This paper proposes a novel video face identification method, named “3D-2D-DCNN cascade” that serially combines 3D and 2D deep convolutional neural networks (DCNNs) for robust video face recognition (FR). In our method, an input video (face) sequence is first divided into a number of sub-video sequences and each of the sub-video sequences is then used as an input to the 3D-DCNN, aiming to obtain a set of class-confidence scores for a given input video sequence. These class-confidence scores are aggregated in a novel way, resulting in the formation of our novel class-confidence matrix. Key characteristic of our method is to make use of this class-confidence matrix for fine-tuning 2D-DCNN, which is serially linked to 3D-DCNN, to obtain the final face identification results. To verify the proposed method, two popular video identification benchmarks, COX Face and YTC databases, were used. Compared to the best reported recognition results on these two benchmarks, our proposed method achieves better or comparable recognition performances.

本文提出了一种新颖的视频人脸识别方法，称为“ 3D-2D-DCNN级联”，该方法将3D和2D深度卷积神经网络（DCNN）串行组合以实现鲁棒的视频人脸识别（FR）。在我们的方法中，首先将输入视频（面部）序列分为多个子视频序列，然后将每个子视频序列用作3D-DCNN的输入，目的是获得一组给定输入视频序列的置信度得分。这些班级信心分数以新颖的方式汇总，从而形成了我们新颖的班级信心矩阵。我们方法的关键特征是利用该类置信矩阵对与3D-DCNN串行连接的2D-DCNN进行微调，以获得最终的人脸识别结果。为了验证所提出的方法，两个流行的视频识别基准，使用了COX Face和YTC数据库。与在这两个基准上报告的最佳识别结果相比，我们提出的方法可实现更好或相当的识别性能。