Siamese Network for Face Recognition in Degraded Conditions

In recent years, face recognition become a primordial technique in computer vision and machine learning. However, masks and many other accessories such as glasses, sunglasses and scarfs lead to face occlusion, which inhibits face recognition and degrades its performance. The automatic recognition of this kind of faces is challenging because: 1) the occluded area hides a significant part from the face, 2) there is a lack of annotated occluded face images for training, and 3) the various degraded conditions make the face recognition task more difficult and complex. To overcome these difficulties, we suggest a new approach for occluded face recognition grounded on the few-shot learning technique. Our suggested approach is based on a Siamese network based on the pre-trained Inception-v3 model for multi-class face recognition under degraded conditions. It aims to represent face images in a new embedding space by extracting the most significant features through the Inception-v3 network. Our proposed network is optimized utilizing the contrastive loss which is calculated between two input images. The suggested network overcomes the existing literature techniques and proves its performance against pre-trained models on both IST-EURECOM LFFD and EKFD datasets.