Loading [MathJax]/jax/output/SVG/jax.js
Research article Special Issues

Enhancing facial recognition accuracy through multi-scale feature fusion and spatial attention mechanisms

  • Received: 15 February 2024 Revised: 02 March 2024 Accepted: 14 March 2024 Published: 21 March 2024
  • Nowadays, advancements in facial recognition technology necessitate robust solutions to address challenges in real-world scenarios, including lighting variations and facial position discrepancies. We introduce a novel deep neural network framework that significantly enhances facial recognition accuracy through multi-scale feature fusion and spatial attention mechanisms. Leveraging techniques from FaceNet and incorporating atrous spatial pyramid pooling and squeeze-excitation modules, our approach achieves superior accuracy, surpassing 99% even under challenging conditions. Through meticulous experimentation and ablation studies, we demonstrate the efficacy of each component, highlighting notable improvements in noise resilience and recall rates. Moreover, the introduction of the Feature Generative Spatial Attention Adversarial Network (FFSSA-GAN) model further advances the field, exhibiting exceptional performance across various domains and datasets. Looking forward, our research emphasizes the importance of ethical considerations and transparent methodologies in facial recognition technology, paving the way for responsible deployment and widespread adoption in the security, healthcare, and retail industries.

    Citation: Muhammad Ahmad Nawaz Ul Ghani, Kun She, Muhammad Usman Saeed, Naila Latif. Enhancing facial recognition accuracy through multi-scale feature fusion and spatial attention mechanisms[J]. Electronic Research Archive, 2024, 32(4): 2267-2285. doi: 10.3934/era.2024103

    Related Papers:

    [1] Fei Hou, Bin Chen . On triple correlation sums of Fourier coefficients of cusp forms. AIMS Mathematics, 2022, 7(10): 19359-19371. doi: 10.3934/math.20221063
    [2] Qiuxia Hu, Bilal Khan, Serkan Araci, Mehmet Acikgoz . New double-sum expansions for certain Mock theta functions. AIMS Mathematics, 2022, 7(9): 17225-17235. doi: 10.3934/math.2022948
    [3] Guangwei Hu, Huixue Lao, Huimin Pan . High power sums of Fourier coefficients of holomorphic cusp forms and their applications. AIMS Mathematics, 2024, 9(9): 25166-25183. doi: 10.3934/math.20241227
    [4] Zhichao Tang, Xiang Fan . Ternary cyclotomic numbers and ternary Jacobi sums. AIMS Mathematics, 2024, 9(10): 26557-26578. doi: 10.3934/math.20241292
    [5] Kwang-Wu Chen, Fu-Yao Yang . Infinite series involving harmonic numbers and reciprocal of binomial coefficients. AIMS Mathematics, 2024, 9(7): 16885-16900. doi: 10.3934/math.2024820
    [6] S. Akansha, Aditya Subramaniam . Exploring Chebyshev polynomial approximations: Error estimates for functions of bounded variation. AIMS Mathematics, 2025, 10(4): 8688-8706. doi: 10.3934/math.2025398
    [7] Lijun Hao . New proofs for three identities of seventh order mock theta functions. AIMS Mathematics, 2023, 8(2): 4806-4813. doi: 10.3934/math.2023238
    [8] Hasanen A. Hammad, Manuel De la Sen . Tripled fixed point techniques for solving system of tripled-fractional differential equations. AIMS Mathematics, 2021, 6(3): 2330-2343. doi: 10.3934/math.2021141
    [9] Jinliang Wang, Fang Wang, Songbo Hu . On asymptotic correlation coefficient for some order statistics. AIMS Mathematics, 2023, 8(3): 6763-6776. doi: 10.3934/math.2023344
    [10] Xiuhai Fei, Zhonghua Wang, Cuixian Lu, Haifang Zhang . Higher Jordan triple derivations on $ * $-type trivial extension algebras. AIMS Mathematics, 2024, 9(3): 6933-6950. doi: 10.3934/math.2024338
  • Nowadays, advancements in facial recognition technology necessitate robust solutions to address challenges in real-world scenarios, including lighting variations and facial position discrepancies. We introduce a novel deep neural network framework that significantly enhances facial recognition accuracy through multi-scale feature fusion and spatial attention mechanisms. Leveraging techniques from FaceNet and incorporating atrous spatial pyramid pooling and squeeze-excitation modules, our approach achieves superior accuracy, surpassing 99% even under challenging conditions. Through meticulous experimentation and ablation studies, we demonstrate the efficacy of each component, highlighting notable improvements in noise resilience and recall rates. Moreover, the introduction of the Feature Generative Spatial Attention Adversarial Network (FFSSA-GAN) model further advances the field, exhibiting exceptional performance across various domains and datasets. Looking forward, our research emphasizes the importance of ethical considerations and transparent methodologies in facial recognition technology, paving the way for responsible deployment and widespread adoption in the security, healthcare, and retail industries.



    In today's world of industrial pollution, toxicants are pervading the air, ecological problems have become increasingly prominent, and environmental pollution has become a major problem. When human activities expose in the wild, they may come into contact with wild animals. In the process, wild animals can easily transmit viruses they carry to humans. In fact, most new infectious diseases come from wild animals [1]. SARS, Ebola virus, AIV, H1N1 influenza, and COVID-19 are threatening the ecological balance as well as the survival of human beings and other creatures. A large amount of toxic and harmful substances are discharged into the atmosphere, and seriously affect the environmental quality. It is necessary to study the effects of toxicants on the ecosystem. Hallam et al. proposed using a dynamic methodology to examine ecotoxicology in [2,3,4]. They established a model of the interaction between toxicants and population, and provided sufficient conditions for the persistence and extinction of a population stressed by a toxicant. Researchers have been studying ecotoxicology since the 1980s, and a large amount of literature has been devoted to problems in the area [5,6,7,8,9,10,11,12]. However, size-structured factor has not been considered in these models. Size here refers to some continuous indices related to individuals in the given population, such as volume, maturity, diameter, length, mass, or other quantities that show its physiological or statistical characteristics.

