details and difficulties in dealing with hair and background blurring, and 3) project the positions of the deformed feature points from the 3D face model to the 2D image, making the results unrealistic because of the misalignment between feature points. In this paper, we propose a novel method named face shape transfer (FST) via semantic warping, which can transfer both the overall face and individual components (e.g., eyes, nose, and mouth) of a reference image to the source image. To achieve controllability at the component level, we introduce five encoding networks, which are designed to learn feature embedding specific to different face components. To effectively exploit the features obtained from semantic parsing maps at different scales, we employ a straightforward method of directly connecting all layers within the global dense network. This direct connection facilitates maximum information flow between layers, efficiently utilizing diverse scale semantic parsing information. To avoid deformation artifacts, we introduce a spatial transformer network, allowing the network to handle different types of semantic warping effectively. To facilitate extensive evaluation, we construct a large-scale high-resolution face dataset, which contains 14,000 images with a resolution of 1024 × 1024. Superior performance of our method is demonstrated by qualitative and quantitative experiments on the benchmark dataset.

Quadratic Matrix Factorization With Applications to Manifold Learning.

Matrix factorization is a popular framework for modeling low-rank data matrices. Motivated by manifold learning problems, this paper proposes a quadratic matrix factorization (QMF) framework to learn the curved manifold on which the dataset lies. Unlike local linear methods such as the local principal component analysis, QMF can better exploit the curved structure of the underlying manifold. Algorithmically, we propose an alternating minimization algorithm to optimize QMF and establish its theoretical convergence properties. To avoid possible over-fitting, we then propose a regularized QMF algorithm and discuss how to tune its regularization parameter. Finally, we elaborate how to apply the regularized QMF to manifold learning problems. Experiments on a synthetic manifold learning dataset and three real-world datasets, including the MNIST handwritten dataset, a cryogenic electron microscopy dataset, and the Frey Face dataset, demonstrate the superiority of the proposed method over its competitors.

Exploring de-anonymization risks in PET imaging: Insights from a comprehensive analysis of 853 patient scans

Due to their high resolution, anonymized CT scans can be reidentified using face recognition tools. However, little is known regarding PET deanonymization because of its lower resolution. In this study, we analysed PET/CT scans of 853 patients from a TCIA-restricted dataset (AutoPET). First, we built denoised 2D morphological reconstructions of both PET and CT scans, and then we determined how frequently a PET reconstruction could be matched to the correct CT reconstruction with no other metadata. Using the CT morphological reconstructions as ground truth allows us to frame the problem as a face recognition problem and to quantify our performance using traditional metrics (top k accuracies) without any use of patient pictures. Using our denoised PET 2D reconstructions, we achieved 72% top 10 accuracy after the realignment of all CTs in the same reference frame, and 71% top 10 accuracy after realignment and mixing within a larger face dataset of 10, 168 pictures. This highlights the need to consider face identification issues when dealing with PET imaging data.

Quaternion-based 2D-DOST and stacked principal component analysis network for multimodal face recognition

Face recognition is widely utilized and has become a ubiquitous part of daily lives. However, original face images are often sensitive and unauthorized access could imperil personal information. To protect privacy and security of multimodal face data, we propose an original approach that integrates differential privacy and a lightweight convolutional neural network with unsupervised predefined quaternion-type filters. Firstly, different modalities of face images are banded together by means of full quaternion matrix representation for simultaneously processing. Then the proposed quaternion two-dimensional discrete orthogonal Stockwell transform is exploited for implementing difference privacy, which is disturbed by virtue of Laplace and exponential mechanism. Afterwards, we devise a stacked quaternion two-dimensional principal component analysis network with three-stages for learning representative features. The proposed network model can enhance the discriminative characteristics through inserting nonlinearity, together with union of feature maps in different levels to achieve information complementation. Experiments conducted on four multimodal face datasets unveil that the proposed method achieves superior recognition performance in comparison with other state-of-the-art approaches.

Cross-Modal Face Super-Resolution Based on Quasi-Siamese Domain Transfer Fusion Network

In this paper, we propose a Cross-Modal Face Super-Resolution (CMFSR) method to construct high-resolution (HR) facial images from low-resolution (LR) cross-modal facial images captured respectively by disjoint visible light (VIS) and near-infrared (NIR) cameras. Due to the coupling of modality transformation and information fusion, CMFSR is more difficult to obtain HR reconstructed results compared with traditional super-resolution. To solve this problem, a Quasi-Siamese Domain Transfer Fusion Network (QSDTFN) for CMFSR is proposed in this paper, whose two branches transfer two LR face modality to HR face modality by domain transfer respectively. Different from two completely independent branches in the traditional pseudo-siamese network, only the HR-to-LR face transfer processes of the two branches in our quasi-siamese network are independent, while the LR-to-HR face transfer processes are coupled. This coupled module called the Adaptive Weighted Domain Transfer Fusion Module (AWDTFM) disentangles the modality and identity information in the two LR faces, thus achieving modality transformation and identity information fusion simultaneously. In order to strengthen the optimization on the process of CMFSR, this method further introduces the backward QSDTFN to form a higher-level bidirectional structure with the forward QSDTFN, and specifically designs two types of losses: intra-network loss and inter-network loss, to constrain the modality and identity consistencies within one QSDTFN and between two QSDTFNs respectively. The experimental results on the challenging LR cross-modal face datasets demonstrate that the proposed method performs favorably against the state-of-the-art methods.

DSU-GAN: A robust frontal face recognition approach based on generative adversarial network

Face recognition technology is widely used in different areas, such as entrance guard, payment etc. However, little attention has been given to non-positive faces recognition, especially model training and the quality of the generated images. To this end, a novel robust frontal face recognition approach based on generative adversarial network (DSU-GAN) is proposed in this paper. A mechanism of consistency loss is presented in deformable convolution proposed in the generator-encoder to avoid additional computational overhead and the problem of overfitting. In addition, a self-attention mechanism is presented in generator–encoder to avoid information overloading and construct the long-term dependencies at the pixel level. To balance the capability between the generator and discriminator, a novelf discriminator architecture based U-Net is proposed. Finally, the single-way discriminator is improved through a new up-sampling module. Experiment results demonstrate that our proposal achieves an average Rank-1 recognition rate of 95.14% on the Multi-PIE face dataset in dealing with the multi-pose. In addition, it is proven that our proposal has achieved outstanding performance in recent benchmarks conducted on both IJB-A and IJB-C.