A NOVEL FEATURE EXTRACTION BASED PERSON RECOGNITION USING LOCAL PHASE QUANTIZATION AND GEOMETRIC FEATURES

— At present, it is simple for everyone to generate digital pictures of their routine life and use them for different purposes. Similarly, facial recognition is a trending technology that can identify or verify an individual from a video frame or digital image from any source. There are numerous techniques involved in the working principle of facial recognition. But the simplified method is feature extraction by comparing the particular facial features of the images from the collected dataset. Multiple algorithms are existing for feature extraction, but they fail to give high accuracy. The proposed algorithm based on deep learning provides a high recognition rate by using a convolutional neural network for classification. For feature extraction, Local Phase quantization, Geometric-based features, and directional graph-based methods are implemented. Various performance metrics, such as recognition rate, classification accuracy, accuracy, precision, recall, F1-score is evaluated. The proposed method achieves high-performance values when it is compared with other existing methods. It is mainly developed to calculate the casual visit of a person to the mall, and it is also deployed for criminal identification.


I. INTRODUCTION
Face recognition can also be defined as biometric artificial intelligence relied on the application that is exclusively designated to recognize a person by exploring sequences based on the individual's facial shape and texture.Even though facial recognition has a wide variety of applications, it is typically essential to use it in the detection of criminal and forensic investigations and inquiries.Applying the digital images for the identification of victims can be challenging at specific occasions.Especially when the individual is wearing a mask or tattoo, or if the environment has background noises, camera distortion, insufficient storage, inadequate computing techniques, occlusion, detection of identical images and lowresolution images (Dutta, Gupta, & Narayan, 2017).
The impression of face recognition is to provide a computer to identify and analyze the human face in a fast way with high precision.Many algorithms are developed to improve its accuracy.Nowadays, deep learning explored the hidden sights of advanced computing applications and executed it in the right place to get an efficient expected output.It is trained similar to the human brain to analyze multiple human faces and store it in a database for future use.Generally, human utilizes all his sensory organs to recollect and analyze the input data, a similar computer process, and the input captured images into various features by a different algorithm.To get the exact desired result and confirm it by comparing it with the original picture by an integral section of biometrics and matching of essential traits.The system involved dual methods, such as face detection and face recognition (Amos, Ludwiczuk, & Satyanarayanan, 2016).Face detection is used to search and find one face by image processing, and face recognition is performed by comparing and analyzing the processed image with similar images from the database.
The method of face recognition in the proposed research comprises feature extraction and feature classification methods.In feature extraction, Phase quantization, directional graphbased process, and Geometric based feature extraction are performed.The feature classification is achieved by using convolutional neural networks for criminal identification.The mandatory methods for effective facial recognition should calibrate the facial expression from the given images.The facial expression can be derived from eyes, mouth, eyebrows, cheeks, and it's represented as salient regions.But few emotions like surprise and sadness acquired with only one salient feature, but anger and fear cannot be obtained from a single salient feature.So feature extraction of the image is mandatory to extract the actual property of the pictures.
In feature extraction, Phase quantization is applied to extract the features from blurred images by short Fourier transform to analyze the structural properties of the image with promising values of accuracy and efficacy (Xiao, Cao, Wang, & Li, 2017).Directional graph-based and geometric based methods in feature extraction are applied to extract the facial feature in ROI along with error filtering techniques to provide prominent face components.It is used to locate the edges and relative size and position of mandatory expression components such as mouth, eyes, nose, and eyebrows.Then it is differentiated into the unimportant and meaningful part and also the conversion of grayscale distribution into feature vector based on the value of the pixel.The feature extraction is based on image pixel distribution, orientation selectivity, and spatial localization.The extracted images are then fed into CNN for further classification by deep learning.It is composed of multiple layers to detect necessary edges, intricate shapes, and by further processing, the final layer can catch the entire face and confirm it by verifying with the high dimensional dataset.Hence it is mostly preferred in real-time applications.
Face recognition is also based on the biometric aspect, which comprises the advantages of high precision and low intrusion in a program that utilizes an individual's face to spontaneously detect and authenticate the individual from a digital picture from any source.It compares the chosen facial attributes from the image and faces the repository database, or hardware can also be employed to authenticate an individual.It is mainly deployed in genetic engineering and biometric devices for identification, authorization, verification, and authentication.

Fig 1-Schematic diagram of Convolutional Neural Network
Many companies have adopted face recognition for security purposes by implementing CCTV to manages and control the organization.Apart from security system, it have other exciting applications such as unlocking the phones, more refreshing advertising, finding missing people, protecting the law enforcement, identifying person on social media, diagnosing various types of diseases, spotting VIP at events, protecting school and government institutions from threat, and controls the access to sensitive regions.• To find the optimal path to achieve the best features using directional graph-based features.

A. The Objective of the research
• The proposed research using novel Convolutional Neural Network can be employed as an effective classifier and applied in the real-time dataset for criminal identification.

B. The Organization of the research
The organization of the research is as follows.Section II explains the related work of face recognition on different feature extraction and classification methods.Part III is based on the detailed description of the proposed work.Section IV represents the analysis of the performance of the proposed method, and section V concludes the results of the research work and explains its future extension of the work.

