- Access to this full-text is provided by Hindawi.
- Learn more

Download available

Content available from Advances in Mathematical Physics

This content is subject to copyright. Terms and conditions apply.

Research Article

High-Capacity Real-Time Face Retrieval Recognition Algorithm

Based on Task Scheduling Model for the Treatment

Area of Hospital

Yi Zhou and Weili Xia

School of Management, Northwestern Polytechnical University, Xi’an 710072, China

Correspondence should be addressed to Yi Zhou; zhouyi1202@mail.nwpu.edu.cn

Received 29 October 2021; Revised 17 November 2021; Accepted 22 November 2021; Published 6 December 2021

Academic Editor: Miaochao Chen

Copyright © 2021 Yi Zhou and Weili Xia. This is an open access article distributed under the Creative Commons Attribution

License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is

properly cited.

This paper presents an in-depth study of face detection, face feature extraction, and face classiﬁcation from three important

components of a high-capacity face recognition system for the treatment area of hospital and a study of a high-capacity real-

time face retrieval and recognition algorithm for the treatment area of hospital based on a task scheduling model. Considering

the real-time nature of our system, our face feature extraction network is modeled by DeepID, and the network is slightly

improved by introducing a central loss veriﬁcation signal to train a DeepID-like network model using central loss and use it to

extract face features. To further investigate and optimize the schedulability analysis problem of the directed graph real-time

task model, this paper proposes a rigorous and approximate response time analysis method for the directed graph real-time

task model with an arbitrary time frame. Based on the theoretical results of the greatly additive algebra, it is shown that the

coherent qualifying function is linearly periodic, i.e., the function can be represented by a ﬁnite nonperiodic part and an

inﬁnitely repeated periodic part, thus calculating the coherent qualifying function independent of the magnitude of the interval

time. The algorithm for high-capacity real-time face retrieval and recognition in the treatment area of hospital based on the

task scheduling model is further investigated, and a face database is established by using the PCA dimensionality reduction

technique. Based on the internal architecture of the processor, image preprocessing and IP core packaging are implemented,

and the hardware engineering of the high-capacity real-time face recognition system for hospital visits is built using the IP-

based design concept. The performance tests of the face detection model and feature extraction network show that the face

detection model has a signiﬁcant reduction in false-positive rate, better ﬁtting of border regression, and improved time

performance. The face feature extraction network has no overﬁtting, and the features are highly discriminative with small

feature extraction time consumption. The high-capacity real-time face recognition system for the treatment area of hospital

combined with the optimized directed graph task scheduling model can approach 25 fps, which meets the real-time

requirements, and the face recognition rate surpasses that of real people. It realizes the intelligence, self-help, and autonomy of

medical services and satisﬁes the medical needs of users in all aspects.

1. Introduction

In the past half-century, biometric identiﬁcation technology

has developed rapidly because each biological individual has

unique biometric points that can be measured, identiﬁed,

and veriﬁed, and biometric identiﬁcation technology iden-

tiﬁes and authenticates individuals based on these unique

points. Unlike traditional identiﬁcation which is easily dam-

aged, easily lost, and easily falsiﬁed, biometric identiﬁcation

technology which utilizes the unique characteristics of bio-

logical individuals is more convenient, fast, safe, reliable,

and accurate. Biometric identiﬁcation technology can be

divided into two categories, one is identiﬁcation by inherent

characteristics of biological individuals, and the other is

identiﬁcation by behavioral characteristics of biologicals [1,

2]. Inherent characteristic identiﬁcation currently has fea-

tures like ﬁngerprint, hand type, iris, retina, pulse, and face

and has been used as biometric identiﬁcation, and

Hindawi

Advances in Mathematical Physics

Volume 2021, Article ID 1547025, 11 pages

https://doi.org/10.1155/2021/1547025

behavioral characteristics like voice and signature have been

used as biometric identiﬁcation. As a kind of biometric tech-

nology, face recognition has its unique advantages such as

nonintrusiveness, convenience, friendliness, noncontact,

and scalability compared with other biometric technologies

in practical use scenarios, which makes it have an important

position in biometric identiﬁcation [3, 4]. If the time con-

straints of the real-time system are met, the system is said

to be schedulable, that is, each task is completed before its

own time limit. Therefore, it is required to be able to deter-

mine in advance whether it is schedulable [5]. If the result is

not schedulable, then you need to increase system resources

or reduce task load so that the system becomes schedulable

again.

Face recognition has become the current research hotspot

in the ﬁeld of artiﬁcial intelligence due to its huge application

prospect. With the development of science and technology

and the progress of society, the technical demand for con-

ducting fast, eﬃcient, and automatic face recognition is

becoming more and more urgent. Face recognition technology

ﬁrst determines whether there is a face inthe image or not, and

if there is a face, the features of the face are extracted and

matched with the known face comparison, and then, the iden-

tity of the face is recognized [6]. The face recognition process

is generally divided into four steps: face detection, face feature

extraction, face matching, and recognition. Among them, face

feature extraction is the most important part of face recogni-

tion, and it is good or bad directly aﬀects the ﬁnal result of

our face recognition. From another perspective, face recogni-

tion is ﬁnding a way to describe faces, and this description

needs to have strong interclass distinguishability and intraclass

stability. There are neural network-based methods, statistical

methods, geometric feature-based methods, model-based

methods, and multiclassiﬁer integration methods for face rec-

ognition, among which the vast majority of traditional face

recognition feature expressions are human selected; however,

in practice, human selection of features is a very time-

consuming and energy-consuming matter, and the good or

bad selection depends largely on experience and luck. How-

ever, with the increasing electronic, automated, and connected

control systems, the functions and performance requirements

undertaken by real-time systems are becoming more and

more demanding. At the same time, the type, number, and a

load of tasks in the system have increased dramatically, which

poses a great challenge to the time veriﬁcation of real-time sys-

tems [7]. Therefore, it is of great signiﬁcance to research sche-

dulability analysis and performance optimization of complex

real-time task systems.