    The effects of environmental pollution on biological population, the dynamical behavioral analysis of ecosystem models, and the control problem have attracted the attention of many scholars [13,14,15]. For many populations, size structure is more appropriate to describe the dynamical evolution than age structure, especially for plants and fishes [16,17]. Population models with age structure have been extensively investigated by many authors as seen in [13] and the references therein. On the other hand, the control problem with size structure has achieved remarkable results through theoretical analysis, numerical calculations, and experimental methods, such as in [14,15,18,19,20,21,22,23,24,25,26]. However, most of these studies have focused on a single species, and only a few have examined interactions among species. Among them, the optimal birth problem has also discussed in detail in [14]. In addition, Hritonenko et al. [15] have established a sized-structured forest system, where the objective function includes the net benefits from timber production and carbon sequestration. Liu et al. [18] have studied the least cost-size problem and the least cost-derivation problem for a nonlinear size-structured vermin population model with separable mortality rate, which takes fertility rate as the control variable. We also mention that Li et al. [21] have considered the optimal harvesting for a size-stage-structured population model. For other types of optimal harvesting problems, refer to [15,22,24,25]. Moreover, the influence of seasonal changes and other factors, the living environment of populations often undergoes periodic changes. Research on optimal harvesting problems dependent on the model of individual size in a periodic environment has been reported in [26,27]. In [27], Zhang et al. have discussed the optimal harvesting in a periodic food chain model by using the size structure of predators. To the best of our knowledge, few studies to date have examined optimal control problems of size-dependent population models and periodic effects in a polluted environment. Inspired by the above work, this paper discusses optimal harvesting for a periodic, competing system that is dependent on size structure in a polluted environment.

    The remainder of this paper is organized as follows: In Section 2, we describe a population model with size structure in a polluted environment and its well-posedness is proved in Section 3. The optimality conditions are established in Section 4. The existence of a unique optimal control pair is obtained in Section 5. Some numerical results are presented in Section 6. At the end of this paper, some brief conclusions are provided.

    In [2,3,4], Hallam et al. proposed the following dynamic population model with toxicant effects:

    {dxdt=x[r0r1C0fx],dC0dt=kCEgC0mC0,dCEdt=k1CEx+g1C0xhCE+u, (2.1)

    where x=x(t) is the population biomass at time t; C0=C0(t) is the concentration of toxicants in the organism at time t; CE=CE(t) is the concentration of toxicants in the environment of the population at time t. The exogenous rate of input of toxicants into the environment was represented by u. They investigated the persistence and extinction of a population in a polluted environment.

    Luo et al. [28] studied optimal harvesting control problem for the following age-dependent competing system of n species:

    {pit+pia=fi(a,t)μi(a,t)pink=1,kiλik(a,t)Pk(t)piui(a,t)pi,,pi(0,t)=βi(t)a2a1mi(a,t)pi(a,t)da,pi(a,0)=pi0(a),Pi(t)=a+0pi(a,t)da,i=1,2,,n,(a,t)Q, (2.2)

    where Q=(0,a+)×(0,+), [a1,a2] is the fertility interval. pi(a,t) represents the density of ith population of age a at time t, and a+ is the life expectancy of individuals; pi0 is the initial age distribution of ith population; ui(x,t) is the harvesting effort function, which is the control variable in the model. The existence of an optimal control, the necessary conditions of optimality for the control problem have been derived.

    By combining (2.1) and (2.2), we consider the following periodic, competing system with size structure in a polluted environment:

    {pit+(Vi(x,t)pi)x=fi(x,t)μi(x,ci0(t))pi3i,k=1,kiλik(x,t)Pk(t)piui(x,t)pi,dci0dt=k1ce(t)g1ci0(t)mci0(t),dcedt=k2ce(t)3i=1Pi(t)+g23i=1ci0(t)Pi(t)hce(t)+v(t),Vi(0,t)pi(0,t)=l0βi(x,ci0(t))pi(x,t)dx,0ci0(0)1,0ce(0)1,pi(x,t)=pi(x,t+T),Pi(t)=l0pi(x,t)dx,i=1,2,3,(x,t)Q, (2.3)

    where Q=(0,l)×R+,lR+ is the maximal size of an individual in the population, TR+ is the period of habitat evolution of the populations. k1,g1,m,k2,g2, and h are nonnegative constants. The meaning of the variables and functional traits are as follows: pi(x,t): the density of the ith population of size x at time t; ci0(t): the concentration of toxicants in the ith population; ce(t): the concentration of toxicants in the environment; Vi(x,t): the average rate of growth for the ith population, that is, dxdt=Vi(x,t) (see [29]); μi(x,ci0(t)),βi(x,ci0(t)): the mortality and fertility rates of the ith population, respectively; v(t): the input rate of exogenous toxicants; Pi(t): total number of individuals in the ith population; fi(x,t): the immigration rate of the ith population; λik(x,t): the interaction coefficient; ui(x,t): function of the harvesting efforts of the ith population of size x harvested at time t; k1ce(t): the organism's net uptake of toxicant from the environment; g1ci0(t) and mci0(t): the egestion and depuration rates of the toxicant in the ith population, respectively. The units of k1,g1 and m are in terms of mem10t1,t1, and t1, respectively. k2ce(t)3i=1Pi(t): the loss of the toxicant in the environment that is due to the uptake of toxicant by the total population. g23i=1ci0(t)Pi(t): the increase in the toxicant in the environment coming from the egestion of the total population. hce(t): the toxicant loss from the environment itself by volatilization and so on. The unit of k2 is in terms of m10t1; g2 is in terms of m1et1; and h is in terms of t1, where me and m0 denote the units of mass of the environment and in the ith population, respectively. The toxicant-population model with size structure is established under the condition of small toxicant capacity in the environment.

    The aim of this paper is to seek the maximum of the following objective functional J(u,v), that is

    Maximize{J(u,v):u=(u1(x,t),,u3(x,t)),v=v(t),(u,v)Ω}, (2.4)

    where

    J(u,v)=3i=1T0l0wi(x,t)ui(x,t)pi(x,t)dxdt123i=1T0l0ciu2i(x,t)dxdt12T0c4v2(t)dt,

    wi(x,t) is the selling price of an individual belonging to the ith population. The positive constants ci and c4 are the cost factors of the ith harvested population and the curbing environmental pollution, respectively. J(u,v) represents the total profit from the harvested populations during period T. The admissible control set Ω is as follows:

    Ω={(u,v)[LT(Q)]3×LT(R+):0ui(x,t)Nia.e.(x,t)Q,0v0v(t)v1a.e.tR+},

    where

    LT(Q)={ηL(Q):η(x,t)=η(x,t+T)a.e.(x,t)Q},
    LT(R+)={ηL(R+):η(t)=η(t+T)a.e.tR+}.

    This paper makes the following assumptions:

    (A1)Vi:[0,l)×R+R+ are bounded continuous functions, Vi(x,t)>0 and Vi(x,t)=Vi(x,t+T) for (x,t)Q, limxlVi(x,t)=0, and Vi(0,t)=1 for tR+. There are Lipschitz constants LVi such that

    |Vi(x1,t)Vi(x2,t)|LVi|x1x2|forx1,x2[0,l],tR+.

    (A2)0βi(x,ci0(t))=βi(x,ci0(t+T))¯βi, ¯βi are constants.

