Invention Summary:
Face recognition is a highly active research area due to increasing security and commercial demands. However, recognition of low-resolution (LR) images in surveillance systems remains a challenge.
Rutgers researchers proposed an improved CNN network that utilizes pre-trained ResNet as the backbone for recognizing both LR and high-resolution (HR) face images. Feature super resolution (FSR) modules are inserted before the classifier of ResNet for low‐resolution face images. They are used to enhance the resolution of effective features extracted from LR face images. The proposed method is effective but computational-efficient. Experimental results show that the recognition accuracy for high‐resolution face images keeps high and the recognition accuracy for low‐resolution face images is improved..
Market Applications:
• Video surveillance used for security demands, commercial applications, and law enforcement applications, help in finding missing people.
Advantages:
- An improved CNN network for face recognition is proposed using ResNet as the backbone with FSR modules inserted between the FEN and classifier for LR and HR images.
- Pre-trained model is fully utilized, with shared parameters in FEN and a memory-efficient FSR module with only one hidden layer.
- Effective and resulting in high recognition accuracy for high-resolution face images and a significant improvement in recognition accuracy for low-resolution face images.
Intellectual Property & Development Status: Patent pending. Available for licensing and/or research collaboration. For any business development and other collaborative partnerships contact: marketingbd@research.rutgers.edu