This paper will focus on the direction of the task sched-

uling model for a high-capacity real-time face retrieval rec-

ognition algorithm for hospital visits. The existing

algorithms for large-capacity real-time face retrieval and rec-

ognition inhospital visit sites match sparse datasets, which

makes the large-capacity heterogeneous face recognition

inhospital visit sites cannot achieve the performance of visi-

ble face recognition when using the same task scheduling

model. At the same time, the existing face recognition task

scheduling model is large and cannot meet the needs of oﬀ-

line inference computation of edge devices nowadays. By

investigating the multitask scheduling method to make the

large-capacity real-time face recognition model complete

the task quickly and accurately on the embedded platform

is the purpose of this paper. Chapter 1 is the introduction,

which brieﬂy introduces the background and signiﬁcance

of the study of the real-time face recognition system and

ﬁnally summarizes the main work of this paper and the

structure of the paper. The second chapter is related work,

brieﬂy describes the status of domestic and international

research, and both face detection and feature extraction are

obtained using task scheduling model, and ﬁnally, it also

introduces how to improve task scheduling performance by

adjusting hyperparameters. Chapter 3 ﬁrstly introduces the

research on the recognition algorithm of large-capacity

real-time face retrieval inhospital visit sites based on task

scheduling model; then, we select the current multitask

scheduling model with very high real time and performance,

improve it, and introduce better recognition methods for

large-capacity real-time face retrieval inhospital visit sites.

Chapter 4 is the result analysis, through the hospital visit

place high-capacity real-time face retrieval recognition algo-

rithm scheduling analysis, hospital visit place high-capacity

real-time face retrieval recognition algorithm performance

analysis, and hospital visit place high-capacity real-time face

retrieval recognition algorithm system analysis, which

proves the better performance and eﬃciency of the algo-

rithm researched in this paper, the relevant test of our face

recognition system, and the analysis of our face recognition

system performance and result presentation. Chapter 5 con-

cludes with a summary of the research work in this paper,

pointing out the shortcomings of the scheme, proposing cor-

responding improvement methods, and looking ahead.

2. Related Work

Since individual task scheduling is computationally limited,

most problems require multiple task schedulers to work

together to solve the target problem, hence the emergence

of multitasking scheduling systems. Multitask scheduling

systems have considerable robustness and adaptability and

are more eﬃcient for solving practical application problems.

Addagarla et al. combined binary particle swarm optimiza-

tion with contract networks and used the synergy of task

scheduling to improve the accuracy of high-capacity real-

time face retrieval recognition and reduce the scheduling

time in the treatment area of hospital. The dynamic path

planning method is accomplished by using the dynamic

and reactive nature of task scheduling and adding it to the

ant colony algorithm [8]. Bartolini and Patella incorporated

a microgrid composed of distributed power sources into a

communication network formed by multitask scheduling

technology, which enabled the system to ensure stable sys-

tem power output even when the load changed drastically

[9]. The multitask scheduling technique is used to coordi-

nate and control the large-capacity real-time face retrieval

and recognition in a hospital visit site, and the multitask

scheduling technique is incorporated into it to enhance the

communication function of the large-capacity real-time face

retrieval and recognition in a hospital visit site and to

2 Advances in Mathematical Physics

maintain a trouble-free retrieval and recognition [10]. The

information captured by task scheduling is processed by

applying convolutional fusion. The task scheduling tech-

nique is incorporated into semantic Web services to make

the discovery, invocation, and combination of Web services

more eﬃcient, and the multitask scheduling technique is

introduced into deep reinforcement learning to solve the

problem that the policy gradient increases with the number

of task scheduling [11].

However, the rapid growth of face recognition technol-

ogy in China in recent years and the booming of task sched-

uling have forced major research institutes, companies, and

universities to invest heavily in face recognition [12]. Choi

and Cha proposed a transformation algorithm using Eigen

to reach the transformation from face sketch to face photo

and then perform face matching on the face photo obtained

from the change [13]. A method using locally linear embed-

ding is proposed to transform the face sketch into a real

photo image and the near-infrared light face image into a

visible light face image, respectively, and then use a general

method to match the face recognition [14]. To consider the

relationship between face facial image regions and their

neighboring regions in a comprehensive way, Markov ran-

dom ﬁeld model is added to the project of heterogeneous

face recognition [15]. In this approach, since only the best

candidate regions are selected for training, this leads to the

deformation of facial images. Kumar et al. proposed a Mar-

kov weight ﬁeld model, which can solve the problem of

image deformation by selecting some selected regions to

form a Markov network model [16]. Muhammad et al. pro-

posed a direct push learning method using Markov random

ﬁelds, which combines sketch images with real photo images

of the test images into the learning process, thus reducing

the error of the model on the test data [17].

