Face recognition using principal component analysis

Abstract

The purpose of this dissertation was to analyze the image processing method known as Principal Component Analysis (PCA) and its performance when applied to face recognition. This algorithm spans a subspace (called facespace) where the faces in a database are represented with a reduced number of features (called feature vectors). The study focused on performing various exhaustive tests to analyze in what conditions it is best to apply PCA. First, a facespace was spanned using the images of all the people in the database. We obtained then a new representation of each image by projecting them onto this facespace. We measured the distance between the projected test image with the other projections and determined that the closest test-train couple (k-Nearest Neighbour) was the recognized subject. This first way of applying PCA was tested with the Leave{One{Out test. This test takes an image in the database for test and the rest to build the facespace, and repeats the process until all the images have been used as test image once, adding up the successful recognitions as a result. The second test was to perform an 8{Fold Cross{Validation, which takes ten images as eligible test images (there are 10 persons in the database with eight images each) and uses the rest to build the facespace. All test images are tested for recognition in this fold, and the next fold is carried out, until all eight folds are complete, showing a different set of results. The other way to use PCA we used was to span what we call Single Person Facespaces (SPFs, a group of subspaces, each spanned with images of a single person) and measure subspace distance using the theory of principal angles. Since the database is small, a way to synthesize images from the existing ones was explored as a way to overcoming low successful recognition rates. All of these tests were performed for a series of thresholds (a variable which selected the number of feature vectors the facespaces were built with, i.e. the facespaces' dimension), and for the database after being preprocessed in two different ways in order to reduce statistically redundant information. The results obtained throughout the tests were within what expected from what can be read in literature: success rates of around 85% in some cases. Special mention needs to be made on the great result improvement between SPFs before and after extending the database with synthetic images. The results revealed that using PCA to project the images in the group facespace is very accurate for face recognition, even when having a small number of samples per subject. Comparing personal facespaces is more effective when we can synthesize images or have a natural way of acquiring new images of the subject, like for example using video footage. The tests and results were obtained with a custom software with user interface, designed and programmed by the author of this dissertation.

O propósito desta Dissertação foi a aplicação da Analise em Componentes Principais (PCA, de acordo com as siglas em inglês), em sistemas para reconhecimento de faces. Esta técnica permite calcular um subespaço (chamado facespace, onde as imagens de uma base de dados são representadas por um número reduzido de características (chamadas feature vectors). O estudo realizado centrou-se em vários testes para analisar quais são as condições óptimas para aplicar o PCA. Para começar, gerou-se um faces- pace utilizando todas as imagens da base de dados. Obtivemos uma nova representação de cada imagem, após a projecção neste espaço, e foram medidas as distâncias entre as projecções da imagem de teste e as de treino. A dupla de imagens de teste-treino mais próximas determina o sujeito reconhecido (classificador vizinhos mais próximos). Esta primeira forma de aplicar o PCA, e o respectivo classificador, foi avaliada com as estratégias Leave{One{Out e 8{Fold Cross{Validation. A outra forma de utilizar o PCA foi gerando subespaços individuais (designada por SPF, Single Person Facespace), onde cada subespaço era gerado com imagens de apenas uma pessoa, para a seguir medir a distância entre estes espaços utilizando o conceito de ângulos principais. Como a base de dados era pequena, foi explorada uma forma de sintetizar novas imagens a partir das já existentes. Todos estes teste foram feitos para uma série de limiares (uma variável threshold que determinam o número de feature vectors com os que o faces- pace é construído) e diferentes formas de pre-processamento. Os resultados obtidos estavam dentro do esperado: taxas de acerto aproximadamente iguais a 85% em alguns casos. Pode destacar-se uma grande melhoria na taxa de reconhecimento após a inclusão de imagens sintéticas na base de dados. Os resultados revelaram que o uso do PCA para projectar imagens no subespaço da base de dados _e viável em sistemas de reconhecimento de faces, principalmente se comparar subespaço individuais no caso de base de dados com poucos exemplares em que _e possível sintetizar imagens ou em sistemas com captura de vídeo.