Desenvolvimento de uma aplicação web com detecção e reconhecimento facial

Abstract

Humans naturally perform the act of detecting and recognizing faces, but for a computer the task of describing objects from images is not trivial. Computer Vision had a revolution after the use of convolutional neural networks to identify patterns. A field benefited is the security field since the systems that exploit the detection of faces and the recognition of them had an increase of precision. This work develops a web application for viewing and organizing information about methods known in the literature, methods of detection and of facial recognition. In addition, the work uses images taken from security cameras, and demonstrates detections and recognitions in an organized manner in the application, since gathering information about such methods can be a exhausting task. This work details the implemented system that uses its methods of public libraries that contribute to the advance of the field and comments the obtained results. Literature in this area generally works with controlled environments and high resolution images to obtain good results, this work describes its results from tests performed on existing image sets and also images collected from the security camera. The detection tests on ready image sets assess the ability of different methods to detect changes of angle and illumination, and it is concluded that extreme angles are not detected using the convolutional neural net method and illumination is a determining factor in detection. In camera image sets detection is evaluated on different methods, but without extreme angle or illumination variations, and at the stipulated distances 100% detection is obtained. The recognition tests on a set of ready images show the difficulty of recognition in real situations, in camera images it was possible to evaluate the distances at which recognition begins to contain false negatives for the same person due to the low resolution of one of the photos.