This paper proposes and studies a task model for

dynamically adjusting the switching speed, which is a strict

extension of the static adaptive rate-of-change task model

[18]. A schedulability analysis method was developed for

such a new real-time task model. This method introduces a

new directed graph task model to safely approximate the

execution load of this type of rotation task. Compared with

the existing static analysis methods, the proposed schedul-

ability analysis method has higher accuracy and can eﬀec-

tively reduce the time complexity [19]. Hospital visit site

high-capacity real-time face retrieval recognition is a process

algorithm used to describe concurrent systems, and its sim-

ple and eﬃcient algorithmic model makes hospital visit site

high-capacity real-time face retrieval recognition the basis

for modeling in several ﬁelds. The CCS is promoted for

high-capacity real-time face retrieval recognition at hospital

sites and can be used to represent communication topologies

with dynamic change capability by transmitting channels

between distributed nodes along other channels, which is

very powerful for formal description. Since a change in the

physical location of processors in a distributed system may

bring about a change in the communication topology, it is

particularly suitable to describe a concurrent programming

model for task scheduling using high-capacity real-time face

retrieval recognition inhospital visit sites. The goal of this

paper is to combine static task scheduling algorithms with

multitask scheduling techniques and high-capacity real-

time face retrieval recognition inhospital visits, so that task

scheduling in a distributed environment can better utilize

environmental resources, not only to enhance concurrency

and reduce communication overhead by using multitask

scheduling techniques but also to combine the simulated

communication capability of high-capacity real-time face

retrieval recognition inhospital visits [20]. It deﬁnes a task

scheduling model for uncertainty-oriented concurrent sys-

tems that reacts faster to network topology changes and is

important for eﬃcient utilization of resources in distributed

systems.

3. Intelligent Analysis of Supply Chain

Coordination Based on the

Internet of Things

3.1. High-Capacity Real-Time Face Information Detection

and Extraction for Hospital Consultation Places. In the face

of multipose face detection task, the organization of the clas-

siﬁer is an important issue, especially when the features of

faces in diﬀerent poses show great diﬀerences, which will

bring a great challenge to the detector; to solve this problem,

a partitioning strategy is usually used, and the classiﬁers are

trained separately for faces in diﬀerent poses and then com-

bined to build a multipose face detector. The partitioning

strategy is the most basic strategy to deal with the multipose

face detection task, but it also needs to take into account the

speed and classiﬁcation accuracy, and the enhancement of

classiﬁcation capability will bring an increase in computa-

tional cost.

We design Equation (1) by combining the localization

error Dc and classiﬁcation error Fc of the samples, ythe pre-

diction probability of classiﬁcation, and xthe intersection

rate between the border and the actual target border and cal-

culate a surrogate value cost for the border containing posi-

tive samples in the grid, based on which we judge whether

the sample is selected to join the training.

Cx,y

ðÞ

=Dc x

ðÞ

+Fc y

ðÞ

:ð1Þ

The positioning error Dc is deﬁned as Equation (2).

Dc x

ðÞ

=1

1 + lim

S⟶∞∑S

i=1ix−t

ðÞ

x+t

ðÞ

:ð2Þ

The classiﬁcation error Fc is deﬁned as Equation (3).

Fc y

ðÞ

=1−y

ðÞ

lim

S⟶∞〠

S

i=1

iy−t

ðÞ

y+t

ðÞ

:ð3Þ

In Equations (2) and (3), tis a certain threshold value of

the intersection rate. iis a coeﬃcient that controls the sensi-

tivity of the localization error Dc to the threshold value t; the

inputs xand yfor both costs take values between 0 and 1.

3Advances in Mathematical Physics

The purpose of DeepID2 for learning is to obtain the

output vector, not to maximize the recognition rate; thus,

the class spacing term is added to the paper. To facilitate

the representation, we express the output vector obtained

by DeepID2 layer extraction as a function, as in Equation

(4), where TðxÞdenotes the extracted DeepID2 feature vec-

tor, the tfunction denotes the feature extraction function, x

denotes the input face slice, and βdenotes the network

parameters.

Tx

ðÞ

=tx,β

ðÞ

:ð4Þ

The function for correct classiﬁcation is the objective

function of softmax, which is aimed at minimizing the

cross-entropy and is deﬁned to identify the signal, as in

Equation (5), where Tis the DeepID2 feature vector of

Equation (4), tdenotes the classiﬁcation target, βidenotes

the softmax layer parameter, pðiÞdenotes the target proba-

bility distribution, and pðtÞdenotes the classiﬁcation predic-

tion probability distribution.

ST,t,βi

ðÞ

= lim

N⟶∞〠

N

i=1

pi

ðÞln pi

ðÞ=−ln pt

ðÞ

:ð5Þ

The objective function for maximizing the sample spac-

ing is as follows, deﬁned as the validation signal, as in Equa-

tion (6), where sij =1 means that fi and fj belong to the

same person, and this is when we need to minimize the

interclass spacing; sij =−1means that Tiand Tjbelong to

diﬀerent people, and this is when we need to minimize the

diﬀerence between mand their distance values. The mis a

parameter that needs to be adjusted manually, and the pur-

pose of proposing mis that the objective function needs to

be minimized, not maximized. Since the validation signal

needs two samples to be calculated, the training process of

the whole network needs to be changed accordingly. Two

samples are randomly selected for each iteration during

training and then trained. Our goal is to learn the parameter

βij updated by stochastic gradient descent.

