Gustavo Bezerra
Recognition of Occluded and Lateral Faces Using MTCNN, DLIB and Homographies
This Short Paper is the result of an improved version of a course final project. The course was lectured at the UFRN during the first semester of 2018. It was an introductory Deep Learning course.
The paper was published in SIBGRAPI 2018’s Workshop of Undergraduate Students; and presented as a poster.
The abstract is as follows. With the advance of technology it is possible to create more robust security systems. For this task, image processing alongside Deep Neural Networks are currently being used in several works for facial recognition. However, occlusions and faces in different angles are a challenge for most algorithms. Attempting to contour this issue, an algorithm for facial recognition combining MTCNN, DLIB and homographies is proposed. In the obtained results, a comparison between the proposed algorithm and basis works indicates that, for some controlled cases, a mean accuracy improvement of 7.4% was obtained, with a maximum of 8.23% for occluded faces and 14.08% for lateral faces.
The results obtained were satisfactory. However, in order to publish a full paper, a lot of work is necessary to improve the project.
The full paper is available at http://sibgrapi.sid.inpe.br/rep/sid.inpe.br/sibgrapi/2018/10.16.17.36?languagebutton=en. The poster can be found in 2018-sibgrapi-wuw.pdf.
Tags: Paper - Computer Vision - Facial Recognition - SIBGRAPI - Conference