Rotation Invariant Face Recognition Using Radial Harmonic Fourier Moments

Face recognition is one of the most important tasks in the biometric domain which still receive more concern due to its immense variety of applications. Face recognition has a lot of challenges, one of these challenges is the rotation variation which highly affects the accuracy of the face recognition task. In this paper, we have proposed rotational invariant face recognition technique based on radial harmonic Fourier moments which have the ability to provide global rotation invariant features. The proposed method’s accuracy evaluated by carried out extensive experiments using two of the standard databases which are ORL and JAFFE. The results of these experiments indicated that the proposed method is rotation invariant and achieved high recognition rates in the presence of different variations


Introduction
Biometrics is an important domain that combines computer science and biological science [1].Biometrics is using a measurable physiological, or behavioral characteristic, e.g., faces, palm veins, fingerprints, DNA, iris, keystroke dynamics, mouse dynamics, gait, voice recognition, etc. [2,3].Face Recognition is one of the extremely important, significant and famous biometric techniques, which had crucial effects in daily life.Face recognition is given more attention because it is employed in many important fields as in digital library, control access, human-intelligent computer interaction, buildings security, security missions (missing children, terrorist and criminal identification), real-time matching of surveillance video images, employee entries and authentication in secure systems like computers or bank ATM [4,5] , as well as other purposes [6].
In the biometric domain, Face recognition field has a lot of advantages over other biometric fields because it doesn't need any special device Rotation Invariant Face Recognition Using Radial Harmonic Fourier Moments ……………………………………………………….Ali Sami Azeez 32 -99 -3102 -88 -comparing by another biometric system like fingerprint and iris.Furthermore, it easy to use as well as it is less hassle than other biometric systems due they do not require contact either the awareness of the subject.It can deal with legacy photography databases, videotape, and other image sources [7].However, Face recognition suffers from different challenges namely rotation, illumination, facial expression, wearing eyeglasses and pose variations, which affects the recognition accuracy [8][9][10][11] The rest of the paper has ordered as follows: An overview of the related work has presented in Section 2. RHFMs discussed in Section 3. Section 4 describes the distance measure used in the proposed method.The utilized databases has described in Section 5. Section 6 describes the proposed system.Details of experimental results are given in Section 7, while Section 8 including the concluded of the paper.

An over view of related work
Face recognition is one of the most important fields in the biometric domain that received more interest in last dedicates.Therefore, many face recognition methods have been existed based on different feature extraction techniques [12][13][14][15].Some of these techniques obtained high recognition rate.However, they aren't invariant to geometric transformation such translation, scaling, and rotation variations.In view of this, some of these methods are weak in the appearance of image noise; hence, they required additional efforts to remove the image noise.Orthogonal rotation invariant moments (ORIMs), are popularly used in digital image processing, there are several moments methods fall under ORIMs, which include Fourier-Mellin Moments (FMMs), Zernike Moments (ZMs), Pseudo Zernike Moments (PZMs) and RHFMs.The different and powerful characteristics of ORIMs prompted the researchers to use them in face recognition field.-89 -Therefore, many face recognition based FMMs, ZMs and PZMs methods have presented.Yee Ming Chen and Jen-Hong Chiang [16], Offered a mixed various features extraction method for face recognition module by using FMMs method, The Fourier-AFMT achievement extract frequency a constant features and OFMM method extracted moment constant features are employed individually, CCM(correlation coefficient method) is classifying and combining certain both kind of features.The test results reveal the average rate of accuracy recognition proposed techniques is far good.Same Authors have been proposed a face recognition based on FMMs by Fusing multiple features [17], a Taylor-AFMT, and Fourier-AFMT are distinguished for face recognition by a comparative study.Then they presented a hybrid face recognition framework based on Fourier-AFMT to extract the presented feature.The First step, extract the intensity facial and edges of the directionality features while the second step is using CCM to combine and classify two characters of features.Sajad Farokhi et al. [18], introduced a new face recognition approach based on Hermite kernels (HKs) and ZMs for coping with alteration expression of the face image, differences pose and scale of the head, the influences of time-lapse and using eyeglasses.The infrared images have employed to process the force of illumination changes on recognition of face image, and the infrared images have employed to process the force of illumination changes on face recognition.In addition, both local and global features combined together and used in the determination fusion step.ZMs have used to provide a global feature as a feature extractor, while the images of the face are divided into various spots and filtered spot-wise with HKs in The local part.The review of the main part which is followed by a linear that differentiate analysis is applied to data vectors to produce outstanding features.The fusion decision is applied to vector feature properly and the combination of both global and local features.The preliminary results achieved showed that the proposed ZMHK method improved the face recognition accuracy.Furthermore, it outperformed some other existing face recognition methods.Tolga Alasag and Muhittin Gokmen [19], Formulated a method that uses Local Zernike Moments (LZM) and Gauss scale space for face recognition in low-resolution face images.The face recognition framework has designed according to the test outcome.The outcome shows that the suggested framework is promising for real-world applications.Also, Evangelos Sarıyanidi et al. [20], proposed a new representation of local LZM for face recognition by measuring ZMs at each pixel of a face image using the local neighborhood of each pixel.Madeena Sultana et al. [21], introduced a technique to solve diverse illumination, pose, and expression Rotation Invariant Face Recognition Using Radial Harmonic Fourier Moments ……………………………………………………….Ali Sami Azeez 32 -99 -3102 -90 -condition which is one of biggest challenges in face recognition systems, by employed lower order PZM based method which can efficiently recognize faces regardless of illumination, pose, and expression change.Due to optimal choice of the features, they introduced a technique obtains much better recognition rate; Extensive experimentation proves the high recognition rate and robustness of the recommended technique under varying conditions.
The RHFMs have attractive characteristics such as rotation invariant, less redundancy, and high resistance against image noise.Thus, they have the capability to provide distinct and rotation features.Therefore, they have used in different applications as will discuss in section 3.Recently RHFMs has implemented in successfully in the biometric field due to their attractive characteristics.Ali Mohammed Sahan [22], has utilized the RHFMs as a feature extraction descriptor in the Palmprint Recognition system.The mentioned attractive features of RHFMs motivate us to employ it in our proposed face recognition system spatially it is not used in face recognition.So our contribution to this work, which we have examined the capabilities of RHFMs in face recognition system.

Radial Harmonic Fourier Moments (RHFMs)
In 2003, H Ren et al. have been proposed the RHFMs [23], RHFMs is one of The orthogonal rotation invariant moments (ORIMs) which have distinct characteristics such as less redundancy, rotation invariant, and robust against noise.As to other ORIMs methods, RHFMs have used to improve the image reconstruction, decrease both computational complexity and noise sensitivity, and magnitude invariance.therefore, RHFMs have used in different image processing applications such as image recognition [23], tumor cell recognition [24], cell image recognition [25], image reconstruction [26], Character reconstruction [27], Geometrically invariant image watermarking [28], Chinese Chess Character [29] and also in biometric palm print recognition [22]…etc.The definition of RHFMs is as following: [27] ∫ ∫ ( ) ( )

Distance measure
In this work, the classification stage has been used the Euclidean distance measure into distinguishing between the face images.The Euclidean distance measure between training and test features vectors can be defined as: [30] Let x; y be two M by N images, ( ), ( ), where ( ) ∑ ( )

Experimental analysis
In this part, the accuracy of the suggested method has been evaluated by carried out extensive experiments on two standard face image databases, which are ORL and JAFFE.We conduct detailed experiments on the above face databases to analyze the recognition performance of our suggested method.The performance has been compared with the similar methods published in the literature.For this purpose, we implement the recommended method and a few existing methods in Microsoft Visual C++ under Windows OS on a personal computer with 6 GB RAM and corei7 CPU.

Assessment the recognition accuracy over different variations
In this experiment, we have examined the accuracy of the proposed method in the presence of different variations such like position, pose, facial expression, illumination, detail, and scale.
For this purpose we have constructed dataset from ORL database by randomly selecting five images from each class for training and the remains five images used for testing,

Assessment the recognition accuracy over facial expressions variation
In order to evaluate the accuracy of the proposed method under facial expression, we have conduct experiment on JAFFE database.In this experiment, the training dataset is constructed by randomly selecting one face image from each of the seven different expression groups which are (surprise, sadness, anger, neutral, fear, happiness, and disgust) for every class, while, the remaining face images used for constructing the testing dataset.Therefore, the total number of the testing dataset is 143 while it is 70 for training datasets.Fig 4 shows some samples of the datasets used in this experiment.The results presented in Table 2 of the above-mentioned experiment refer that the highest recognition rate is %98.6 obtained at the order and repetition 10 which provides 121 features while the orders and repetitions 6 and 8 are achieved %97.9.

Assessment the recognition accuracy over rotation variation
The rotation variation is an important and big challenge in face recognition.Therefore, we have conduct experiments to assess the accuracy of the suggested method over different rotation angles.In these experiments, we have utilized the same training dataset in Section 7.1, while the testing dataset is constructed by rotating the testing dataset in Section 7.1 by three rotation angles which are

Conclusions
In This work, we propose an accurate and rotational invariant face method has been introduced by utilizing the RHFMs as feature descriptor.The experimental part of this work refers that the RHFMs can extract distinct global features, which have the ability to distinguish between different face images.Furthermore, these features are slightly affected by different rotation angles, which mean they are rotation invariant.The highest recognition rate obtained over different variations is %95.5 at the order and repetition 8, while its %98.6 at the order and repetition 10 over facial expression variation.The analysis of the outcome shows that the suggested method is robust against image noise.The comparison between the suggested method and other face recognition methods based on other ORIMs techniques indicates that the proposed method outperforms other face recognition methods that utilized other ORMs techniques such as ZMs and PSZMs.
. The face recognition techniques can be divided based on feature extraction method type into 3 categories, which are local, global and combined local and global technique.In local, face recognition technique the face image features are extracted from facial regions by measuring the distance between eyes, mouth, side of the nose, corner points, goatee, etc.While the global technique is based on extracting the features from the entire face image.The combined local and global technique is based on extract both local and global features.In this paper, an accurate and rotation invariant face recognition method has proposed based on RHFMs as a feature extraction descriptor.The proposed method exploits the capabilities of RHFMs in terms of providing rotation invariant global features of the human face image.The distance between the training and testing face images has measured by utilizing the Euclidean distance measure.

Fig 4 :
Fig 4: (a) samples of training dataset, (b) samples of testing dataset these experiments, we have considered the order and repetition 8.The results of these experiments, which are presented in table3, indicated that the recognition rates after rotation (%95.3,%95.1, and %95.2) are close to the recognition rate before a rotating basis (%95.5) which means that the offered system is rotation invariant.

Table 1 :
Recognition rates achieved using different RHFMs orders and repetitions under different variations

Table 2 :
Recognition rates achieved using different RHFMs orders and repetitions under facial expressions variation

Table 3 :
Recognition rates achieved using different RHFMs orders and repetitions over different rotation angles