HT

i,Tj,sij,βij

=

max 0, Ti−Tj

2/2, m−Ti−Tj

2

2:

ð6Þ

3.2. High-Capacity Real-Time Face Retrieval Recognition

Algorithm Based on Task Scheduling Model for the

Treatment Area of Hospital. The multiframe task ican be

described by the minimum release interval Ti, the execution

time vector Ci =ðCi,1;⋯⋯ ;Ci,miÞ, and the relative time

frame Di =Ti. A typical example of a multiframe task model

is an MPEG video stream decoding application using multi-

ple frame types, capable of simulating the periodic arrival of

video frames and diﬀerent frames with diﬀerent decoding

times. As shown in Figure 1, a multiframe task can be repre-

sented by a graph with a chain structure (Figure 1(a)), where

the solid nodes represent the entry points for task execution.

In addition, the execution time of the multiframe task is var-

iable, but the release pattern and the relative time frame are

ﬁxed (Figure 1(b)).

For nindependent multiframe tasks running on a pro-

cessor with a scheduling policy using the preemptive static

priority RM algorithm, the system is schedulable if Equation

(7) is satisﬁed.

Ux

ðÞ

=1+

1

x

1/n

−1

"#

x+n

ðÞ

x−n

ðÞ

:ð7Þ

Each node xi∈X(or task type) is represented by an

ordered pair heðxiÞ,dðxiÞi, where eðxiÞand dðxiÞare the

worst-case execution time and the relative time frame,

respectively. The edges represent the possible control ﬂow

directions, i.e., the order of releasing tasks. An edge ðxi,yjÞ

⊆Kis described by the parameter pðxi,yjÞ, which is the

A

B

C

DD

BDF

CE

AG

Graphical representation of multi-frame tasks

(a)

Release mode of multi-frame tasks

T1,C1 T2,C2 T3,C3

T2,C2 T3,C3T1,C1

Second

frame

First frame ird

frame

Second

frame

ird

frame

First frame

Task model one

Task model two

(b)

Figure 1: Graph representation and release pattern for multiframe tasks.

4 Advances in Mathematical Physics

minimum release interval time for xiand yj. We assume that

all parameters are positive integers, and that the WCET of

each node is not greater than the minimum release interval

of all its outgoing edges, i.e., it satisﬁes Equation (8).

∃xi,yj

⊆K,ex

i

ðÞ

≐px

i,yi

ðÞ

:ð8Þ

For the same path t=fx1,x2⋯⋯xngin task Tfunc-

tions f, there exists a vertical to slope one substitution rela-

tion deﬁned as follows: for any vertical part of fas in

Equation (9).

fT,x

ðÞ

=T∗ft

ðÞ

+ψ:ð9Þ

For task T, the maximum cumulative execution request

of its task release in an interval of arbitrary length tis

deﬁned as its request-limited function, as in Equation (10).

gT,x

ðÞ

= max

β⊆DT

ðÞTf t

ðÞ1−β

ðÞ

1+β

ðÞðÞ

:ð10Þ

For task T, its task release is deﬁned as its coherence-

limited function in terms of the maximum number of execu-

tion loads in an interval of arbitrary length t, as in Equation

(11).

hT,x

ðÞ

= max

β⊆DT

ðÞ Tg t

ðÞ 1−β

ðÞ

1+β

ðÞ

lim

N⟶∞∑N

i=1βi

0

@1

A

0

@1

A:ð11Þ

For task T, whose tasks are released and the time frame

is within interval t, its maximum cumulative execution

demand is deﬁned as its demand-limited function, as in

Equation (12).

yT,x

ðÞ

= max

β⊆DT

ðÞht

ðÞ1−β

ðÞ

1+β

ðÞðÞ

:ð12Þ

Intuitively, the fðÞ and gðÞ functions quantify the maxi-

mum cumulative execution request and load, respectively,

while the yðÞ function measures the size of the execution

time that a task is released and must be completed within

a given interval. Like the fðÞ function, the gðÞ function is a

nondecreasing step function whose horizontal part is left-

open and right-closed. At the same time, the yðÞ function

inherits the properties of yand is also a continuous sloping

step function. Like the gðÞ function, the gðÞ function is a

step function whose horizontal part is left-closed and right-

open.

To reduce the complexity of computation and to achieve

the purpose of real-time processing, dimensionality reduc-

tion methods are used in face recognition. The commonly

used mainstream linear transformation method for dimen-

sionality reduction is principal component analysis (PCA),

which is derived from K-L transformation, and its essence

is to protect the data in the high-dimensional space into

the low-dimensional space through a linear transformation

while representing the original data as well as possible. In

the process of real-time face recognition, assuming that the

fps (frames per second) of the currently captured video is

30 Hz and the resolution of the video image is m. Then, at

a certain time, there will be nimages, and mis the dimen-

sion of each image. That is, in the m-dimensional space,

there exist nsamples as in Equation (13). The dimensional-

ity of the data is too high, it is time-consuming and laborious

to process, so I wonder if I can only process part of the

dimensions, and the results obtained are consistent with

the results of all dimensions. To put it simply, a picture

has 2000 feature dimensions, but in fact, only 100 of them

(or even less) have a huge impact on the result.

