Selective Knowledge Distillation for Low-Resolution Face Recognition in the Wild

PROJECT TITLE :

Low-Resolution Face Recognition in the Wild via Selective Knowledge Distillation

ABSTRACT:

Low-resolution faces must be identified with minimal computational expense before face recognition algorithms may be used in the wild. Compressing a face model in order to increase speed and reduce memory consumption is a viable solution for this issue. This work presents a method for learning to detect low-resolution faces by selective knowledge distillation, inspired by that. Convolutional neural networks (CNNs) are used to recognise high-resolution and low-resolution faces with a teacher stream and a student stream, respectively, in this method. For high-accuracy recognition, the teacher stream is represented by a complicated CNN, whereas the student stream is represented by a much simpler CNN. As a result, we apply a sparse graph optimization problem to extract only the most relevant facial features from the instructor stream, which are then used to fine-tune the fine-tuning process of the student stream. Through the use of feature regression and low-resolution face classification to both approximate and recover the most informative facial cues, the student stream is actually trained to handle two jobs at once with limited computational resources in this manner. In tests, the student stream performs well in detecting low-resolution faces and consumes only 0.15 MB of memory and operates at 418 faces per second on CPU and 9433 faces per second on GPU.

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here