Automated Attendance System Based on Face Recognition & Live Video Processing

Automated Attendance System Based on Face Recognition & Live Video Processing

Varun Gopinatp

Student, Dept. of Computer Science & Engineering Mangalam College of Engineering

Kottayam, India

Rakhi R Nair2

Student, Dept. of Computer Science & Engineering Mangalam College of Engineering

Kottayam, India

Meharban3

Student, Dept. of Computer Science & Engineering Mangalam College of Engineering

Kottayam, India

Divya S. B5

Lekshmipriya M. B4

Student, Dept. of Computer Science & Engineering Mangalam College of Engineering

Kottayam, India

Assistant Professor, Dept. of Computer Science & Engineering Mangalam College of Engineering

Kottayam, India

Abstract:- Tremendous growth in the Face Recognition Technology paved a new way for making the normal life easier and more secure. Face Recognition can be used as a Security & Privacy authentication factor and at the same time, it can be used as an identification methodology. Here, an automated attendance marking system based on face recognition technology has been made with the aid of live video processing. The video will be captured using a camera module and that live video will get processed at the same time. The processing means nothing but searching for a human face in the captured video. If there is a human face in the live video, that face will be detected by the face detection module and will be given to the face recognition module. The face recognition module is connected to a database which contains multiple images of a persons face with various facial expressions. Each face in the database has a unique identification number. After the face has been detected, it will be compared with each and every face in the database system. If the detected face is present in the database, the unique value of the image will be popped out and the name of the person whose face which is identified will be marked. This popped out name will be given to the attendance marking module so that the attendance can be marked real time.

Keywords Face Recognition, Linear Discriminant Analysis, Face Identification, Attendance System

1. INTRODUCTION

   The objective is to build an automatic face recognition attendance system using machine learning method. The detection & recognition of face will be carried out using an algorithm called as Fisher LDA (Fisher Linear Discriminant Analysis). These process are trained using Haar Cascade implementation. Haar Cascade implementation is a machine learning methodology with which both the positive and negative images are gets used to train the identifier aka classifier module or system. The positive image consists of the images that the system wants to detect and recognize. That is in this case, the positive images consists of human faces that are to be detected by

   the system. But in the case of negative images, it consists of the images that the system wants to be neglected, that is faces other than human faces. Images of various animals, various buildings, trees etc. can be included in this. The main aim of haar cascade implementation is to train the system to easily distinguish between a human and non- human face. Thus by the accuracy of the system will get increased and the error rate will be diminished. Face Recognition is normally used for security purposes. Here it is used as an identification & recognition method. The scope of the system is that, the time and effort for taking attendance in conventional mode can be saved using this method. A criterion can be set for attendance marking, for e.g., the person who attended the session more than 50 minutes will get attendance, so by this way, the mentality of class bunking can be reduced.

2. RELATED WORKS

   The Face Recognition Techniques are divided into 4 categories

   • Knowledge Based Methods

   • Feature Invariant Methods

   • Template Matching Methods

   • Appearance Based Methods

   All the above described technologies has both efficiency & deficiency. The techniques are easier to implement and on the same time they are much more susceptible to error. Detection of face with various facial expressions are also difficult by using the above mentioned methodologies [4][9].

   Longitudinal Study & Analysis of Permanence Property provides stability to the system during overall analysis phase. The projected output using the above mentioned methodology is less error susceptible. The major drawback is that this methodology is much more complex to implement and the magnitude of the performance of the face recognition system is based on the amount of longitudinal analysis gets performed [5].

   Superposed Linear Representation Classifier (SLRC) is another technology used for face detection. This method successfully overlays the advantages of the robustness of the collaborative representation which makes the system much more user friendly and easier to amend. Processing of uncontrolled and underdamped datasets, CR Based methodologies will result in the chance of occurring more error [7] [10].

   While using Convolutional Neural Networks as training data, the face recognition system acts in a different dimension. It is common to occur image blur in the captured video or image. Here by using CNN, the image blur is treated by applying artificial blurriness to make up for the shortage of real world video training data using CNN training data. The above mentioned process can only be implemented as a single CNN Model. To implement in large scale multiple number of single CNN must be needed to concatenate together which results in the increase in cost and time of calculation. It also increases the complexity of the overall system [1].