M= lim

n⟶∞〠

n

i=1

xi

n

:ð13Þ

Assuming that 1 line is to be used to represent these n

samples and let edenote the unit direction vector of this line

through the mean, the equation of the line can be expressed

as Equation (14), where ais a real scalar representing the

distance of a point on the line from the point m.

dx

ðÞ

=dm

ðÞ

+α∗ψ:ð14Þ

A linear projection of the nsamples in the direction of

the eigenvector corresponding to the maximum eigenvalue

of the walk matrix yields a 1-dimensional representation ia

in the least square error sense, and this projection transfor-

mation is the transformation of the bases. In the m-dimen-

sional space, the mbases are unit vectors ψiin the

direction of each coordinate axis ði=1,2⋯nÞ, and a certain

sample xiðxi1,xi2⋯ximÞin the space can be represented by

this set of bases as Equation (15).

xi= lim

m⟶∞〠

m

i=1

xijψi+ lim

n⟶∞〠

n

j=1

xijψi:ð15Þ

A nonlinear activation function with a segmented linear

activation function (Hard_Swish) is introduced in the Light-

facenet model, which can signiﬁcantly improve the accuracy

of the neural network for classiﬁcation tasks when using

Hard_Swish instead of as ReLU, which is improved from

the Swish nonlinear activation function. Swish is deﬁned in

Equation (16).

Swish = lim

M⟶∞〠

M=mn

i=1

xi−xi−1

ðÞ

2

:ð16Þ

The bottleneck is a bottleneck structure, which can be

used in the network to reduce the number of parameters

and computation of the network with no loss of accuracy.

The structure is shown in Table 1.

In the traditional method, the Euclidean distance is usu-

ally used to determine the identity similarity has a great lim-

itation. The characters classiﬁed by the softmax function will

have two diﬀerent identities that are closer than the same

identity, thus failing to achieve the classiﬁcation eﬀect, and

adding intraclass constraints through the cross-entropy loss

5Advances in Mathematical Physics

function can make the network have a better classiﬁcation

eﬀect, and the improved formula is shown in Equation (17).

Y=ln ψWx

ðÞ β+1

ðÞ

β−1

ðÞ

βlim

n⟶∞∑n

i=1ψWx

ðÞ

+xi

:ð17Þ

By normalizing the weights and features, the prediction

of the network depends only on the learning of the weights

and feature angles ψ. All features are distributed on a hyper-

sphere of radius s. A hyperparameter is added to the cross-

entropy loss function for the edge penalty of the weights

and feature angles ψ. This is done to increase the edge angles

between categories and at the same time make the categories

more compact internally, resulting in a better performance

for facial recognition, as in Equation (18).

Y=ln ψsin xi+T

ðÞ

lim

n⟶∞∑n

i=1ψWx

ðÞ

+xi+ lim

n⟶∞∑n

i=1ψsin xi+T

ðÞ

:ð18Þ

If two task instance sequences have the same task instance

sequence, then as far as the coherence function is concerned,

one task instance sequence with the shortest release interval

will dominate the other task instance sequences. Therefore,

we can divide the release time area of the task instance so that

each node is ﬁxed in each release interval.

3.3. Design of High-Capacity Real-Time Face Retrieval

System for Hospital Consultation Places. The design of the

hospital visit place high-capacity real-time face retrieval sys-

tem architecture is oriented to all users, the overall goal of

the system as the basis of the design, standing in the global

overview perspective of the system’s overall perspective for

the construction of the hospital visit place high-capacity

real-time face retrieval of the business aspects of the needs

and technical aspects of the needs from multiple manage-

ment and constraints, from a comprehensive perspective

requires the construction of a uniﬁed platform, the global

model with uniﬁed speciﬁcations. The system is designed

for a ﬁve-layer structure, as shown in Figure 2.

The planning analysis is mainly carried out in two

aspects: on the one hand, it is data-oriented and analyzes

the current situation of data application for each department

and each data source; the data quality analysis is mainly

focused on the usability of data and the usage of data. Based

on data application status and data quality analysis, the sys-

tem’s shared data model is constructed by combining appli-

cation analysis. From another aspect, planning analysis is

oriented to applications and processes, and planning analy-

sis is conducted for applications, processes, and data

exchange and sharing involving cross-department and

cross-system, and on top of that, an application process

model is constructed. Based on these two aspects, a feasible

planning and analysis report is formed and the outline

design and detailed solution design are written and loaded

into the integration and development and implementation

phase of the project.

The self-built face database contains the face data of 5

people in the lab, in which each person’s face data contains

10 images with diﬀerent angles and each face image has a

pixel size of 112 ∗92. The self-built face database uses 9 face

images of each person as training and 1 face image as testing.

In this paper, the validation of the PCA algorithm dimen-

sionality reduction is implemented by Matlab. Firstly, the

data from the self-built face database is read, and the sample

matrix is generated; secondly, the dimensionality reduction

is implemented by the PCA algorithm, and ﬁnally, the prin-

cipal component faces are displayed. The main parameters

include the number of people read in, the number of samples

nFacePerPerson read in for each person, and the container

FaceContainer to store the sample matrix. In the process of

implementing the PCA algorithm, the diﬃculty is to calcu-

late the sample covariance matrix of eigenvalues and eigen-

vectors. The custom function named CalculationPCA

calculates the eigenvectors and eigenvalues of the matrix by

calling the library function. The input parameters are the

stored sample matrix FaceContainer and the dimension K

to be reduced to; the output parameters are the sample

eigenvector matrix and the principal component vector after

reducing to Kdimensions; ﬁnally, the reduced dimensional

principal component face is displayed. The output of the

reduced dimensional data is saved. Since this paper adopts

the functions of OpenCV to realize the detection and recog-

nition function of the human face, the XML ﬁle is completed

by OpenCV to ensure uniformity.

