毕业生知识库

    随着深度学习技术的快速发展，人脸识别技术取得了重大突破并迎来了运用的井喷期，目前已被广泛地应用于银行金融、公共安防以及社交媒体等场景。此外，在大数据时代背景下海量人脸数据的采集变得更加方便快捷，丰富的人脸图像给人脸识别技术的研究奠定了数据基础，并带来了空前的机遇。然而，在实际应用中，人脸数据的采集通常是在非受控的自然场景下进行的，采集到的图像常含有大量的噪声以及分辨率、光照条件、姿态角度、面部表情、遮挡等变化，导致识别的性能急剧下降，给人脸识别技术的研究带来了巨大的挑战。如何有效地抑制这些噪声和变化，提取准确鲁棒的人脸图像特征，进而实现在非受控环境下有效的人脸识别，仍是当前人脸识别技术面临的重大挑战。

    本文以对抗学习为技术基础，借鉴对抗学习能隐式地约束分布之间的距离和生成逼真图像的特性，探索非受控环境下人脸识别问题中颇具代表性的视频人脸识别和跨姿态人脸识别问题。具体研究内容如下：

    1. 提出了一种基于注意集度量学习的视频人脸识别方法，通过结合最大均值差异与注意力机制更好地利用了视频中不同帧之间的关系，抑制了训练数据中的噪声，促进了卷积神经网络提取到更加鲁棒的人脸特征。同时，它也为我们后续更加深入地研究视频人脸识别打下了坚实的基础。

    2. 提出了一种基于对抗嵌入和变分融合的视频人脸识别方法，通过自对抗过程和一个变分推断结构来学习更有判别性的人脸特征和准确的融合模型。对抗嵌入和变分融合可以取代在视频人脸识别中广泛使用的深度度量学习方法，如对比损失、三元组损失等，而无需设置距离阈值超参数和复杂的采样策略，并可以取得更准确的视频人脸识别效果。

    3. 提出了一种基于偶代理姿态引导生成对抗网络的跨姿态人脸识别方法，通过将侧面姿态人脸依据先验域知识合成为正面姿态，进而进行跨姿态的人脸识别。偶代理姿态引导生成对抗网络不仅可以将侧面人脸正面化以提升人脸识别性能，也可以旋转正面人脸到任意的姿态来增广数据。

    With the rapid development of deep learning, face recognition has achieved major breakthroughs and ushered in the blowout period, which has been widely used in

banking finance, public security, and social media. In addition, collecting massive face data becomes more convenient and faster in the era of big data. Abundant face images have lain a foundation and brought unprecedented opportunities for the research of face recognition. However, in practical applications, face data is usually collected in uncontrolled environment, and the collected images often contain a large amount of noises and appearance variations incurred by resolution changes, illumination conditions, pose shifts, facial expressions, occlusions, etc. As a result, the performance of recognition drops drastically, which brings great challenges to the research of face recognition. How to effectively suppress these noises and variations, extract robust face features, and then achieve effective face recognition in uncontrolled environment are still major challenges in current field of face recognition.
    Inspired by the fact that adversarial learning has the ability to implicitly constrain the distances of distributions and generate photorealistic images, this thesis takes adversarial learning as the technical foundation and explores video face recognition and cross-view face recognition, which are two representative problems of face recognition in uncontrolled environment. The contributions of this thesis are summarized as follows:

    1. This thesis proposes an attention-set based metric learning method for video face recognition. By combining maximum mean discrepancy and attention mechanism, it makes better use of the relationships between different frames in a video, suppresses the noises in the training data, and promotes the convolutional neural network to extract more robust face features. In addition, attention-set based metric learning method also lies a solid foundation for us to further explore video face recognition.

    2. An adversarial embedding and variational aggregation method is proposed for video face recognition. Adversarial embedding and variational aggregation aims to learn discriminative embeddings and a precise aggregation model by a self-adversary process and a variational inference structure. It can replace widely used deep metric learning methods such as contrastive or triplet loss in video face recognition without margin hyper-parameters and sophisticated sampling strategies, and also achieve better performance on challenging benchmarks.

    3. An couple-agent pose-guided generative adversarial network is proposed for cross-view face recognition. It first frontalizes profile faces based on the priori knowledge, and then performs cross-view face recognition. Couple-agent poseguided generative adversarial network can not only promote the performance of face recognition by frontalizing profile faces, but also can rotate the frontal faces to arbitrary poses for data augmentation.