Algoritmos para análise da qualidade de imagens faciais de acordo com requisitos do padrão ISO/ICAO

• Autor(a) principal: Parente, Rodrigo Loureiro
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFPB
• Texto Completo: https://repositorio.ufpb.br/jspui/handle/123456789/12913
• topic: Qualidade de imagem facial; ICAO; ISO/IEC 19794-5; Facial image quality; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

The face has traditionally been used in identity documents for visual recognition of a person, being this way one of the most used physical characteristics for biometric recognition. The task of confirming identity with traditional documents is done visually by a human being. Such a task demands time, human resources and precludes the agility of critical scenarios such as airports etc. In this context, the International Civil Aviation Organization (ICAO) has initiated a project for automatic validation of personal identity through an electronic passport. However, the accuracy of facial recognition is directly dependent on the quality of the captured face samples. Later, the International Organization for Standardization (ISO) has developed an international quality standard for facial imaging to be used in electronic passports. This work aims to present algorithms for evaluating the quality of facial images according to the four ISO/ICAO standard requirements described ahead. The blurring level analysis is performed by exploring the high frequency information, calculating the variance of the pixels after Sobel high-pass filtering. This proposed method achieved an error rate of 6,2%. The requirement for analysis of the presence of hair in front of the eyes is evaluated by detecting borders in the eyes region, aiming to detect hair threads in front of the eyes. This proposed method achieved an error rate of 11,9%. The detection of the presence of a hat/cap is performed through skin segmentation in the forehead region. This proposed method achieved an error rate of 7,2%. Finally, the detection of eyeglass frame covering eyes is ascertained by detecting horizontal edges and detecting straight line segments in the eye region. This proposed method achieved an error rate of 11,6%.