When a noisy image is acquired, the next step is to process

the image to achieve the desired eﬀect. Second, to increase the

detectability of the object to be detected in the image and to

improve the image quality, grayscale processing, ﬁltering,

noise reduction, and contrast enhancement operations are

required for the acquired image. The ARM architecture has

increased the clock speed by simplifying the instruction set

and also uses pipelining to increase the speed of the processor

instruction stream, but the nature of ARM for image process-

ing is serial, especially for embedded low-power products, due

to power constraints, the clock speed needs to be reduced,

which makes it diﬃcult to ensure the real-time requirements.

This is where the signiﬁcant acceleration of many image pro-

cessing algorithms can be achieved. In this paper, we take

advantage of the high-speed parallelism of FPGAs to perform

grayscale processing, median ﬁltering, and histogram equali-

zation on the acquired images through the PL part of Zynq

and package the module as an IP core.

4. Results and Analysis

4.1. Scheduling Analysis of High-Capacity Real-Time Face

Retrieval Recognition Algorithm for Hospital Consultation

Table 1: Bottleneck structure.

Serial number Input Operate Output

1h×w×dConv 1 × 1 h×w×d

2h×w× Ltd Conv 2 × 2 th × w× Ltd

3h×tw×dConv 3 × 3 th × tw × d

4th × w×dConv 4 × 4 th × tw × Ltd

6 Advances in Mathematical Physics

Places. As shown in Figure 3, the scheduling success rate of

both the grouped moderate algorithm and the improved

algorithm is higher than the saving algorithm in the same

feasibility check window. When the Wnd value is small,

the scheduling success rate of the improved algorithm is

not as good as that of the grouped moderate algorithm, but

as the Wnd value increases, the success rate of the improved

algorithm increases and exceeds that of the grouped

Infrastructure layer

Wechat users

System

administrator Doctor

Hospital decision

makers

Patients Department

Face retrieval

management

Hospital information

management

Face recognition

management

Number source

management

Doctor management Interactive data

management

Face data Permission dataBasic data

Multi-task model neural

network cloud

Task 1

Task 2

Training data

Training data

Scheduling model

User layer

Application layer

Network layer

Data layer

Figure 2: Overall system architecture diagram.

012345678910 11

Real-time acquisition size

300

280

260

240

220

200

Number of successful tasksAlgorithm success rate (%)

88

84

80

76

72

Saving algorithm success task

Group moderate algorithm success task

Improved algorithm success task

Saving algorithm success task

Improved algorithm success task

Moderate grouping algorithm success rate

Figure 3: Comparison of scheduling success rate of diﬀerent scheduling algorithms.

7Advances in Mathematical Physics

moderate algorithm when Wnd is equal to 5. This is because

the improved algorithm needs to group the tasks in the win-

dow, and the grouping algorithm of the improved algorithm

is more sensitive to the size of the feasibility check window

because it takes into account the sharing of all resources of

the tasks compared to the grouping method in the grouping

moderate. When the window value is small, the diﬀerence

between the improvement algorithm and the saving algo-

rithm is not too big, and the eﬀect of grouping will slowly

appear when the window value becomes larger.

It can be seen from Figure 4 that among the processor

utilization rates of several algorithms, the average core utili-

zation rate of the improved algorithm is higher than that of

the other two, but the improved algorithm ﬂuctuates more

sharply than other algorithms. After calculation, the average

utilization rate of the saving algorithm, the appropriate

grouping algorithm, and the improved algorithm is

82.32%, 82.76%, and 86.12%, and the standard deviations

are 2.18, 1.64, and 3.25, respectively. One reason for this sit-

uation is that the improved algorithm will try to put tasks

with a high degree of correlation into the same core as much

as possible, but this will lead to a decrease in the degree of

load balancing.

4.2. Performance Analysis of High-Capacity Real-Time Face

Retrieval Recognition Algorithm for Hospital Consultation

Places. To validate the task scheduling-based high-capacity

real-time face retrieval and recognition algorithm for hospi-

tal visits, the neural network framework is used to imple-

ment the algorithm result analysis, and comparative

experiments are done in the training phase of the network

in terms of network structure, the number of parameters,

running speed, and recognition rate. The training samples

of the data set were divided by 11 : 2 to generate the training

set and test set for the experiments. The comparison results

of the experimental results are shown in Figure 5.

The time required by the task scheduling model-based

high-capacity real-time face retrieval recognition algorithm

and the traditional real-time face retrieval recognition algo-

rithm in the hospital visit site were tested separately. As

shown in Figure 6, it can be seen that the time required by

the traditional real-time face retrieval recognition algorithm

is about three times longer than that required by the task

scheduling model-based high-capacity real-time face

retrieval recognition algorithm for hospital visits, and it

can be seen that the speed of face detection is improved by

the task scheduling model, and the design objective is

achieved.

85

84

83

81

82

86

87

88

02468101214161820

Number of experiments

Processor utilization (%)

Saving algorithm

Moderate grouping algorithm

Improve algorithm

Figure 4: Processor utilization of the algorithm.

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Number of iterations

100

96

94

92

90

88

98

Retrieval and recognition accuracy (%)

Task scheduling face retrieval and recognition algorithm

Traditional face retrieval and recognition algorithm

Figure 5: Algorithm retrieval recognition eﬀect.

