Siamese Network and Facial Ratios for Deformed Facial Matching

Facial Recognition is a technique that uses the face and its features to identify/verify a person. It is a Biometric Application, which is playing an integral role in today’s world in a wide variety of areas like Criminal Identification, Visitor Verification, and many other Real Time Identification systems. In this paper, we present a system that does facial matching for deformed faces. The proposed method combines the extraction of high-level feature representations using Deep Learning and the calculation of facial ratios using Image Processing to generate a robust model that can be used to identify/verify deformed faces. Whereas traditional face recognition systems show poor results when there is a face deformation. We use a function that tells us how similar or how different the two input images are. So, for this Siamese network is used. This network takes in an image as input and gives its feature vector. Then, we find the distance between the computed feature vectors of the two images which will give us a similarity score. Next, we use an Image Processing technique to calculate the facial ratios from macro features like eyes, lips, etc. These facial ratios are calculated based on facial landmarks spread across the face. We use multiple facial ratios, so even if deformities occur in a part of the face it will not have much effect on the system rendering it immune to facial changes. Hence, this paper tries to close the gap in the previously devised facial matching methods which were not designed for deformed faces.