Yaşlar arası yüz iz çıkarımı ve tanıması

In this work we focus our study on two subjects: face track retrieval and recognition across age, and the analysis of Fisher vector on face verification. Age invariant face recognition lacks enough researches and dataset based on video. In our study, we present a novel face track dataset called "Harry Potter Movies Aging Dataset (Accio). The dataset is harvested from Harry Potter movie series. It provides face recognition challenges such as variant face pose and expression, illumination, and more importantly a challenge of facial appearance changes due to age progression since the movies span a period of ten years. As a result, this dataset introduces a great environment for studying face track retrieval and recognition across age. Previous datasets conducted on age invariant face recognition use still images (FGNET, MORPH, CACD) either contains small number of data or the data spans small range of years. Our dataset contains large number of face tracks (nearly 38K face tracks) with variant age groups that span an age range between 10 to 88 years. Typically face track recognition is harder than still image face recognition due to effects of tracking, motion blur and pose variation. Each movie of the dataset has different number of face tracks and characters. Nearly 60\% of the face tracks are named as they appear in the movies and the rest of face tracks act as distractors or in retrieval evaluation. Our dataset contains great distribution for age group and most of face tracks belong to young characters. This is an important property to study the impact of aging factor since facial appearance changes is more in early ages than older ones. For face track representation, we use the state-of-the-art descriptor: Fisher vector. Fisher vector encoding aggregates large set of local descriptors of all face images into one high dimensional single vector representation. We define two primary tasks for retrieval and recognition: within and across movies evaluation. (i) \emph{Within movies} task uses the data of the individual movies for training and testing once to evaluate the performance against challenges such as pose and illumination variation. (ii) \emph{Across movies} task adopts evaluation between pairwise movies. Such as training data from HP-1 and testing data from HP-2 or vice versa. The purpose of this task is to assess the impact of aging factor across movies as the years' gap between movies increases. following the definition of evaluation tasks, we suggest three benchmark protocols for the dataset for the evaluation for future studies and researches: (i) \emph{restricted}: that uses only face track query for training and retrieval or recognition models should not use other Accio face tracks or any external data. The aim of this protocol is to compare the performance of face track descriptors. (ii) \emph{Unrestricted}: this protocol allows only for the usage of external data to train learning model for the retrieval and recognition evaluation on Accio dataset. (iii) \emph{Free-for-all} protocol which allows the usage of internal and external data for training. Using the suggested retrieval and recognition tasks, we introduce an extensive study for face track retrieval and recognition across age. The retrieval evaluation uses two popular performance measures: (ii) Mean Average Precision (mAP) and (ii) precision @k. We apply different experiments on Accio dataset. First Experiment is the across and within movies evaluation on all face tracks with full dimensional Fisher vector. The mAP and Precision @k results of this experiment show clearly the impact of age progression across movies. The performance of within movies evaluation is better than the performance of across movies evaluation due to the facial changes between movies. Following the baseline experiment, and in order to capture the similarity features of face tracks across movies and to improve the performance of this evaluation, we benefit from three metric learning algorithms: (i) \emph{low-rank Mahalanobis metric learning}, (ii) \emph{joint metric learning}, and (iii) \emph{diagonal metric learning}. These metric learning techniques serve two aims: (i) reducing the dimensions of Fisher vectors to make learning more applicable on large datasets, and (ii) increasing the face track retrieval and recognition performance by making the Fisher vectors more discriminative in the new projected subspace. Prior to the experiments of within and across movies retrieval using metric learning, we reduce the dimensionality of Fisher vectors of face tracks by PCA and then apply the learning algorithms. We use metric learning approaches in two scenarios: (i) metric learning on pairwise movies which combines percentage of face tracks for training a metric model and the uses it for the evaluation on the rest tracks for the two tasks of the evaluation, and (ii) metric learning on data from all movies which combines a percentage of face tracks from all Accio movies and uses them to obtain metric model. Then this model is used to evaluate on the rest of test face tracks for two evaluation tasks. In both scenarios of metric learning, low-rank and joint metric learning improve the results of the evaluation for both measures, mAP and precision @k which reflects their ability to capture variant facial changes through movies and to learn the similarity features between them. However diagonal metric is more basic and leads to slightly worse results than baseline results. For example, the average improvement in between the baseline and the joint learned from all movies is $42.7\%$. In addition there is a great improvement in precision @k such that the gaps between query and database movies is minimized. This shows the efficiency of these two approaches in learning the similarity between variant facial features across movies. Unlike the baseline where the performance declines for larger k values, results of precision @k is stable for different values in different k values. The last experiment on Accio dataset is face recognition across age. In this experiment and similar to retrieval tasks, we have within and across movies evaluation. We apply 5 fold across validation test on movies using linear SVMs models. SVM models are obtained from each movies' characters, and we use this model for the validation within the same and across movies. As expected, recognition accuracy in within movies evaluation is higher than accuracy in across movies' evaluation, because of the facial appearance variation between the data of training model and testing data. In the second part of this work, we evaluate and measure the performance of Fisher vector that we use in face track retrieval and recognition. Fisher vector is successful method and gives great results on face verification task using LFW dataset. However, in our work, the aim is to assess the performance of Fisher vector, using different dataset and features with various parameters. We use FRGC dataset rather than LFW. Specifically we evaluate its fourth experiment which has one set for training and two sets for test: query and target sets. We use Gabor filter as local descriptors rather than SIFT. We study the effect of spatial, scale and orientation augmentation to the features on the results of face verification. In addition we change the parameters of Fisher vectors such as the number of Gaussian Mixture Models and Feature-PCA dimension. In each experiment, we change one parameter while keeping the others fixed during the evaluation to see the impact of that parameter on the performance of Fisher vectors on face verification task. Since Fisher vector is efficient encoding system in features space, augmentation of the feature information improve the performance greatly. Furthermore, PCA dimension of features has an influence on the results when the dimensions are low, while GMM size does not have big impact on the results.