Face recognition method for preventing non-living body attack
A face recognition, non-living technology, applied in the field of face recognition, can solve problems such as loss of verification effect, misidentification of individual A, failure of face recognition scheme, etc., to achieve the effect of improving recognition performance and applicable scale, and improving immunity
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0031] Such as Figure 1-3 As shown, a face recognition method to prevent non-living attacks includes the following steps:
[0032] Step 1: Training 2D face recognition model and 3D face recognition model;
[0033] Step 2: Record the face;
[0034] Step 3: Recognize the face,
[0035] When recognizing the face, the thermal imaging module is used to determine whether the detected face is a real face or a three-dimensional model face, and combined with 2D and 3D face recognition models for joint judgment.
[0036] The 2D face data is an ordinary RGB face image, and the 3D face data is the depth map extracted from the point cloud and the pseudo-RGB data that is post-processed from the depth map. In the step 1, the 2D face recognition model and the 3D face recognition model The recognition training process is as follows: First, prepare the cropped face data and their corresponding person labels. Different labels represent different people, and the same person has one or more fa...
Embodiment 2
[0047] Such as Figure 1-3 As shown, a face recognition method to prevent non-living attacks, the complete process is as follows:
[0048] S1. Prepare data: organize a large amount of ID data, one ID corresponds to one person, and each ID contains multiple face data of the person;
[0049] S2. Model preparation: prepare the model (such as resnet50), and set the training loss function to ArcFace Loss;
[0050] S3. Model training: input data into the model (such as resnet50) for training, and use the gradient descent method to iterate the network to converge;
[0051] S4. Model extraction: Remove the last classification layer from the trained network, and use the remaining part as a face featuremap extractor;
[0052] S5. Enter a new face: input a face image, extract it as a feature map and store it in the face database (the face database contains n IDs);
[0053] S6. Face recognition: input a face image of a certain ID, extract it as a feature map through the model, compare ...
Embodiment 3
[0059] In the process of face recognition, assuming that the 3D face model is misrecognized, the joint recognition and correction results are shown in the table:
[0060]
[0061]
[0062] The theoretical basis for the joint recognition model to have high performance:
[0063] When the 3D face recognition model has a small recognition scale and misrecognition occurs during the large-scale crowd recognition process, it will be corrected by the 2D face recognition model with a wider recognition scale. Because the 2D distance difference between different individuals is greater than the distance difference between the wrong ID and the correct ID when the 3D model is misidentified, even if the 3D model is misidentified, the individual with the smallest joint distance will still be correctly identified. At the same time, the distance between the real person and the plane calculated by the 3D face recognition model will be far greater than the distance between the real person a...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More