8 Advances in Mathematical Physics

4.3. Analysis of High-Capacity Real-Time Face Retrieval

Recognition Algorithm System for Hospital Consultation

Places. The features extracted by the feature extraction net-

work are subjected to SVM classiﬁcation to see the overall

recognition performance of the network, and we perform

two experiments to verify the network performance. We ﬁrst

extract the database of 1000 people and select the database

containing 5 to 100 a priori face images to extract face fea-

tures to train the SVM classiﬁer, leave 5 images for each per-

son as a test to test the classiﬁcation performance, and test

the average value three times, respectively, and the results

are shown in Figure 7. As can be seen in Figure 7, our net-

work recognition performance increases in classiﬁcation

accuracy in the early stage according to the increase of a

priori pictures, the highest reached 98.76%.

In the test set, we use CASIA Face dataset for testing, and

the network is based on task scheduling to do migration

learning. This experiment will be done by covering more

combinations randomly and uniformly before the training

starts, and the anchor points can be NIR or visible images.

The positive samples have the same id as the anchor points,

and the subsamples have diﬀerent ids from the anchor

points, according to which about 160,000 sets of training

data can be obtained. Using these data, the trained visible

face recognition network is set up as a three-channel net-

work structure, and the network is trained by a triple loss

function, and the batch size is set to 256 and goes in for

training. The experimental results are shown in Figure 8.

From Figure 8, it can be seen that the scheme in this paper

has surpassed the existing mainstream algorithmic schemes.

When using multitask scheduling as the base network for

training, the recognition rate in the CASIA Face dataset

reaches a level of 99.2% and the veriﬁcation rate reaches

98.9%. When using the traditional method as the base net-

work, the recognition rate reaches 98.8% and the veriﬁcation

rate reaches 98.5% in the dataset. The task scheduling net-

work structure in this paper achieves similar performance

while using a fewer number of parameters and computations

than a standard convolutional network such as the tradi-

tional network. This results in a high-capacity real-time face

recognition network for hospital visits, which illustrates the

method of pretraining a perfect face recognition model by

task scheduling the face dataset and then using a triple loss

70

60

50

40

30

10

20

70

80

60

50

40

30

10

20

0

Retrieval recognition time (ms)Retrieval recognition time (ms)

123

4

5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Retrieval recognition times

Task scheduling face retrieval and recognition algorithm

Traditional face retrieval and recognition algorithm

Task scheduling face retrieval and recognition algorithm

Traditional face retrieval and recognition algorithm

Figure 6: Time required for face detection.

99.0

98.5

98.0

97.5

97.0

96.5

96.0

0 102030405060708090100

Number of face images

1.0

0.6

0.4

0.2

0.0

0.8

94.0

93.5

93.0

92.5

92.0

91.5

91.0

90.5

90.0

Task scheduling face retrieval and recognition algorithm (%)

Traditional face retrieval and recognition algorithm (%)

Figure 7: System recognition performance.

9Advances in Mathematical Physics

function with a suitable triplet for small-scale datasets for

migration learning, which has a good eﬀect for learning

small-scale datasets.

5. Conclusion

This paper investigates the task scheduling-based high-

capacity real-time face retrieval recognition algorithm for

hospital visit places, which is not convenient for the equip-

ment due to a large number of scheduling parameters. The

algorithm of large-capacity real-time face retrieval and rec-

ognition for hospital consultation places based on task

scheduling is proposed. Firstly, the weaknesses and short-

comings of the existing commonly used algorithms in doing

face recognition are analyzed, and they are improved in

terms of the multitask scheduling network structure. A mul-

titask scheduling approach is proposed to replace the exist-

ing high-capacity real-time face retrieval inhospital

consultation places to recognize facial information with dif-

ferent degrees of importance in diﬀerent locations. For the

triple loss function training method, the impact of poor

training data selection and construction method on neural

network training is analyzed, and a reasonable method of

training data triad selection and construction is proposed.

A reasonable training data triad selection and construction

method are proposed to solve the problem that the network

cannot converge quickly due to the random selection of data

to build the training set and then do migration learning with

the triple angular loss function to improve the training eﬀect

of the network. Further, the acceleration factors between

RPM divisions with diﬀerent granularity are investigated

and abounded upper bound on the acceleration factors of

two eﬀective RPM divisions with subset relations which are

given to demonstrate the bounded pessimism of the pro-

posed analysis method. The core of the algorithm is the

use of hybrid schedulability analysis techniques with diﬀer-

ent accuracy and complexity. Experimental results show that

the optimization method proposed in this chapter provides

better and near-upper bound performance while signiﬁ-

cantly reducing the running time of the algorithm compared

to the best available related research results. Although the

system designed in this paper improves the speed of face

detection and combines soft and hard codesign to accom-

plish the face recognition function, there are still some short-

comings. Firstly, it is diﬃcult to improve the recognition rate

of face recognition by relying on the simple task scheduling

algorithm, and the SVM algorithm will be added in the next

work to improve the recognition rate of face recognition.

These shortcomings will be improved in future work and

practice.

Data Availability

The data used to support the ﬁndings of this study are avail-

able from the corresponding author upon request.

Conflicts of Interest

The authors declare that they have no known competing

ﬁnancial interests or personal relationships that could have

appeared to inﬂuence the work reported in this paper.

References

[1] P. Porambage, J. Okwuibe, M. Liyanage, M. Ylianttila, and

T. Taleb, “Survey on multi-access edge computing for internet

of things realization,”IEEE Communications Surveys & Tuto-

rials, vol. 20, no. 4, pp. 2961–2991, 2018.

