There are approximately 2.2 billion people around the world with varying degrees of visual impairments. Among them, individuals with severe visual impairments predominantly rely on hearing and touch to gather external information. At present, there are limited reading materials for the visually impaired, mostly in the form of audio or text, which cannot satisfy the needs for the visually impaired to comprehend graphical content. Although many scholars have devoted their efforts to investigating methods for converting visual images into tactile graphics, tactile graphic translation fails to meet the reading needs of visually impaired individuals due to image type diversity and limitations in image recognition technology. The primary goal of this paper is to enable the visually impaired to gain a greater understanding of the natural sciences by transforming images of mathematical functions into an electronic format for the production of tactile graphics. In an effort to enhance the accuracy and efficiency of graph element recognition and segmentation of function graphs, this paper proposes an MA Mask R-CNN model which utilizes MA ConvNeXt as its improved feature extraction backbone network and MA BiFPN as its improved feature fusion network. The MA ConvNeXt is a novel feature extraction network proposed in this paper, while the MA BiFPN is a novel feature fusion network introduced in this paper. This model combines the information of local relations, global relations and different channels to form an attention mechanism that is able to establish multiple connections, thus increasing the detection capability of the original Mask R-CNN model on slender and multi-type targets by combining a variety of multi-scale features. Finally, the experimental results show that MA Mask R-CNN attains an 89.6% mAP value for target detection and 72.3% mAP value for target segmentation in the instance segmentation of function graphs. This results in a 9% mAP improvement for target detection and 12.8% mAP improvement for target segmentation compared to the original Mask R-CNN.
Citation: Jiale Lu, Jianjun Chen, Taihua Xu, Jingjing Song, Xibei Yang. Element detection and segmentation of mathematical function graphs based on improved Mask R-CNN[J]. Mathematical Biosciences and Engineering, 2023, 20(7): 12772-12801. doi: 10.3934/mbe.2023570
[1] | Yilin Tu, Jin-E Zhang . Event-triggered impulsive control for input-to-state stability of nonlinear time-delay system with delayed impulse. Mathematical Biosciences and Engineering, 2025, 22(4): 876-896. doi: 10.3934/mbe.2025031 |
[2] | Siyu Li, Shu Li, Lei Liu . Fuzzy adaptive event-triggered distributed control for a class of nonlinear multi-agent systems. Mathematical Biosciences and Engineering, 2024, 21(1): 474-493. doi: 10.3934/mbe.2024021 |
[3] | Chaoyue Wang, Zhiyao Ma, Shaocheng Tong . Adaptive fuzzy output-feedback event-triggered control for fractional-order nonlinear system. Mathematical Biosciences and Engineering, 2022, 19(12): 12334-12352. doi: 10.3934/mbe.2022575 |
[4] | Wenjing Wang, Jingjing Dong, Dong Xu, Zhilian Yan, Jianping Zhou . Synchronization control of time-delay neural networks via event-triggered non-fragile cost-guaranteed control. Mathematical Biosciences and Engineering, 2023, 20(1): 52-75. doi: 10.3934/mbe.2023004 |
[5] | Dong Xu, Xinling Li, Weipeng Tai, Jianping Zhou . Event-triggered stabilization for networked control systems under random occurring deception attacks. Mathematical Biosciences and Engineering, 2023, 20(1): 859-878. doi: 10.3934/mbe.2023039 |
[6] | Duoduo Zhao, Fang Gao, Jinde Cao, Xiaoxin Li, Xiaoqin Ma . Mean-square consensus of a semi-Markov jump multi-agent system based on event-triggered stochastic sampling. Mathematical Biosciences and Engineering, 2023, 20(8): 14241-14259. doi: 10.3934/mbe.2023637 |
[7] | Guodong Zhao, Haitao Li, Ting Hou . Survey of semi-tensor product method in robustness analysis on finite systems. Mathematical Biosciences and Engineering, 2023, 20(6): 11464-11481. doi: 10.3934/mbe.2023508 |
[8] | Na Zhang, Jianwei Xia, Tianjiao Liu, Chengyuan Yan, Xiao Wang . Dynamic event-triggered adaptive finite-time consensus control for multi-agent systems with time-varying actuator faults. Mathematical Biosciences and Engineering, 2023, 20(5): 7761-7783. doi: 10.3934/mbe.2023335 |