    (A3){μi(x,ci0(t))=μi0(x)+¯μi(x,ci0(t))a.e.(x,t)Q,whereμi0L1loc([0,l)),μi0(s)0a.e.x[0,l),l0μi0(s)ds+,¯μiL(Q),¯μi(x,ci0(t))0,¯μi(x,ci0(t))=¯μi(x,ci0(t+T))a.e.(x,t)Q.

    (A4)fiL(Q),0fi(x,t)=fi(x,t+T).0λik(x,t)¯λi,0wi(x,t)wi(x,t+T)¯wi,¯λi and ¯wi are constants.

    (A5) There exist constants Lβ>0,Lμ>0such that|βi(x,c1i0(t))βi(x,c2i0(t))|Lβ|c1i0(t)c2i0(t)|, |μi(x,c1i0(t))μi(x,c2i0(t))|Lμ|c1i0(t)c2i0(t)|.

    (A6)g1k1g1+m,v1h. (see [30])

    Definition 3.1. For i=1,2,3, the unique solution x=φi(t;t0,xi0) of the initial value problem x(t)=Vi(x,t),x(t0)=xi0 is said to be a characteristic curve of the hybrid system (2.3) through (t0,xi0). Let zi(t):=φi(t;0,0) denote the characteristic curve through (0,0) in the xt plane.

    For any point (x,t)[0,l)×[0,T] such that xzi(t), define the initial time τ:=τ(t,x), in order that φi(t;τ,0)=xφi(τ;t,x)=0. The solution of (2.3) is

    pi(x,t)=pi(0,tz1i(x))Πi(x;x,t)+x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Πi(x;x,t)Πi(r;x,t)dr, (3.1)

    where

    Πi(s;x,t)=exp{s0μi(r,ci0(φ1i(r;t,x)))Vi(r,φ1i(r;t,x))+3i,k=1,kiλik(r,φ1i(r;t,x))Pk(φ1i(r;t,x))Vi(r,φ1i(r;t,x))=+ui(r,φ1i(r;t,x))+Vix(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))dr}.
    ci0(t)=ci0(0)exp{(g1+m)t}+k1t0ce(s)exp{(st)(g1+m)}ds. (3.2)
    ce(t)=ce(0)exp{t0[k23i=1Pi(τ)+h]dτ}+t0[g23i=1ci0(s)Pi(s)+v(s)]exp{st[k23i=1Pi(τ)+h]dτ}ds. (3.3)

    By assumption (A1), we have Vi(0,t)=1. Let bi(t)=pi(0,t). Then, by noting that φ1i(0;t,x)=τ=tz1i(x), we have

    bi(t)=Fi(t)+l0Ki(t,x)bi(tz1i(x))dx, (3.4)

    where

    Ki(t,x)=βi(x,ci0(t))Πi(x;x,t), (3.5)
    Fi(t)=l0βi(x,ci0(t))x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Πi(x;x,t)Πi(r;x,t)drdx. (3.6)

    Define the linear and bounded operator Ai:LT(R+)LT(R+) given by

    (Aiq)(t)=l0Ki(t,x)qi(tz1i(x))dx. (3.7)

    As a consequence (3.4) can be written in LT(R+) as the following abstract equation

    bi=Aibi+Fi, (3.8)

    with FiLT(R+) defined by (3.6). We denote by r(Ai) the spectral radius of the operator Ai. If r(Ai)<1, then (3.8) has unique solution in LT(R+).

    Remark 3.1. If we denote by

    ˆβi(x)=esssuptR+βi(x,ci0(t))a.e.x[0,l),

    then (A2) and (3.7) allow us to conclude that

    r(Ai)l0ˆβi(x)dx.

    Theorem 3.1. Assume that (A1)(A6) hold. Then, the hybrid system (2.3) has a nonnegative and unique solution (p1(x,t),,p3(x,t),c10(t),,c30(t),ce(t)), such that

    (i)(pi(x,t),ci0(t),ce(t))L(Q)×L(0,T)×L(0,T).

    (ii)0ci0(t)1,0ce(t)1,t(0,T),0pi(x,t),l0pi(x,t)dxM,(x,t)Q,i=1,2,3.

    where M=M2l+fi(,)L(Q).

    Proof. Without loss of generality, we assume that ui(x,t)0. p(x,t)=(p1(x,t),,p3(x,t)),c0(t)=(c10(t),,c30(t)). When t is so large that t>z1i(l), from (3.5) it follows that

    |K1i(t,x)K2i(t,x)|=|βi(x,c1i0(t))Π1i(x;x,t)βi(x,c2i0(t))Π2i(x;x,t)||βi(x,c1i0(t))βi(x,c2i0(t))|+|βi(x,c2i0(t))||Π1i(x;x,t)Π2i(x;x,t)|Lβ|c1i0(t)c2i0(t)|+¯βix0|μi(r,c1i0(φ1i(r;t,x)))μi(r,c2i0(φ1i(r;t,x)))|Vi(r,φ1i(r;t,x))dr+¯βix03i,k=1,kiλik(r,φ1i(r;t,x))|P1k(φ1i(r;t,x))P2k(φ1i(r;t,x))|Vi(r,φ1i(r;t,x))drLβ|c1i0(t)c2i0(t)|+¯βitφ1i(0;t,x)|μi(φi(σ;t,x),c1i0(σ))μi(φi(σ;t,x),c2i0(σ))|dσ+¯βitφ1i(0;t,x)3i,k=1,kiλik(φi(σ;t,x),σ)|P1k(σ)P2k(σ)|dσLβ|c1i0(t)c2i0(t)|+¯βiLμt0|c1i0(σ)c2i0(σ)|dσ+¯βi¯λikl0t03i,k=1,ki|p1k(x,σ)p2k(x,σ)|dσdx.

    Let

    M1=max{Lβ,¯βiLμ,¯βi¯λik},
    W(t)=|c1i0(t)c2i0(t)|+t0|c1i0(σ)c2i0(σ)|dσ+l0t03i,k=1,ki|p1k(x,σ)p2k(x,σ)|dσdx.