[2] Z. H. Ali and H. A. Ali, “Towards sustainable smart IoT appli-

cations architectural elements and design: opportunities, chal-

lenges, and open directions,”The Journal of Supercomputing,

vol. 77, no. 6, pp. 5668–5725, 2021.

[3] E. Indrawan, N. Jalinus, and S. Syahril, “Project based learning

in vocational technology education study of literature,”Inter-

national Journal of Scientiﬁc & Technology Research, vol. 9,

no. 2, pp. 2821–2825, 2020.

[4] S. Abirami and P. Chitra, “Energy-eﬃcient edge based real-

time healthcare support system,”Advances in Computers,

vol. 117, no. 1, pp. 339–368, 2020.

[5] W. M. Alenazy and A. S. Alqahtani, “Gravitational search

algorithm based optimized deep learning model with diverse

set of features for facial expression recognition,”Journal of

Ambient Intelligence and Humanized Computing, vol. 12,

no. 2, pp. 1631–1646, 2021.

[6] X. Zhang, W. Huang, X. Lin, L. Jiang, Y. Wu, and C. Wu,

“Complex image recognition algorithm based on immune ran-

dom forest model,”Soft Computing, vol. 24, no. 16, pp. 12643–

12657, 2020.

[7] S. Baek, “System integration for predictive process adjustment

and cloud computing-based real-time condition monitoring of

vibration sensor signals in automated storage and retrieval sys-

tems,”The International Journal of Advanced Manufacturing

Technology, vol. 113, no. 3-4, pp. 955–966, 2021.

[8] S. K. Addagarla, G. K. Chakravarthi, and P. Anitha, “Real time

multi-scale facial mask detection and classiﬁcation using deep

transfer learning techniques,”International Journal, vol. 9,

no. 4, pp. 4402–4408, 2020.

[9] I. Bartolini and M. Patella, “A general framework for real-time

analysis of massive multimedia streams,”Multimedia Systems,

vol. 24, no. 4, pp. 391–406, 2018.

[10] F. M. Talaat, S. H. Ali, A. I. Saleh, and H. A. Ali, “Eﬀective load

balancing strategy (ELBS) for real-time fog computing

180

180

160

160

140

140

120

120

100

100

80

80

60

60

40

40

20

20

092

91

93

94

95 80 81

82

83

84

85

86

87

88

89

90

Traditional system algorithm

e system algorithm of this paper

Figure 8: Face recognition experiment results.

10 Advances in Mathematical Physics

environment using fuzzy and probabilistic neural networks,”

Journal of Network and Systems Management, vol. 27, no. 4,

pp. 883–929, 2019.

[11] Y. Xu, S. Wu, M. Wang, and Y. Zou, “Design and implementa-

tion of distributed RSA algorithm based on Hadoop,”Journal

of Ambient Intelligence and Humanized Computing, vol. 11,

no. 3, pp. 1047–1053, 2020.

[12] Q. Zhang, H. Sun, X. Wu, and H. Zhong, “Edge video analytics

for public safety: a review,”Proceedings of the IEEE, vol. 107,

no. 8, pp. 1675–1696, 2019.

[13] W. Choi and Y. J. Cha, “SDDNet: real-time crack segmenta-

tion,”IEEE Transactions on Industrial Electronics, vol. 67,

no. 9, pp. 8016–8025, 2019.

[14] L. J. Rubini and E. Perumal, “Hybrid kernel support vector

machine classiﬁer and grey wolf optimization algorithm based

intelligent classiﬁcation algorithm for chronic kidney disease,”

Journal of Medical Imaging and Health Informatics, vol. 10,

no. 10, pp. 2297–2307, 2020.

[15] C. Deng, R. Guo, C. Liu, R. Y. Zhong, and X. Xu, “Data clean-

sing for energy-saving: a case of cyber-physical machine tools

health monitoring system,”International Journal of Produc-

tion Research, vol. 56, no. 1-2, pp. 1000–1015, 2018.

[16] S. N. Kumar, A. L. Fred, and P. S. Varghese, “An overview of

segmentation algorithms for the analysis of anomalies on med-

ical images,”Journal of Intelligent Systems, vol. 29, no. 1,

pp. 612–625, 2018.

[17] G. Muhammad, M. F. Alhamid, and X. Long, “Computing and

processing on the edge: smart pathology detection for con-

nected healthcare,”IEEE Network, vol. 33, no. 6, pp. 44–49,

2019.

[18] S. Chen, Z. Zhang, J. Yang et al., “Fangcang shelter hospitals: a

novel concept for responding to public health emergencies,”

The Lancet, vol. 395, no. 10232, pp. 1305–1314, 2020.

[19] S. Jain, R. Khera, Z. Lin, J. S. Ross, and H. M. Krumholz,

“Availability of telemedicine services across hospitals in the

United States in 2018: a cross-sectional study,”Annals of Inter-

nal Medicine, vol. 173, no. 6, pp. 503–505, 2020.

[20] P. Yang, K. M. Treurniet, L. Zhang et al., “Direct intra-arterial

thrombectomy in order to revascularize AIS patients with

large vessel occlusion eﬃciently in Chinese tertiary hospitals:

a multicenter randomized clinical trial (DIRECT-MT)—pro-

tocol,”International Journal of Stroke, vol. 15, no. 6,

pp. 689–698, 2020.

11Advances in Mathematical Physics