This project intends to propose an accurate and fast face recognition technology for the uncontrolled situation. In this scenario, the enrolled face images are usually taken in a controlled situation and have high quality while the test face images are captured in an uncontrolled environment and facing several severe challenges such as pose variations, occlusions, different illuminations, and expressions, etc. With the development of deep learning techniques and some large scale face datasets becoming available, it is possible for us to address the face recognition in the uncontrolled situation. To end up with this, we propose 20facenet and the results show that we can surpass state-of-the-art performance.
Existing deep learning techniques:
We start the research with state-of-the-art deep learning techniques which are Inception based deep networks such as inception-v3 and inception-v4 in the above figure.
The facenet baseline consists of two components, Multi-task CNN Face Detection (MTCNN) network for face detection, and Resnet-Inception-V1 Face Recognition for accurate and fast face recognition. We first introduce technique details in Resnet-Inception-V1 deep mode. There are three main strengths for the network architecture of Resnet-Inception-V1 which are:
- using residue connection
- efficient inception module
- deeper networks with computations.
We propose to combine several strategies to improve state-of-the-art performance, including:
- select proper data augmentation
- change the probability ratio of embedding needs to keep and
- search for better convergency of the model training.
The results demonstrate that our method achieves the best accuracy on public training datasets. CosFace obtains the best accuracy on private training datasets.
NWO - Toegepaste en Technische Wetenschappen