    Then, we can obtain

    |K1i(t,x)K2i(t,x)|M1W(t). (3.9)

    By (3.6) and a similar procedure, we have

    |F1i(t)F2i(t)|=|l0βi(x,c1i0(t))x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Π1i(x;x,t)Π1i(r;x,t)drdxl0βi(x,c2i0(t))x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Π2i(x;x,t)Π2i(r;x,t)drdx|l0|βi(x,c1i0(t))βi(x,c2i0(t))|x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))drdx+l0βi(x,c2i0(t))x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))xr(|μi(δ,c1i0(φ1i(δ;t,x)))μi(δ,c2i0(φ1i(δ;t,x)))|Vi(δ,φ1i(δ;t,x))+3i,k=1,kiλik(δ,φ1i(δ;t,x))|P1k(φ1i(δ;t,x))P2k(φ1i(δ;t,x))|Vi(δ,φ1i(δ;t,x)))dδdrdxLβl0|c1i0(t)c2i0(t)|t0fi(φi(σ;t,x),σ)dσdx+¯βil0t0fi(φi(σ;t,x),σ)(Lμt0|c1i0(σ)c2i0(σ)|dσ+¯λikl0t03i,k=1,ki|p1k(x,σ)p2k(x,σ)|dσdx)dσdxfi(,)L1(Q)(Lβ|c1i0(t)c2i0(t)|+¯βiLμt0|c1i0(σ)c2i0(σ)|dσ+¯βi¯λikl0t03i,k=1,ki|p1k(x,σ)p2k(x,σ)|dσdx).

    Consequently,

    |F1i(t)F2i(t)|fi(,)L1(Q)M1W(t). (3.10)

    Since

    exp{x0(Vi)x(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))dr}=1Vi(x,t),

    and thanks to the periodicity of bi(t), we need only to consider the case t[z1i(l),z1i(l)+T]. By (3.4)–(3.6), we have

    bi(t)=Fi(t)+l0Ki(t,x)bi(tz1i(x))dx=l0βi(x,ci0(t))x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Πi(x;x,t)Πi(r;x,t)drdx+l0βi(x,ci0(t))Πi(x;x,t)bi(tz1i(x))dxl0βi(x,ci0(t))x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))drdx+l0βi(x,ci0(t))Vi(x,t)bi(tz1i(x))dxl0βi(x,ci0(t))tφ1i(0;t,x)fi(φi(s;t,x),s)dsdx+¯βil0bi(tz1i(x))Vi(x,t)dxl0βi(x,ci0(t))t0fi(φi(s;t,x),s)dsdx+¯βittz1i(l)b(s)ds¯βifi(,)L1(Q)+¯βit0b(s)ds.

    From Bellman's lemma, we have

    bi(t)¯βif(,)L1(Q)exp{t0¯βidr}¯βif(,)L1(Q)exp{¯βi(T+z1i(l))}=:M2.

    From (3.4), we get

    |b1i(t)b2i(t)||F1i(t)F2i(t)|+l0|K1i(t,x)b1i(tz1i)K2i(t,x)b2i(tz1i)||F1i(t)F2i(t)|+l0|K1i(t,x)K2i(t,x)|b1i(tz1i)dx+l0K2i(t,x)|b1i(tz1i)b2i(tz1i)|dxfi(,)L1(Q)M1W(t)+M2l0M1W(t)dx+¯βit0|b1i(s)b2i(s)|dsM3W(t)+¯βit0|b1i(s)b2i(s)|ds,

    where M3=M1(fi(,)L1(Q)+M2l). It follows from generalized Gronwall Bellman inequality that

    |b1i(t)b2i(t)|M3W(t)+¯βiexp{¯βiT}M3t0W(s)dsM4W(t),

    where M4 is a positive constant independent of pi(x,t).

    Denote X=[LT(R+,L1(0,l))]3×[L(R+)]4, then we define the state space

    Y={(p,c0,ce)Xpi(x,t)0a.e.(x,t)Q,l0pi(x,t)dxM,0ci0(t)1,0ce(t)1}.

    Define a mapping

    G:YX,G(p,c0,ce)=(G1(p,c0,ce),G2(p,c0,ce),,G7(p,c0,ce)),

    where

    Gi(p,c0,ce)(x,t)=pi(0,tz1i(x))Πi(x;x,t)+x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Πi(x;x,t)Πi(r;x,t)dr,i=1,2,3. (3.11)
    Gj(p,c0,ce)(t)=cj0(0)exp{(g1+m)t}+k1t0ce(s)exp{(st)(g1+m)}ds,j=4,5,6. (3.12)
    G7(p,c0,ce)(t)=ce(0)exp{t0[k23i=1Pi(τ)+h]dτ}+t0[g23i=1ci0(s)Pi(s)+v(s)]exp{st[k23i=1Pi(τ)+h]dτ}ds. (3.13)

    Then, we have

    l0|Gi(p,c0,ce)|dx=l0bi(tz1i(x))Πi(x;x,t)dx+l0x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))Πi(x;x,t)Πi(r;x,t)drdxl0bi(tz1i(x))dx+l0t0fi(φi(s;t,x),s)dsdxM2l+fi(,)L(Q)=M.

    It is trivial to show that G(p,c0,ce)Y. We now discuss the compressibility of G. By (3.11), we have

    l0|Gi(p1,c10,c1e)Gi(p2,c20,c2e)|dx(i=1,2,3)l0b1i(tz1i(x))|Π1i(x;x,t)Π2i(x;x,t)|dx+l0|b1i(tz1i(x))b2i(tz1i(x))|Π2i(x;x,t)dx+l0x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))|Π1i(x;x,t)Π1i(r;x,t)Π2i(x;x,t)Π2i(r;x,t)|drdxM2l0x0(|μi(r,c1i0(φ1i(r;t,x)))μi(r,c2i0(φ1i(r;t,x)))|Vi(r,φ1i(r;t,x))+3i,k=1,kiλik(r,φ1i(r;t,x))|P1k(φ1i(r;t,x))P2k(φ1i(r;t,x))|Vi(r,φ1i(r;t,x)))drdx+l0|b1i(tz1i(x))b2i(tz1i(x))|Vi(x,t)dx+l0x0fi(r,φ1i(r;t,x))Vi(r,φ1i(r;t,x))xr(|μi(δ,c1i0(φ1i(δ;t,x)))μi(δ,c2i0(φ1i(δ;t,x)))|Vi(δ,φ1i(δ;t,x))+3i,k=1,kiλik(δ,φ1i(δ;t,x))|P1k(φ1i(δ;t,x))P2k(φ1i(δ;t,x))|Vi(δ,φ1i(δ;t,x)))dδdrdxM2l0t0(Lμ|c1i0(s)c2i0(s)|+¯λikl03i,k=1,ki|p1k(x,s)p2k(x,s)|dx)dsdx+fi(,)L1(Q)t0(Lμ|c1i0(s)c2i0(s)|+¯λikl03i,k=1,ki|p1k(x,s)p2k(x,s)|dx)ds+M4t0(|c1i0(s)c2i0(s)|+(z1i(l)+T)|c1i0(s)c2i0(s)|)ds+M4(z1i(l)+T)l0t03i,k=1,ki|p1k(x,s)p2k(x,s)|dsdxM5(t0|c1i0(s)c2i0(s)|ds+l0t03i,k=1,ki|p1k(x,s)p2k(x,s)|dsdx),

    where M5=max{M2lLμ+fi(,)L1(Q)Lμ+M4(1+z1i(l)+T),M2¯λikl+fi(,)L1(Q)¯λik+M4(z1i(l)+T)}.