[9] | Qiushi Wang, Hongwei Ren, Zhiping Peng, Junlin Huang . Dynamic event-triggered consensus control for nonlinear multi-agent systems under DoS attacks. Mathematical Biosciences and Engineering, 2024, 21(2): 3304-3318. doi: 10.3934/mbe.2024146 |
[10] | Mingxia Gu, Zhiyong Yu, Haijun Jiang, Da Huang . Distributed consensus of discrete time-varying linear multi-agent systems with event-triggered intermittent control. Mathematical Biosciences and Engineering, 2024, 21(1): 415-443. doi: 10.3934/mbe.2024019 |
There are approximately 2.2 billion people around the world with varying degrees of visual impairments. Among them, individuals with severe visual impairments predominantly rely on hearing and touch to gather external information. At present, there are limited reading materials for the visually impaired, mostly in the form of audio or text, which cannot satisfy the needs for the visually impaired to comprehend graphical content. Although many scholars have devoted their efforts to investigating methods for converting visual images into tactile graphics, tactile graphic translation fails to meet the reading needs of visually impaired individuals due to image type diversity and limitations in image recognition technology. The primary goal of this paper is to enable the visually impaired to gain a greater understanding of the natural sciences by transforming images of mathematical functions into an electronic format for the production of tactile graphics. In an effort to enhance the accuracy and efficiency of graph element recognition and segmentation of function graphs, this paper proposes an MA Mask R-CNN model which utilizes MA ConvNeXt as its improved feature extraction backbone network and MA BiFPN as its improved feature fusion network. The MA ConvNeXt is a novel feature extraction network proposed in this paper, while the MA BiFPN is a novel feature fusion network introduced in this paper. This model combines the information of local relations, global relations and different channels to form an attention mechanism that is able to establish multiple connections, thus increasing the detection capability of the original Mask R-CNN model on slender and multi-type targets by combining a variety of multi-scale features. Finally, the experimental results show that MA Mask R-CNN attains an 89.6% mAP value for target detection and 72.3% mAP value for target segmentation in the instance segmentation of function graphs. This results in a 9% mAP improvement for target detection and 12.8% mAP improvement for target segmentation compared to the original Mask R-CNN.
The Navier-Stokes equations are a typical nonlinear system, which model the mechanics law for fluid flow and have been applied in many fields. There are many research findings on the Navier-Stokes system, involving well-posedness, long-time behavior, etc., [1,2,3,4,5,6,7,8,9,10]. Furthermore, to simulate the fluid movement modeled by the Navier-Stokes equations, some regularized systems are proposed, such as the Navier-Stokes-Voigt equations. The Navier-Stokes-Voigt equations were introduced by Oskolkov in 1973, which describe the motion of Kelvin-Voigt viscoelastic incompressible fluid. Based on the global well-posedness of 3D Navier-Stokes-Voigt equations in [11], many interesting results on long-time behavior of solutions have been obtained, such as the existence of global attractor and pullback attractors, determining modes and estimate on fractal dimension of attractor [12,13,14] and references therein for details.
The influence of past history term on dynamical system is well known, we refer to [15,16,17,18] for interesting conclusions, such as the global well-posedness, the existence of attractors and so on. In 2013, Gal and Tachim-Medjo [17] studied the Navier-Stokes-Voigt system with instantaneous viscous term and memory-type viscous term, and obtained the well-posedness of solution and exponential attractors of finite dimension. In 2018, Plinio et al. [18] considered the Navier-Stokes-Voigt system in [18], in which the instantaneous viscous term was completely replaced by the memory-type viscous term and the Ekman damping βu was presented. The authors showed the existence of regular global and exponential attractors with finite dimension. The presence of Ekman damping was to eliminate the difficulties brought by the memory term in deriving the dissipation of system.