II. REVIEW OF EXISTING TECHNIQUES
In The detection and localization of blurred face using local phase quantization applying Fourier transform phase in resident neighborhoods (Wechsler & El Khiyari, 2018).Under certain conditions, the period can be represented as a blur invariant attribute.So in the analysis, a histogram of local phase quantization labels calculated in resident areas is executed as face descriptor, related to broadly applied local binary pattern for a description of face images.But to limit the impact of lighting alteration in face pictures, procedure on illumination normalization is applied that comprises of contrast equalization, gamma correction, and Gaussian filtering that increases the result of recognition accuracy.
Some investigations made on a pair of parent and child to verify the effectiveness of the local phase quantization.The standardized and cropped images of parent and child were fed as input and converted to grayscale.Then local phase quantization is applied to retrieve the local features of every mode (Garg & Kaur, 2016).Encrypted images are differentiated into k non overlap rectangle scratches, and histogram values define each stretch.Each histogram has allocated to the high dimensional feature vector.Then cosine similarity is analyzed between the images by projecting into transformed subspace, and the resultant score is compared with threshold value from ROC by performance measures.With the obtained result, it is decided whether the person belongs to the same family or not.
The face is an essential attribute considered in a security system.Even though it is applied vastly, there is some limitation such as illumination, pose, and condition of the pictures(Ghimire, Lee, Li, & Jeong, 2017).To overcome this issue, face recognition is performed under non-uniform illumination by CNN that can learn exclusively local patterns from information is utilized for identification of a face.The symmetry of face data is processed to eliminate and enhance the performance of the system by accepting the horizontal reflection of the face images.The experiment was made using the Yale dataset under various illumination settings and showed moderate enhancement in the performance of CNN.The production remains undisturbed by the horizontal reflection of images.But it is failed when it is applied with fisher discriminant analysis and fisher linear discriminant analysis that is based on Gabor wavelet transform.
Human gait is soft biometric that support to identify an individual by their walking gestures(C.Li, Min, Sun, Lin, & Tang, 2017).To enhance performance recognition, the videobased sensor is applied by using convolutional features.The depth gait is introduced by applying a pre-trained deeper network dataset without any apt tuning.It is performed by the integration of more soft biometric attributes by the stochastic model.The height and stride length is considered as input information for the gait recognition system and improves its performance.But the stochastic model has some issues in weighting computations in calibrating the score of soft biometrics.Hence the model comprises the bunch framework of gait feature extraction, Gait score estimation, estimation of soft biometric trait, and probabilistic smooth biometric system.The combined score is delivered from the probabilistic system.
The gender-based classification is also made possible using a supervised machine learning approach.The diverse classifier used in this approach is the support vector machine, convolutional neural network, and adaptive boost algorithm.The input images are fed into pre-processing for removing noisy information, and feature extraction is based on a geometric approach to extract meaningful data from the processed images.It is fed into the combination of a classifier to get the resultant output.It is used to recognize the gender of the picture.

III. PROPOSED METHODOLOGY
This section explains the proposed work of the research.The features from the image are extracted using Local phase quantization and geometric based features and classified using novel based convolutional neural network.It is executed in a shopping mall to identify the individual to evaluate the person's number of visit to that mall.
Here,   represents the basis vector of frequency q.   Vector represents all image samples from neighborhood   .
Local Phase quantization is defined as follows, -The original image is compared with the input image, and it is formulated as follows, Where ℎ(, ) is represented as a Point Spread Function (PSF), (, ) Is described as an additive noise -Then the Fourier transformation is performed to convert the image from the range of high intensity to low intensity.It is formulated as below, (, ) = (, ).(, ) Where (, ), (, ), (, )Fourier are transforms of (, ), (, ),  ℎ(, ) respectively.
-To analyze the impulsive response of an image, Point Spread Function is calibrated as below, To calculate the relative translation between two pixels, phase correlation is estimated as below, ∠(, ) = ∠(, ), for all (, ) ≥ 0.
Finally, the distribution histogram of the encoded values x within the image is constructed to yield local phase quantization in 256 dimensions.It is formulated as below Where   , the quantization of the i th element in W is given by, Ga = V T Fa∑ ()(, ) Although local phase quantization quantizes the coefficients of short term Fourier transform, this process tends to ensure blurinsensitivity, but somewhat loss discriminative power.It is only the sign of the STFT coefficients that affect texture characterization.

B. Directional Graph-based Features
In this process, the input image is represented as a graph where each pixel in the picture is mapped as times series of respective vertex (Mutlu & Oghaz, 2019) Where   implies the presence of edge.If   = 0, implies that the edge does not exist.There are several steps in assessing directional graphs, which are explained further.

Construction of Directional graph:
Step

n -Number of vertices
Step 4: Average Distribution Weight of the Graph To detect the similarities or dissimilarities between the vertices connected by the edges, weight is calculated using the below equation (17), Where W can be represented as the weight of all edges in a graph.