    By (3.12)–(3.13), we have

    |Gj(p1,c10,c1e)Gj(p2,c20,c2e)|(t)(j=4,5,6)M6t0|c1e(s)c2e(s)|ds,

    where M6=k1.

    where M7=max{k2+g2+k2hT+k2g2MT,g2M}.

    We now use the Banach fixed point theorem to demonstrate that the mapping G has only one fixed point. Due to the periodicity of elements in the set Y, we consider the case t[0,T] only. Define a new norm in L(0,T) by

    (p,c0,ce)=esssupt[0,T]eλt{3i=1l0|pi(x,t)|dx+3i=1|ci0(t)|+|ce(t)|},

    where λ>0 is large enough. Then, we have

    where M8=max{M5,M6,M7}. Thus, choosing λ>M8 yields that G is a strict contraction on (Y,). The unique fixed point (p,c0,ce) of G must be solution to (2.3).

    Theorem 3.2. If T is small enough, then there are constants Kj(T) with limT0Kj(T)>0,j=1,2, such that

    3i=1p1ip2iL(0,T;L1(0,l))+3i=1c1i0c2i0L(0,T)+c1ec2eL(0,T)K1(T)T(3i=1u1iu2iL(0,T;L1(0,l))+v1v2L(0,T)). (3.14)
    3i=1p1ip2iL1(Q)+3i=1c1i0c2i0L1(0,T)+c1ec2eL1(0,T)K2(T)T(3i=1u1iu2iL1(Q)+v1v2L1(0,T)). (3.15)

    This proof process of Theorem 3.2 is similar to that of Theorem 4.1 in [27], and is omitted here.

    In this section, we employ tangent-normal cone techniques in nonlinear functional analysis to deduct the necessary conditions for the optimal control pair.

    Theorem 4.1. If (u,v) is an optimal control pair and (p,c0,ce) is the corresponding optimal state, then

    ui(x,t)=Fi([wi(x,t)ξi(x,t)]pi(x,t)ci),i=1,2,3,a.e.(x,t)Q, (4.1)
    v(t)=F4(ξ7(t)c4)a.e.t(0,T), (4.2)

    in which the truncated mappings Fj are given by

    Fj(η)={0,η<0,η,0ηNj,j=1,2,3,4,Nj,η>Nj, (4.3)

    and (ξ1,ξ2,,ξ7) is the solution of the following adjoint system corresponding to (u,v):

    {ξit+Viξix=[μi(x,ci0(t))+3i,k=1,kiλikPk(t)+ui]ξi+[k2ce(t)g2ci0(t)]ξ7ξi(0,t)βi(x,ci0(t))+wiui,dξi+3dt=l0μi(x,ci0(t))ci0piξidx+(g1+m)ξi+3g2Pi(t)ξ7ξi(0,t)l0βi(x,ci0(t))ci0pidx,dξ7dt=k13i=1ξi+3+[k23i=1Pi(t)+h]ξ7,ξi(l,t)=0,ξi(x,t)=ξi(x,t+T),i=1,2,3,ξj(T)=0,j=4,,7. (4.4)

    Proof. The existence of a unique, bounded solution to the adjoint system (4.4) can be treated in the same manner as the state system (2.3). For any given (ν1,ν2)TΩ(u,v) (the tangent cone of Ω at (u,v)), u=(u1,,u3),ν1=(ν11,,ν31), (u+εν1,v+εν2)Ω provided that ε is small enough. Then, from J(u+εν1,v+εν2)J(u,v), we derive

    3i=1T0l0wi(ui+ενi1)pεidxdt123i=1T0l0ci(ui+ενi1)2dxdt12T0c4(v+εν2)2dt3i=1T0l0wiuipidxdt123i=1T0l0ciu2idxdt12T0c4v2dt,

    and then deduce that

    3i=1T0l0wi(uizi+νi1pi)dxdt3i=1T0l0ciuiνi1dxdtT0c4vν2dt0, (4.5)

    where 1ε(pεipi)zi,1ε(cεi0ci0)zi+3,1ε(cεece)z7,asε0. By Theorem 3.2, we get the existence of z1,z2,,z7. (pε,cε0,cεe) is the state corresponding to (u+εν1,v+εν2). It follows from the state system (2.3) that (z1,z2,,z7) satisfies

    {zit+Vizix=[μi(x,ci0(t))+3i,k=1,kiλikPk(t)+Vix+ui]zi3i,k=1,kiλikZk(t)piμi(x,ci0(t))ci0pizi+3vi1pi,dzi+3dt=k1z7g1zi+3mzi+3,dz7dt=k2ce(t)3i=1Zi(t)+g23i=1[ci0Zi(t)+zi+3Pi(t)][k23i=1Pi(t)+h1]z7+ν2,Vi(0,t)zi(0,t)=l0βi(x,ci0(t))zidx+l0βi(a,ci0(t))ci0pizi+3dx,zi(x,t)=zi(x,t+T),zi+3(0)=z7(0)=0,Pi(t)=l0pi(x,t)dx,Zi(t)=l0zi(x,t)dx,i=1,2,3. (4.6)

    We multiply the first three equations in (4.6) by ξ1,ξ2,,ξ7, respectively, and integrate on Q and (0,T). By using (4.4), we have

    3i=1T0l0wiuizidxdt=3i=1T0l0νi1ξipidxdt+T0ν2ξ7dt. (4.7)

    Substituting (4.7) into (4.5) gives

    3i=1T0l0[(wiξi)piciui]νi1dxdt+T0(c4v+ξ7)ν2dt0,

    for any (ν1,ν2)TΩ(u,v). Consequently, the structure of normal cone tells us that ((wiξi)piciui,c4v+ξ7)NΩ(u,v) (the normal cone of Ω at (u,v)), which gives the desired result.

    Theorem 4.2. If T is small enough, then there is a constant K3, such that

    3i=1ξ1iξ2iL(Q)+3i=1ξ1i+3ξ2i+3L(0,T)+ξ17ξ27L(0,T)K3T(3i=1u1iu2iL(Q)+v1v2L(0,T)). (4.8)

    The proof process of Theorem 4.2 is similar to that of Theorem 3.2, and is omitted here.

    In order to show that there exists a unique optimal control pair by means of the Ekeland variational principle, we embed the functional ˜J(u,v) into [L1(Q)]3×L1(0,T). We define

    ˜J(u,v)={J(u,v),(u,v)Ω,,otherwise.