   Another methodology to implement face recognition is by using Helmert Contrast. Here signers are much less capable of identifying the human face and recognizing it by compared with non-signers. Response time varies with the age, gender and facial expression of the collected dataset [3].

   Heterogeneous Joint Bayesian (HJB) is the much more advanced technology in the field of face recognition while by compared to all the former defined methodologies. Here Separation of the Equivalent/Diverse Face Matches Precise is done more rapidly. The visible drawback of the system is that the procedure of implementation is very high and small mistakes leads to overall damage of the output [2].

   The metric calculations are normally expressed as vectors in Proximity Based Clustering. Time taken to complete the process is high because the vector input data is to be converted to scalar data to get processed and again needed to be converted to vector form [6].

3. PROPOSED METHODOLOGY

   Face Recognition can be performed in so many ways that each way gives each result accuracy. The proposed method uses a methodology called as Fisher Linear Discriminant Analysis along with Principal Component Analysis. The system is classified into 5 Modules.

   1. MODULES

      Camera

      Camera

      1. Video Capturing Module

         Memory Area

         Fig. 1. Video Capturing Module

         Video Capturing Module consists of only 2 items. Camera used to capture the data or video & Memory Device to store the captured data. Whenever the camera turns on, the memory device also gets turned on and the simultaneous storage will gets takes place. The video

         stored inside the memory area will be used for frame by frame reading and image pre-processing.

      2. Image Pre-processing Module

         Image Pre-processing improves the data tored in the image. Improvement doesnt means increase in data, but it reduces the data distortion of the image. In the proposed system, image pre-processing consists of the 2 steps: Image Crop and Resize & Gabor Wavelet Calculation.

      3. Training Module

         All Images are pre-processed so that the obtained data can be used for creating tensor. The tensor is a 3D representation which describes the relationship between various sets to a vector space. The created tensor is then given to perform Principal Component Analysis (PCA). The result of PCA is then used to create various projections of face. The projection is created by using various Color lines for each expression in order easily distinguish between them. A Class Mean Matrix and Global Mean is created so that the further step of creating a Scatter Matrix will gets boosted.

         Scatter Matrix is mainly created when the covariance matrix calculation is much harder or too costly to calculate. Scatter Matrix plays a huge role in the process of dimensionality reduction. Corresponding Fisher Face is gets calculated as the next step by which the scatter matrix as input. Fisher Face has superiority over Eigen faces because of the effort in maximise the separation between various domains or classes in the training pairs. Recognizer function has been derived from the calculated Fisher Face and this function is used for comparing the unique id with the actual inputs.

      4. Registration Module

         Load Image from Dataset

         Face Detection & Cropping

         Pre-process Faces

         Load Image from Dataset

         Face Detection & Cropping

         Pre-process Faces

         Save Representation

         Create LDA Representation

         Save Representation

         Create LDA Representation

         Fig. 2. Registration Module

         The primary aim of registration module is to make the system able to train the datasets or the training sets. Initially, the images resides in the databases is gets retrieved using SQL queries. The retrieved images are the cropped and given for haar cascade implementation. This is for detecting a human face [8]. After detecting human face, the image is gets pre-processed as described in the pre- processing module. Pre-processed image is then trans- formed to some mathematical expressions based on the data obtained from the image. These mathematical data is given to LDA for Fisher Face creation. After creating the fisher face, corresponding unique ID is created so that this unique id is then compared with other unique id which is the final step for obtaining the persons identity.

      5. Attendance Marking Module

         Load Image from Dataset

         Face Detection & Cropping

         Pre-process Faces

         Load Image from Dataset

         Face Detection & Cropping

         Pre-process Faces

         Compare Representation

         Load Face Representation

         Create LDA Representation

         Compare Representation

         Load Face Representation

         Create LDA Representation

         Predict Person

         Load Face Representation

         Predict Person

         Load Face Representation

         Fig.3. Attendance Marking Module

         In the case of attendance module, live video is gets captured and frame by frame image is gets read out from the video. This frame by frame video is then cropped and pre-processed and corresponding fisher faces are calculated. As the next phase, the calculated fisher representations are loaded and compare with the already stored fisher face values. Based on this comparison results, predictions are happened. That is, when 2 values are gets compared and shown that they are same, details of the corresponding person is retrieved from the database and this details are used to mark the attendance.