Some convergence results of solutions or attractors as perturbation vanishes for the non-autonomous dynamical systems without memory can be seen in [19,20,21,22]. However, there are few convergence results on the system with memory. Therefore, our purpose is to study the tempered pullback dynamics and robustness of the following 3D incompressible Navier-Stokes-Voigt equations on the bounded domain Ω with memory and the Ekman damping:
{∂∂t(u−αΔu)−∫∞0g(s)Δu(t−s)ds+(u⋅∇)u+βu+∇p=fε(t,x), (t,x)∈Ωτ,divu=0, (t,x)∈Ωτ,u(t,x)=0, (t,x)∈∂Ωτ,u(τ,x)=u(τ), x∈Ω,u(τ−s,x)=φ(s,x), (s,x)∈Ω0, | (1.1) |
where Ωτ=(τ,+∞)×Ω, ∂Ωτ=(τ,+∞)×∂Ω, Ω0=(0,∞)×Ω, τ∈R+ is the initial time, α>0 is a length scale parameter characterizing the elasticity of fluid, β>0 is the Ekman dissipation constant, u=(u1(t,x),u2(t,x),u3(t,x)) is the unknown velocity field of fluid, and p is the unknown pressure. The non-autonomous external force is fε(t,x)=f1(x)+εf2(t,x) (0≤ε<ε0), where ε0 is a fixed constant small enough. In addition, u(τ) is the initial velocity, and φ(s,x) denotes the past history of velocity. The memory kernel g: [0,∞)→[0,∞) is supposed to be convex, smooth on (0,∞) and satisfies that
g(∞)=0, ∫∞0g(s)ds=1. |
In general, we give the past history variable
η=ηt(s)=∫s0u(t−σ)dσ, s≥0, |
which satisfies
∂∂tη=−∂∂sη+u(t). |
Also, η has the explicit representation
{ηt(s)=∫s0u(t−σ)dσ, 0<s≤t,ηt(s)=η0(s−t)+∫t0u(t−σ)dσ, s>t, | (1.2) |
and
ητ(s)=∫s0φ(σ)dσ. |
Next, we give the main features of this paper as follows.
1) Inspired by [18,23], we provide a detailed representation and Gronwall type estimates for the energy of (1.1) dependent on ε in Lemma 2.1, with a focus on the parameters ω, Λ and the increasing function J(∗). Using these parameters, we construct the universe D and derive the existence of D−pullback absorbing sets, see Lemma 4.9.
2) Via the decomposition method, we show that the process of the system has the property of D−κ−pullback contraction in the space NV, and the D−pullback asymptotic compactness is obtained naturally. Based on the theory of attractor in [1,24], the D-pullback attractors for the process {Sε(t,τ)} in NV are derived, see Theorem 3.3.
3) When the perturbation parameter ε→0 with the non-autonomous external force, the robustness is obtained via the upper semi-continuity of pullback attractors of (1.1) by using the technique in [18,21,22], see Theorems 3.2 in Section 3.
This paper is organized as follows. Some preliminaries are given in Section 2, and the main results are stated in Section 3, which contains the global well-posedness of solution, the existence of pullback attractors and robustness. Finally, the detailed proofs are provided in Sections 4 and 5.
● The Sobolev spaces
Let E={u|u∈(C∞0(Ω))3,divu=0}, H is the closure of E in (L2(Ω))3 topology with the norm and inner product as
|u|=‖u‖H=(u,v)1/2, (u,v)=3∑j=1∫Ωuj(x)vj(x)dx, ∀ u,v∈H. |
V is the closure of E in (H1(Ω))3 topology with the norm and inner product as
‖u‖=‖u‖V=((u,u))1/2V, ((u,v))V=3∑i,j=1∫Ω∂uj∂xi∂vj∂xidx, ∀ u,v∈V. |
Also, we denote
((u,v))Vα=(u,v)+α((u,v))V, ||u||2Vα=|u|2+α||u||2. |
H and V are Hilbert spaces with their dual spaces H and V′ respectively, ‖⋅‖∗ and ⟨⋅,⋅⟩ denote the norm in V′ and the dual product between V and V′ respectively, and also H to itself.