    Lemma 5.1. ˜J(u,v) is upper semi-continuous with respect to (u,v) in [L1(Q)]3×L1(0,T).

    Proof. Let (un,vn)(u,v) as n, (pn,cn0,cne) and (p,c0,ce) be the states of (2.3) corresponding to (un,vn) and (u,v), respectively. By Riesz theorem, there is a subsequence, denoted still by (un,vn) such that

    [un(x,t)]2u2(x,t)a.e.(x,t)Q,[vn(t)]2v2(t)a.e.t(0,T),asn.

    Thus, from the Lebesgue's dominated convergence theorem yields

    limnT0l0[uni(x,t)]2dxdt=T0l0u2i(x,t)dxdt,limnT0[vn(t)]2dt=T0v2(t)dt.

    On the other hand, it follows from (3.15) that

    |T0l0wi(x,t)uni(x,t)pni(x,t)dxdtT0l0wi(x,t)ui(x,t)pi(x,t)dxdt|T0l0wi(x,t)pni(x,t)|uni(x,t)ui(x,t)|dxdt+T0l0wi(x,t)ui(x,t)|pni(x,t)pi(x,t)|dxdtM¯wiuniuiL1(Q)+Ni¯wipnipiL1(Q)M¯wiuniuiL1(Q)+Ni¯wiK2(T)T(uniuiL1(Q)+v1v2L1(0,T)).

    Therefore,

    limnT0l0wi(x,t)uni(x,t)pni(x,t)dxdt=T0l0wi(x,t)ui(x,t)pi(x,t)dxdt.

    In a word, we have proved that lim supn˜J(un,vn)˜J(u,v).

    Theorem 5.1. If T is sufficiently small, there exists one and only one optimal control pair (u,v), which is in feedback and is determined by (4.1)–(4.4) and (2.3), where C1 and C2 are the supremum of |pi| and |ξj|,i=1,2,3,j=1,2,,7, respectively.

    Proof. Define the mapping L:ΩΩ as follows:

    L(u,v)=F((w1ξ1)p1c1,,(w3ξ3)p3c3,ξ7c4)=(F1((w1ξ1)p1c1),,F3((w3ξ3)p3c3),F4(ξ7c4)),

    where (p,c0,ce) and (ξ1,ξ1,,ξ7) are the state and adjoint state, respectively, corresponding to the control (u,v). We show that L admits a unique fixed point, which maximizes the functional L.

    From Lemma 5.1 and the Ekeland variational principle, for any given ε>0, there exists (uε,vε)Ω such that

    ˜J(uε,vε)sup(u,v)Ω˜J(u,v)ε, (5.1)
    ˜J(uε,vε)sup(u,v)Ω{˜J(u,v)ε(3i=1uεiuiL1(Q)+vεvL1(0,T))}. (5.2)

    Thus, the perturbed functional

    ˜Jε(u,v)=˜J(u,v)ε(3i=1uεiuiL1(Q)+vεvL1(0,T)),

    attains its supremum at (uε,vε). Then, we argue as in Theorem 4.1:

    (uε,vε)=L(uε,vε)=(F1((w1ξε1)pε1+εθε1c1),,F3((w3ξε3)pε3+εθε3c3),F4(ξε7+εθε4c4)), (5.3)

    where (pε,cε0,cεe) and (ξε1,ξε2,,ξε7) are the state and adjoint state, respectively, corresponding to the control (uε,vε), θε1,,θε3L(Q),θε4L(0,T), and with |θεi|1,i=1,2,3,4.

    First, we show that L has only one fixed point. Let (pj,cj0,cje) and (ξj1,ξj2,,ξj7) be the state and adjoint state corresponding to the control (uj,vj),j=1,2. By (3.14) and (4.8), we have

    L(u1,v1)L(u2,v2)=3i=1Fi((wiξ1i)p1ici)Fi((wiξ2i)p2ici)L(Q)+F4(ξ17c4)F4(ξ27c4)L(0,T)3i=1(wiξ1i)p1ici(wiξ2i)p2iciL(Q)+ξ17c4ξ27c4L(0,T)3i=1wi(p1ip2i)ci+|ξ1i|(p1ip2i)ci+|p2i|(ξ1iξ2i)ciL(Q)+ξ17ξ27c4L(0,T)T(3i=11ci(¯wiK1+C2K1+C1K3)+K3)(3i=1u1iu2iL(Q)+v1v2L(0,T)).

    Clearly, L is a contraction if T is sufficiently small. Hence, L has a unique fixed point (u,v).

    Next, we prove (uε,vε)(u,v) as ε0+. The relations (4.1), (4.2) and (5.3) lead to

    it is easy to derive that

    (u,v)(uε,vε)L(u,v)L(uε,vε)+L(uε,vε)(uε,vε)T(3i=11ci(¯wiK1+C2K1+C1K3)+K3)(3i=1uiuεiL(Q)+vvεL(0,T))+ε4i=11ci.

    So, if T is small enough, the following result holds:

    3i=1uiuεiL(Q)+vvεL(0,T)ε4i=11ci1T(3i=11ci(¯wiK1+C2K1+C1K3)+K3),

    which gives the desired result.

    Finally, passing to the limit ε0+ in the inequality of (5.2) and using Lemma 5.1 yield ˜J(u,v)lim sup(u,v)Ω˜J(u,v), which finishes the proof.

    In this section, our goal is to obtain a numerical approximation for the nonnegative T-periodic solution of the system (2.3). We numerically study the evolution of a single species in a polluted environment as a simplification of the complete model (2.3). If the harvest effort term and the summation term are considered, it will be transformed into the optimization problem (2.3)–(2.4), which is complex.

    Suppose the computational domain ˜Q=[0,l]×[0,˜T] is divided into an J×N mesh with the spacial step size h=lJ=0.01 in the x direction and time step size τ=˜TN=0.02. The grid points (xj,tn) are defined by

    xj=jh,j=0,1,2,,J;
    tn=nτ,n=0,1,2,,N,

    where J and N are two integers. The pnj and fnj terms denote the solution p(jh,nτ) and source term f(jh,nτ) of the finite difference equation, respectively.

    Based on the state system (2.3), the finite difference scheme can be written as follows:

    pnjpn1jτ+Vpnjpnj1h+Vxpnj+μpnjfnj=0, (6.1)

    where j=1,2,,J;n=1,2,,N. It follows from (6.1) that

    dVpnj1+[1+dV+τ(Vx+μ)]pnj=pn1j+τfnj, (6.2)

    where d=τh.

