多尺度注意力融合和抗噪声的轻量点云人脸识别模型

在低质量点云人脸数据集上，判别性特征的提取和模型对噪声的鲁棒性是解决点云人脸识别问题 的关键。针对现有轻量点云人脸识别算法不能充分提取判别性特征和数据集中存在大量噪声而影响模型训练的问 题，设计轻量高效的网络模型，提出了基于多尺度注意力融合和抗噪声的自适应损失函数的点云人脸识别算法。 首先通过不同卷积模块获得不同感受野大小的特征图。然后进行多尺度的注意力特征提取，并使用高层的注意力 权重来引导低层注意力权重的生成，最后进行通道融合得到多尺度融合的特征，提升了模型捕获人脸细节特征的 能力。其次，根据低质量点云人脸图像的噪声信息特点，设计了一种新颖的抗噪声的自适应损失函数(anti-noise adaptive loss)，以应对数据集大量噪声对模型训练过程中可能造成的负面影响，提升模型的鲁棒性和泛化能力。 在开源数据集 Lock3DFace 和本文提出的 KinectFaces 数据集上的实验结果表明，与当前的主流算法相比该算法模 型在低质量点云人脸识别任务中具有更好的识别效果。

1. School of Information and Electronic Engineering,Shandong Technology and Business University,Yantai Shandong 264005, China; 2. Institute of Information Engineering,Chinese Academy of Sciences,Beijing 100195, China

The key to achieving point cloud face recognition is discriminative feature extraction and noise robustness for low quality data. To address the problems that the existing lightweight point cloud face recognition algorithms cannot adequately extract discriminative features and that the large amount of noise in the dataset affects model training, we designed a lightweight and efficient network model and proposed a point cloud face recognition algorithm based on multi-scale attention fusion and noise-resistant adaptive loss function. Firstly, the features of receptive fields of different sizes were generalized. Then, the multi-scale attention features were extracted, and high-level attention weights were utilized to guide the generation of low-level attention weights. Finally, channel fusion was performed to obtain multi-scale fusion features, which improved the model’s ability to capture face details. Meanwhile, according to the noise information characteristics of low-quality point cloud face images, a novel anti-noise adaptive loss function was designed to deal with the possible negative impact of the large amount of noise in the dataset on the model training process, thus enhancing the robustness and generalization ability of the model. Experiments on open-source datasets such as Lock3Dface and KinectFaces show that the proposed method yields better performance on low-quality 3D face recognition accuracy.