Dematting face images for use in a face recognition system

Abstract

Face recognition is in the fields of biometrics, machine vision and pattern recognition and has a wide application, including issues related to security systems. Since various factors such as ambient lighting, noise, and image opacity are more or less effective in the performance of face recognition methods, therefore, investigating the methods of removing blur from face images used in face recognition algorithms, in order to improve recognition accuracy, is of particular importance.

Given that the basic problem in all image enhancement methods is to understand the type and characteristics of the point spread function (PSF) related to the image blurring factor. Therefore, in a part of the proposed method in this thesis, by learning the previous knowledge from training examples that include artificially blurred face images, the PSF of the blurring agent of face images has been identified. According to the proposed method in this thesis, first in the training phase, we artificially matte the sets of facial images related to the ORL database using a few specific PSFs and then add white noise with an average power of 30 dB to them. Now, we put the features consisting of the maximum information related to the size of the frequency components of the matted images with the same PSF in one category and use the MLP neural network to learn the knowledge from the created feature space. Then, in the test phase, the input opaque face image, which has an unknown opaque PSF, is mapped to the feature space of the training phase, and we extract the previous features from the mapped image. Now, with the help of previously trained neural network, we select the category closest to this image from among the trained categories, and consider the blurring PSF of the images related to this category as the blurring PSF of the input matte face image. In the following, according to this PSF and using deconvolution method (convolution photo), we improve the input image and deliver the improved image to the face recognition system for recognition.

During the proposed method in this thesis, by creating a special feature space consisting of the maximum information related to the size of the frequency components of opaque images, we succeeded in increasing the PSF identification accuracy (recognition accuracy higher than 80% in noisy conditions). And as a result, we have increased the accuracy of the face recognition system (increasing the accuracy of the recognition system from 19.833% to 90.837%) by this method. Also, the use of neural network to identify PSF, on the one hand, has reduced the average execution time of this method by 172.41% compared to the modern methods presented in this field, and on the other hand, it has increased the hardware implementation capability of this method compared to existing methods.

Key words: removing blur from face images, face recognition systems, point expansion function, feature space learning, MLP neural network

1-1 Introduction

The topic of face recognition is widely discussed in fields such as image processing, machine vision, pattern recognition, neural networks and machine learning. The face recognition system is a biometric system (biometer) [1] that identifies or confirms the identity of a human using intelligent and automatic methods.

Many methods for face recognition have been provided, and these methods are generally divided into the following two categories:

a) Pattern-based methods

Pattern-based methods They operate based on the comparison of the input image with sets of patterns related to the structure of the face. These patterns are created using statistical tools such as support vector machine (SVM[2]), principal component analysis (PCA[3]), linear discriminant analysis (LDA[4]) and independent component analysis (ICA[5]) from face images related to the training set. to show the differences in the face. In these methods, people's faces are matched with a predetermined model and the obtained data are stored as extracted features.

One ??of the advantages of the face recognition system is the appropriate power of identification, harmlessness, friendliness and naturalness of the method to recognize people. On the other hand, it should be mentioned that, unfortunately, automatic face recognition with the help of a machine still remains a scientific challenge. Among the reasons for this lack of success, we can mention the large amount of image data and, accordingly, the wide range of changes in this data, as well as the nature of the data itself. For example, the identification system should be less sensitive to age conditions and face makeup, hair style, facial expressions, facial angle changes, etc. be resistant To the mentioned factors, one can add the imaged opacity, light conditions, the background of the images and other imaging parameters. As variant factors such as environmental lighting, noise and image blurring can more or less (approximately) affect the face recognition performance, thus assessing deblurring methods in order to enhance the recognition accuracy has significant importance.

Since inferring the type and characteristics of the point spread function (PSF) of image blurring factor is the main problem in all image enhancement methods, therefore in this thesis, in part of proposed method, by learning the prior knowledge over the training set including the artificially degraded images, inferring the PSF is discussed.

According to the proposed method in this thesis, first a set of facial images on the ORL database are artificially blurred and white Gaussian noise of 30 dB is then added. Then we put the features comprised of maximum information related to magnitude of frequency domain components of images degraded by same PSF in a group and trained an MLP neural network over such constructed feature space in learning phase. Then at testing phase, we mapped blurred input facial image with an unknown blurring PSF to the learning stage feature space and extracted the features over the mapped image. Now using trained neural network, we selected the nearest group to this image, among learned groups and considered the blurring PSF of this group as the blurring PSF of facial input image. Finally, according to this PSF and using deconvolution, we improved the input image and delivered the improved image to a face recognition system.

With proposed method in this thesis, constructing an especial feature space comprised of maximum information related to magnitude of frequency domain components of degraded image, we have enhanced the PSF inference accuracy (inferring accuracy more than 80% in noise condition) and face recognition system accuracy (accuracy was improved from 19.833% to 90.837%) by this method. Also, because of using neural network to infer the PSF, running time is reduced by 41.172 percent compared to an examined novel method in this field.