    Since V(0,t)=1, then the boundary condition p(0,t)=l0β(x,c0(t))p(x,t)dx and initial condition p(x,0)=p0(x) can be discretized as

    {p0j=p0j,pn0=Jj=1βjpnjh. (6.3)

    From (6.2) and (6.3), we have the matrix associated with the system of linear equations of the finite difference method

    APn=Pn1+τF, (6.4)

    where

    A=[1+dV+τ(Vx+μ)dVβhdVβhdVβhdVβhdV1+dV+τ(Vx+μ)0000dV1+dV+τ(Vx+μ)000dV],
    Pn=(pn1,pn2,,pnJ)T,F=(fn1,fn2,,fnJ)T.

    Note that A is an upper triangular matrix, so the nonlinear algebraic equations (6.4) have solutions. In this paper, we choose the following parameters:

    {β(x,c0(t))=100x2(1x)(1+sin(πx))|sin2πc0(t)T|,μ(x,c0(t))=e4x(1x)1.4(2+cos2πc0(t)T),V(x,t)=1x,f(x,t)=2+(1+x)sin(2πtT),p0(x)=ex,u(x,t)=0,x=1,T=13,˜T=6T.

    In this paper, we used the backward difference scheme and chasing method, and (6.4) was solved through programming. The fertility rate, mortality rate, and immigration rate were T-periodic and were all greater than zero, which is consistent with the assumptions. We considered T=13. Their graphs are given in Figures 13, respectively. The fertility rate was the highest when the size was half and the mortality rate was the highest when the size was the maximum, which conformed to the empirical situation. Therefore, the selection of parameters β, μ, and f was reasonable.

    Figure 1.  Fertility rate of the population.
    Figure 2.  Mortality rate of the population.
    Figure 3.  Immigration rate of the population.

    The graphic of the numerical solution p is given in Figure 4. Over time, solution p showed T-periodic changes. We take the numerical solution of (2.3), corresponding to an arbitrary positive initial datum p0, on some interval [kT,(k+1)T], where k is large enough. On such an interval, the solution p was already stable. We can then get the periodic solution of (2.3) by extending the numerical solution p. During computation we found that any positive initial datum p0 was appropriate for use.

    Figure 4.  Numerical solution of the system.

    The study of time periodic models is of great importance due to the fact that the vital rates and the inflow are often time periodic. In the foregoing, we have established the existence and uniqueness of a nonnegative solution of the hybrid system (2.3). The necessary conditions for optimal controls were provided. The existence of the unique optimal control pair was investigated. Some numerical results were finally presented. The results implied that the solution of (2.3) always maintains the pattern of increasing periodically, and any positive initial datum p0 is appropriate. Over time, the density of the population increased first and then decreased in a cycle. The bang-bang structure of solutions is much more common in optimal population management.

    Furthermore, if Vi(x,t)=1 for Q=(0,l)×R+,i=1,2,3, the state system degenerates into an age-structured model, and our results cover the corresponding results [5,6,7]. Note that the individual price factor wi(x,t) plays an important role in the structure of the optimal controller (4.1). However, as we do not have a clear biological meaning for the solutions ξi(i=1,2,,7) of the adjoint system (4.4), it is difficult to give a precise explanation of the threshold conditions (4.1) and (4.2). In specific applications, the optimal population density and optimal policy are calculated by combining the state system and the adjoint system. This is a challenging problem, and future work in the area should address it.

    The authors thank the referees for their valuable comments and suggestions on the original manuscript that helped improve its quality. The work was supported by the National Natural Science Foundation of China under grant 11561041.

    None of the authors has a conflict of interest in the publication of this paper.



