Research On Face Recognition Technology For Low-Quality 3D Point Cloud Data

3D point cloud face recognition has attracted great attention in computer vision due to its important theoretical research and application value.It has been widely used in areas such as security,monitoring and human-computer interaction.Point cloud face images only retain the 3D spatial information of feature points,without color information.Therefore,their performance in facial recognition is more stable than RGB images under scenarios such as changes in illumination and posture.Additionally,it can protect user privacy and security.Deep learning-based 3D point cloud face recognition has been extensively studied in recent years,and a series of rich achievements have been made in this field.However,it is still very difficult to face the challenges of low-quality point cloud face recognition(such as data noise,occlusion of target faces and changes in expression and posture,etc.).In our work,by utilizing the design of multi-scale attention fusion and loss functions,as well as the design of the deep residual equivariant mapping network model,the face recognition network is adjusted,combined,reorganized,and optimized.This has led to some progress in the issues of discriminative feature extraction,model robustness and generalization,and how to explore the relationship between multi-pose faces and frontal faces.The contributions of this paper are summarized as follows:(1)To address the issues of discriminative feature extraction and noise interference in the dataset for lightweight 3D point cloud face recognition algorithms,we design a lightweight and efficient network model and propose a point cloud face recognition algorithm based on multi-scale attention fusion and noise-resistant loss function.Firstly,the features of receptive fields of different sizes are generalized.Then,the multi-scale attention features are extracted,and high-level attention weights are utilized to guide the generation of low-level attention weights.Finally,channel fusion is 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 loss function is designed to deal with the possible negative impact of a large amount of noise in the dataset on the model training process,thus enhancing the robustness and generalization ability of the model.To alleviate the scarcity of low-quality point cloud data,this paper uses a Kinect device to collect and produce a point cloud face dataset containing 60 people and 9600 images.Experimental results on open-source datasets such as Lock3 DFace and Kinect Faces show that the proposed method achieves better performance on low-quality 3D face recognition accuracy.(2)To address the problem of imbalanced training data between frontal faces and multipose faces,as well as the difficulty in multi-pose face recognition caused by the loss of a large number of feature points due to pose changes in point cloud face recognition tasks,we proposed a direction-adaptive multi-pose point cloud face recognition method.This method uses the rotation angles of the face in three directions in space as thresholds to adjust the feature input to the direction-adaptive deep residual equivariant mapping module,achieving the mapping of multi-pose faces to frontal faces at the feature level.To address the issue that the center loss function cannot fully solve the inter-class constraint of multi-pose faces,we propose a pose loss function based on the center loss function,which expands the distance between the mapped faces of different classes and improves the model’s ability to recognize multi-pose faces.In addition,this paper chooses to improve on the basis of the backbone network Light CNN and designs a convolutional module that is more suitable for extracting feature information from 3D face synthesis images to extract feature information containing richer key points.This paper generates a multi-pose 3D face dataset containing10,000 identities based on 3DMM and converts the generated point cloud faces into depth maps,pitch angles and azimuth angle three-channel images to provide data support for the training of multi-pose point cloud face recognition models.