Step 5: Graph Hilbert Transform (GHT):
To convert the image and to extract the global features, the Hilbert transform method is applied in this work.The Hilbert transform of the graph of a function f(x) can be defined as below, The Hilbert transform of a constant is zero.If you compute the Hilbert transform in more than one dimension and one of the dimensions does not vary (is a constant), the transform will be zero (or at least numerically close to zero).

C. Geometric Features:
In a geometric-based approach, the local features that are based on local statistics and locations like bodily features are extracted.It can be implemented in the following ways.
• Convex Area Of Image

i. Convex Area of Image
The binary image covering the convex area gives the outline of the curved image.The convex area can be denoted as several pixels in a convex image.The convex area can be computed as shown in fig 3

Fig 3 -Convex area of Image j. The eccentricity of an image:
Eccentricity is the ratio of length pf short or minor axis to the length of the major or long axis of an image.For a linked region of a digital image, it is defined by neighbor graph and calibrated metrics.The number that measures and look of the ellipse by illustrating its flatness is also called as eccentricity.The eccentricity can be calculated by using the given equation, eccentricity = lenght of the short axis length of the long axis k.Filled Area: Full Image can be denoted as a binary image of the same size similar to the bounding box of the region.Filled Area is defined as the number of pixels in that particular filled image.

l. Major Axis Length:
The major axis is endpoints of the lengthiest line that is constructed through the image or object.The major axis endpoints are derived by calculating the pixel distance between each grouping of boundary pixels in the object periphery and finding the pair with increased length.Major Axis Length gives the value of the image or object length.

m. Minor Axis Length:
The minor axis endpoint is the longest line that can be constructed through the image or object, whereas remaining perpendicular along with the major axis.Minor Axis Length gives the value of the image or object width.

n. Orientation
The major-axis angle is calculated from the angle between the X-axis of the image and the major axis.It can vary from 0° to 360°.Orientation is defined as the entire direction of the shape of the object in the image.

D. Convolutional Neural Network
The convolutional neural network uses stochastic gradient descent to calculate the minima for the loss function, and it uses the high dimensional dataset to reduce efficiently.Hence it is a greedy technique.It doesn't assure you to find global minima.So the proposed research used an optimization algorithm to minimize the loss function, and it is trained using the back propagation method.The back propagation method initially calculates the dot products of the input signal and its respective weights by activation function to the summation of products.It then transforms the signal to design the complex non-linear functions.The introduced non-linear functions train the design to learn complete random functional mapping.Then back propagation is performed in the system with error terms with updated weight values by Gradient Descent.By applying this, we measure the gradient error, including weight and update in alternate direction of the gradient of the loss function concerning the design parameters.
The extracted features are grouped and passed into the following classifiers

E. Recall
The recall is termed as sensitivity that is the ratio of total quantity of related instances that are truly retrieved.

Fig 2 -
Fig 2 -Architecture of the proposed system
D. PrecisionPrecision is a significant performance metric calculated to find the positive predicted values.It is the fraction of related instances from the extracted cases.From the fig 2 and 3, it is shown that the value of precision is high in the proposed method using a real-time dataset from the image fig 6 and compared to other existing methods such as Face Net, CDA, TDL, KCFT, G-FST, and HOD+SVM by using LFW and ORL dataset.
From the fig 2 and 3, it is shown that the value of recall is high in the proposed method using real-time dataset from the image fig 6 and compared to other existing methods such as Face Net, CDA, TDL, KCFT, G-FST and HOD+SVM by using images from the fig 4 and 5 of ORL and LFW dataset F. F1-Score It is one of the critical performance metrics to test the accuracy of the proposed method.It is shortly defined as the weighted harmonic mean of tested values of recall and precision.From the fig 2 and 3, it is shown that the value of recall is high in the proposed method by using input real-time dataset images from the fig 6.It is compared to other existing methods such as Face Net, CDA, TDL, KCFT, G-FST, and HOD+SVM by using the pictures from the fig 4 and 5 of ORL and LFW dataset.

Fig 2 -Fig 3 -Fig - 4 ORLFig - 6
Fig 2 -Comparison of Performance metrics using the different existing method and proposed method in ORL dataset In similarity-based models, it follows global, local, and Quasi-local approaches by calculating the similarity between nodes.Depending on the structure of the network, a statistical model is designed to calculate the possibility of unnoticed connections to happen by using maximum likelihood models such as Hierarchical structural design and stochastic block model.These techniques are time-consuming and derive an accurate result.In feature learning methods, CNN extract features from the graph to check the connection between linked nodes and extract the data from the topology of the system(El Khiyari & Wechsler, 2017).The awareness of this learning model is to establish the structure of the localized neighbor by disintegrated feature data as an alternative to creating the entire graph by multiple source link prediction of input images.
(Laiadi et al., 2018)method involves face area acquisition, localization, and segmentation to define the facial expression, face shape, and position of components, and the quality of the image(Laiadi et al., 2018).Then feature detection and Point   is mapped into a node   ∈ .The relation between any two points (  ,   ) is represented as an edge   ∈ , and the value is defined as below, Typically, a time series {  } (=1,…,) is mapped into a graph (, ) where a time  = � 1,   <   ∧   <   ∀ ∈ (, ) ∧  <  0, ℎ