    [1] S. Kumar, Rishabh, K. Bhatia, A review on face identification systems in computer vision, WoS, 2 (2023), 230–238. Available from: https://innosci.org/wos/article/view/1474.
    [2] W. Yang, S. Wang, J. Hu, G. Zheng, C. Valli, A fingerprint and finger-vein based cancelable multi-biometric system, Pattern Recognit., 78 (2018), 242–251. https://doi.org/10.1016/j.patcog.2018.01.026 doi: 10.1016/j.patcog.2018.01.026
    [3] K. Conger, R. Fausset, S. F. Kovaleski, San Francisco bans facial recognition technology, in The New York Times, 14 (2019).
    [4] L. Li, X. Mu, S. Li, H. Peng, A review of face recognition technology, IEEE Access, 8 (2020), 139110–139120. https://doi.org/10.1109/ACCESS.2020.3011028 doi: 10.1109/ACCESS.2020.3011028
    [5] N. Zeng, H. Zhang, B. Song, W. Liu, Y. Li, A. M. Dobaie, Facial expression recognition via learning deep sparse autoencoders, Neurocomputing, 273 (2018), 643–649. https://doi.org/10.1016/j.neucom.2017.08.043 doi: 10.1016/j.neucom.2017.08.043
    [6] N. Zeng, X. Li, P. Wu, H. Li, X. Luo, A novel tensor decomposition-based efficient detector for low-altitude aerial objects with knowledge distillation scheme, IEEE/CAA J. Autom. Sin., 11 (2024), 487–501. https://doi.org/10.1109/JAS.2023.124029 doi: 10.1109/JAS.2023.124029
    [7] J. M. Mase, N. Leesakul, G. P. Figueredo, M. T. Torres, Facial identity protection using deep learning technologies: an application in affective computing, AI Ethics, 3 (2023), 937–946. https://doi.org/10.1007/s43681-022-00215-y doi: 10.1007/s43681-022-00215-y
    [8] X. Jin, Y. Xie, X. S. Wei, B. R. Zhao, Z. M. Chen, X. Tan, Delving deep into spatial pooling for squeeze-and-excitation networks, Pattern Recognit., 121 (2022), 108159. https://doi.org/10.1016/j.patcog.2021.108159 doi: 10.1016/j.patcog.2021.108159
    [9] X. Lian, Y. Pang, J. Han, J. Pan, Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation, Pattern Recognit., 110 (2021), 107622. https://doi.org/10.1016/j.patcog.2020.107622 doi: 10.1016/j.patcog.2020.107622
    [10] D. Yang, X. Wang, N. Zhu, S. Li, N. Hou, MJ-GAN: Generative adversarial network with multi-grained feature extraction and joint attention fusion for infrared and visible image fusion, Sensors, 23 (2023), 6322. https://doi.org/10.3390/s23146322 doi: 10.3390/s23146322
    [11] Z. Shao, X. Wang, B. Li, Y. Zhang, Y. Shang, J. Ouyang, Cancelable color face recognition using trinion gyrator transform and randomized nonlinear PCANet, Multimedia Tools Appl., (2024), 1–15. https://doi.org/10.1007/s11042-023-17905-2 doi: 10.1007/s11042-023-17905-2
    [12] Z. Shao, L. Li, Z. Zhang, B. Li, X. Liu, Y. Shang, et al., Cancelable face recognition using phase retrieval and complex principal component analysis network, Mach. Vision Appl., 35 (2024), 12. https://doi.org/10.1007/s00138-023-01496-x doi: 10.1007/s00138-023-01496-x
    [13] H. Tao, Q. Duan, Hierarchical attention network with progressive feature fusion for facial expression recognition, Neural Networks, 170 (2024), 337–348. https://doi.org/10.1016/j.neunet.2023.11.033 doi: 10.1016/j.neunet.2023.11.033
    [14] H. Tao, Q. Duan, A spatial-channel feature-enriched module based on multi-context statistics attention, IEEE Internet Things J., 2023. https://doi.org/10.1109/JIOT.2023.3339722 doi: 10.1109/JIOT.2023.3339722
    [15] M. Ren, Y. Wang, Y. Zhu, K. Zhang, Z. Sun, Multiscale dynamic graph representation for biometric recognition with occlusions, IEEE Trans. Pattern Anal. Mach. Intell., 45 (2023), 15120–15136. https://doi.org/10.1109/TPAMI.2023.3298836 doi: 10.1109/TPAMI.2023.3298836
    [16] S. B. Chaabane, M. Hijji, R. Harrabi, H. Seddik, Face recognition based on statistical features and SVM classifier, Multimedia Tools Appl., 81 (2022), 8767–8784. https://doi.org/10.1007/s11042-021-11816-w doi: 10.1007/s11042-021-11816-w
    [17] J. S. Talahua, J. Buele, P. Calvopiña, J. Varela-Aldas, Facial recognition system for people with and without face mask in times of the COVID-19 pandemic, Sustainability, 13 (2021), 6900. https://doi.org/10.3390/su13126900 doi: 10.3390/su13126900
    [18] J. Wu, W. Feng, G. Liang, T. Wang, G. Li, Y. Zheng, A privacy protection scheme for facial recognition and resolution based on edge computing, Secur. Commun. Netw., 2022 (2022), 4095427. https://doi.org/10.1155/2022/4095427 doi: 10.1155/2022/4095427
    [19] M. Zhang, L. Wang, Y. Zou, W. Yan, Analysis of consumers' innovation resistance behavior to facial recognition payment: an empirical investigation, WHICEB 2022 Proc., 32 (2022). Available from: https://aisel.aisnet.org/whiceb2022/32/.
    [20] E. Farooq, A. Borghesi, A federated learning approach for anomaly detection in high performance computing, in 2023 IEEE 35th International Conference on Tools with Artificial Intelligence (ICTAI), (2023), 496–500. https://doi.org/10.1109/ICTAI59109.2023.00079
    [21] M. H. B. Alhlffee, Y. Huang, Y. A. Chen, 2D facial landmark localization method for multi-view face synthesis image using a two-pathway generative adversarial network approach, PeerJ Comput. Sci., 8 (2022), e897. https://doi.org/10.7717/peerj-cs.897 doi: 10.7717/peerj-cs.897
    [22] S. Cen, H. Luo, J. Huang, W. Shi, X. Chen, Pre-trained feature fusion and multidomain identification generative adversarial network for face frontalization, IEEE Access, 10 (2022), 77872–77882. https://doi.org/10.1109/ACCESS.2022.3193386 doi: 10.1109/ACCESS.2022.3193386
    [23] A. Ullah, H. Elahi, Z. Sun, A. Khatoon, I. Ahmad, Comparative analysis of AlexNet, ResNet18 and SqueezeNet with diverse modification and arduous implementation, Arabian J. Sci. Eng., 47 (2022), 2397–2417. https://doi.org/10.1007/s13369-021-06182-6 doi: 10.1007/s13369-021-06182-6
    [24] A. Ullah, H. Xie, M. O. Farooq, Z. Sun, Pedestrian detection in infrared images using fast RCNN, in 2018 Eighth International Conference on Image Processing Theory, Tools and Applications (IPTA), (2018), 1–6. https://doi.org/10.1109/IPTA.2018.8608121
    [25] O. Basystiuk, N. Melnykova, Z. Rybchak, Machine Learning Methods and Tools for Facial Recognition Based on Multimodal Approach, 2023. Available from: https://ceur-ws.org/Vol-3426/paper13.pdf.
    [26] B. Thaman, T. Cao, N. Caporusso, Face mask detection using mediapipe facemesh, in 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO), (2022), 378–382. https://doi.org/10.23919/MIPRO55190.2022.9803531
    [27] S. Bhatlawande, S. Shilaskar, T. Gadad, S. Ghulaxe, R. Gaikwad, Smart home security monitoring system based on face recognition and android application, in 2023 International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT), (2023), 222–227. https://doi.org/10.1109/IDCIoT56793.2023.10053558
    [28] C. S. Hsu, S. F. Tu, P. C. Chiu, Design of an e-diploma system based on consortium blockchain and facial recognition, Educ. Inf. Technol., 27 (2022), 5495–5519. https://doi.org/10.1007/s10639-021-10840-5 doi: 10.1007/s10639-021-10840-5
    [29] S. Rizwan, M. Zubair, A. Ghani, S. Ahmed, B. Fayyaz, Decentralized voting system based on regions using facial recognition, J. Independent Stud. Res. Comput., 20 (2022). https://doi.org/10.31645/JISRC.22.20.1.8 doi: 10.31645/JISRC.22.20.1.8
    [30] F. Schroff, D. Kalenichenko, J. Philbin, Facenet: a unified embedding for face recognition and clustering, in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2015), 815–823. https://doi.org/10.1109/CVPR.2015.7298682
    [31] J. Hu, L. Shen, G. Sun, Squeeze-and-excitation networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2018), 7132–7141.
    [32] Z. Wen, W. Lin, T. Wang, G. Xu, Distract your attention: multi-head cross attention network for facial expression recognition, Biomimetics, 8 (2023), 199. https://doi.org/10.3390/biomimetics8020199 doi: 10.3390/biomimetics8020199
    [33] A. R. Revanda, C. Fatichah, N. Suciati, Utilization of generative adversarial networks in face image synthesis for augmentation of face recognition training data, in 2020 International Conference on Computer Engineering, Network, and Intelligent Multimedia (CENIM), (2020), 396–401. https://doi.org/10.1109/CENIM51130.2020.9297899
  • This article has been cited by:

    1. Bin Feng, Chengbo Ai, Haofei Zhang, Fusion of Infrared and Visible Light Images Based on Improved Adaptive Dual-Channel Pulse Coupled Neural Network, 2024, 13, 2079-9292, 2337, 10.3390/electronics13122337
  • Reader Comments
  • © 2024 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2776) PDF downloads(105) Cited by(1)

Figures and Tables

Figures(4)  /  Tables(4)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog