Ensemble of Deep Siamese Neural Network for Large Scale Kinship Verification in the Wild
- 0 Collaborators
This project is a proof of concept of the technology of kinship verification based on face analysis. The proposed research technique compares favorably with the best results obtained in RFIW2020. The optimized models are used to build our demo and realize the kinship verification task. ...learn more
Project status: Under Development
Intel Technologies
Other
Overview / Usage
Automatic kinship verification is intended to determine whether certain individuals belong to the same family, it can be considered as an application of face recognition systems with additional restrictions. These restrictions have proven to be difficult to manage, therefore face recognition is an important source domain from which we can transfer knowledge to achieve better performance for kinship recognition. A lot of progress has been made after a decade of research work, yet automatic kinship verification remains a very difficult task. Although some very interesting results have been obtained on small-scale datasets, kinship verification on a larger-scale in unconstrained environments is far from being solved. The emergence of a new dataset ”FIW” has encouraged researchers to give more important results, but the best ones so far have only achieved an accuracy of 70%. This motivated the creation of the RFIW, a large-scale annual challenge that supports visual problems based on kinship on larger scales. It is a platform thatpermits the publishing of original work and the gathering and discussions between experts. It was first organized as a data challenge Workshop hosted in conjunction with ACM Multimedia 2017.
Methodology / Approach
Within the framework of this project, we used the FIW database for the implementation of our system and the evaluation of the performance of our algorithms. We suggest a Deep Siamese Neural Network to quantify the relative similarity between two individuals. The concept here is to extract features from two input face images by the deep siamese network, fuse these features by combining and concatenating, then the kinship is verified by using the fused features that are fed into a fully-connected network to obtain the similarity score between two faces. A multi-model fusion is used to improve the performance of our system, the suggested method is effective as illustrated by the experiments and according to the results we have obtained our architecture has improved the state of the art of automatic kinship verification (we achieved an average accuracy of 77.5% in the RFIW 2020 challenge)
Technologies Used
Python, TensorFlow, Keras, Google Colab