         Face Detection using Haar Cascade

         Face Detection using Haar Cascade

         Crop & Resize Image

         Crop & Resize Image

   2. SYSTEM ARCHITECTURE

      Database

      PCA

      Create Tensor

      Create Gabor Wavelet

      PCA

      Create Tensor

      Create Gabor Wavelet

      Fisher Face Creation

      Save Tensor

      Comparison

      Fisher Face Creation

      Save Tensor

      Comparison

      Face Detection Using Haar Cascade

      Crop & Resize Image

      PCA

      Face Detection Using Haar Cascade

      Crop & Resize Image

      PCA

      Live Video

      Live Video

      Fig. 4. System Architecture

      The Above Diagram shows us the actual architectural explanation of the proposed system. The combination of 2

      • Dataset (Database) Stores data that are extracted using feature extraction technology. These Data is used to characterise the face.

      • Face Detection using Haar cascade Implementation Both Positive & Negative Images are trained so that the face detection process will be much more efficient and error free.

      • Crop & Resize Face This phase reduces the size of the image and make the face recognition process much more concentrated to the needed areas.

      • Create Gabor Wavelet Using Gabor Feature Extraction, features are directly extracted from the Gray-scale image.

      • Create Tensor Create a Rank 1 Tensor (Vector Tensor)

      • PCA the Tensor Principal Component Analysis is the process of reducing the dimension. That is the dimensionality reduction is performed so that as much as the information is retained on the image or dataset. The Tensor gets PC Analysed

      • Fisher Face Creator Face Detection is performed using the fisher algorithm so that the efficiency will be high and easier to make separation between classes during training.

      • Create & Save Tensor The New values are inserted into a tensor and this tensor is saved for future usage.

      1. Haar Cascade Implementation for Face Detection Haar Cascade implementation is a ML method used to learn and train systems. In Haar Cascade implementation, both the positive and negative image are used to train the system. The Positive image consists of the images that we want to detect. That is in the case of human face detection, positive images consists of human faces along with some random images. In the case of negative images, it doesnt contains the type of images we want to detect. Haar Cascade Implementation makes the system much more accurate in predicting the output. This is because, the system gets much more calibre to distinguish between the training set which is necessary and not necessary. In other words, the capability of detecting wanted images from unwanted images gets increased using haar cascade

         implementation.

      2. Gabor Wavelet Creation

         2

         2

         Face detection is quite possibly the main utilizations of Gabor wavelets. The face image is convolved with a bunch of Gabor wavelets and the subsequent pictures are additionally handled for acknowledgment reason. The Gabor wavelets are typically called Gabor channels or Gabor Filters in the extent of utilizations. The major reason for using Gabor wavelet is because it reduces the amount of standard deviation with respect to its time and frequency values. The 1-D equation used to calculate the Gabor wavelet is

         architecture forms a single structure to identify and mark

         f(x)= e

         -(x-x0)2/a

         /e-ik0(x-x0)

         (1)

         the persons attendance. Following are the steps performed inside the system.

      3. Fisher Face Calculation

         Fisher Face is a mathematical modelling of images. After the tensor is gets PC analysed, the next step is to

         create a fisher face based on the obtained data. Fisher Face has much advantage over Eigen Face because, Fisher face is less susceptible to error & also because of the effort to maximize the separation between various classes in the dataset over the time of training.

      4. Principal Component Analysis

         Principal Component Analysis is a dimensionality reduction process with hich the larger size sets have been reduced to lower size sets with retaining the maximum amount of data it can hold. Sometimes the dimensionality reduction may results in the accuracy of the output. By compared to larger sets, smaller sets are always comfortable to scan, traverse and project the data values. So the process of the dimensionality reduction results in making the system less time consuming. It is also proved that for machine learning based calculations, it is always better to use smaller sets, which is dimensionally reduced sets. There are 5 steps for explaining the concept of PCA. They are,

         • Standardization

         • Covariance Matrix Computation

         • Eigen Value & Eigen Vector to calculate the principal component

         • Feature Vector

         • Recast The Data Along The Principal Components Axes

4. RESULT & PERFORMANCE ANALYSIS

   As taking the whole output response as reference, the proposed system produce much more accurate output than any other existing face recognition system. Fisher Linear Discriminant Analysis combined with Principal Component Analysis makes the system less susceptible to error and more user friendly. Compared with the existing systems, proposed system possess much more stability while performing the operations, that is the system is stable and there is only less chance of having error. Amount of junk value generation is reduced to a certain level, so the output will not be affected with junk. Occluded Face Recognition can be performed by using the proposed system with 80% more accuracy than any other system. Computational Time has been reduced drastically so that all the operations are done rapidly.

   Performance Analysis Chart

   TABLE 1. DATA TABLE

   Test Set Count	Accuracy
   	SLRC	Helmert Contrast	Fisher LDA
   10 ~ 20	0.79	0.79	0.87
   20 ~ 30	0.78	0.73	0.85
   30 ~ 40	0.76	0.73	0.85
   40 ~ 50	0.73	0.73	0.84
   50 ~ 60	0.73	0.73	0.81

5. CONCLUSION

   Face recognition can be used as a security method. But here it act as an identification factor. The proposed system provides an output that is much more accurate than any other system. Unlike all the systems, the proposed system has less mathematical processing background, so it is much easier to implement and easier to amend. Accurate face recognition is done so that no unwanted or misleading prediction will happen and the system will always remain error free.

6. REFERENCES

1. C. Ding and D. Tao, Trunk-branch ensemble convolutional neural networks for video-based face recognition, IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 4, pp. 10021014, Apr. 2018

2. C. Stoll, R. Palluel-Germain, R. Caldara, J. Lao, M. W. G. Dye, F. Aptel, and O. Pascalis, Face recognition is shaped by the use of sign language, J. Deaf Stud. Deaf Educ., vol. 23, no. 1, pp. 19, 2018

3. H. Shi, X. Wang, D. Yi, Z. Lei, X. Zhu, and S. Z. Li, Cross- modality face recognition via heterogeneous joint Bayesian, IEEE Signal Process. Lett., vol. 24, no. 1, pp. 8185, Jan. 2017

4. K. Solanki and P. Pittalia, Review of face recognition techniques,

   Int. J. Comput. Appl., vol. 133, no. 12, pp. 2024, Jan. 2016

5. L. Best-Rowden and A. K. Jain, Longitudinal study of automatic face recognition, IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 1, pp. 148162, Jan. 2018

6. V. B. Nemirovskiy, A. K. Stoyanov, and D. S. Goremykina, Face recognition based on the proximity measure clustering, Inst.

   Cybern. Tomsk Polytech. Univ., vol. 40, no. 5, pp. 740745, 2016

7. W. Deng, J. Hu, and J. Guo, Face recognition via collaborative representation: Its discriminant nature and superposed representation, IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 10, pp. 25132521, Oct. 2018

8. K. Taniya, M. Nidhi, and T. Nandini, Automated human resource and attendance management system based on real time face recognition, IJSRSET, vol. 16, no. 4, pp. 847853, 2016

9. D. Wu, Y. Tang, G. Lin, and H. Hu, Roboust face recognition based on significance local directional pattern and deep learning, J. Optoelectron. Laser, vol. 27, no. 6, pp. 655661, 2016.

10. Y. Sun, J. Zhao, and Y. Hu, Supervised sparsity preserving projections for face recognition, Proc. SPIE, vol. 8009, no. 4, pp. 357366, 2017.

0.9

0.85

ACCURACY

ACCURACY

0.8

0.75

0.7

0.65

10 20 30 40 50

TEST SET COUNT

SLRC Helmert Contrast Fisher LDA

Fig. 5. Performance Analysis Based on Accuracy