● The fractional power functional spaces
Let PL be the Helmholz-Leray orthogonal projection in (L2(Ω))3 onto H [3,7], and
PL: H⊕H⊥→H, |
where
H⊥={u∈(L2(Ω))3; ∃ χ∈(L2loc(Ω))3: u=∇χ}. |
A=−PLΔ is the Stokes operator with eigenvalues {λj}∞j=1 and orthonormal eigenfunctions {ωj}∞j=1.
Define the fractional operator As by
Asu=∑jλsj(u,ωj)ωj, s∈R, j∈Z+ |
for u=∑j(u,ωj)ωj with the domain D(As)={u|Asu∈H}, and we use the norm of D(As) as
‖u‖22s=|Asu|2=∑jλ2sj|(u,ωj)|2. |
Especially, denote W=D(A), and V=D(A1/2) with norm ‖u‖1=|A1/2u|=‖u‖ for any u∈V.
● The memory spaces
For any s∈(0,∞), we define μ(s)=−g′(s), which is nonnegative, absolutely continuous, decreasing (μ′≤0 almost everywhere) and
κ=∫∞0μ(s)ds>0. | (2.1) |
Also, there exists δ>0 such that
μ′(s)+δμ(s)≤0, a.e. s∈(0,∞). | (2.2) |
Let
MX=L2μ(R+;X), X=Vor W, |
which is a Hilbert space on R+ with inner product and norm
((η,ζ))MX=∫∞0μ(s)((η(s),ζ(s))Xds, ‖η‖MX=(∫∞0μ(s)||η(s)||2Xds)1/2. |
Moreover, the extended memory space can be defined as
NX=X×MX |
equipped with the norm
‖(u,η)‖2NX=‖u‖2X+‖η‖2MX. |
● The bilinear and trilinear operators
The bilinear and trilinear operators are defined as follows [8]
B(u,v):=PL((u⋅∇)v), ∀ u,v∈V, | (2.3) |
b(u,v,w)=<B(u,v),w>=3∑i,j=1∫Ωui∂vj∂xiwjdx. | (2.4) |
Denote B(u)=B(u,u), B(u,v) is a continuous operator from V×V to V′, and there hold
b(u,v,v)=0, b(u,v,w)=−b(u,w,v), ∀ u,v,w∈V. | (2.5) |
● Some useful lemmas
Lemma 2.1. ([23]) Assume that
1) A nonnegative function h is locally summable on R+, and for any ε∈(0,ε0] and any t≥τ≥0 there holds
ε∫tτe−ε(t−s)h(s)ds≤85supt≥0∫t+1th(s)ds<∞. |
2) The nonnegative function yε(t) is absolutely continuous on [τ,∞), and satisfies for some constants R,C0≥0 that
yε(t)≤Re−ε(t−τ)+εp∫tτe−ε(t−s)h(s)yε(s)qds+C0ε1+r, |
where p,q,r≥0, and p−1>(q−1)(1+r)≥0.
3) Let z(t)≥0 be a continuous function on (0,∞) equivalent to yε(t), which means there exist some constants M≥1, L≥0 such that
z(t)≤Myε(t)≤M(z(t)+L). |
Then, there exist ω, Λ>0 and an increasing function J(∗): R+→R+ such that
z(t)≤J(MR)e−ω(t−τ)+Λ(MC0+L). |
Remark 2.1. Under the assumptions in Lemma 2.1, there exists a constant θ∈(0,1) satisfying
p1=pθ−θ+1−q>0, p2=1−θ−rθ>0. |
Denote
p3=max{ε−1/θ0,(2supt≥0∫t+1th(s)ds)1/p1,C1/p20}, p4=2max{6Rp−13,1}, |
then
ω=ωθ,p,q,r,C0=12p−θ3, Λ=Λθ,p,q,r,C0=5p1−p23, |
J(R)=Jθ,p,q,r,C0(R)=2pq4p3exp(pθ41−2−θln(6pq4)). |
Lemma 2.2. ([15]) Let η be the past history variable and (1.2) holds. Then
‖ηt(s)‖2MX−‖ητ(s)‖2MX≤2∫tτ((ησ,u(σ)))MXdσ. |
We assume that f1(x) and f2(t,x) satisfy the following hypotheses:
(C1) The function f1∈H.
(C2) f2(t,x) is translation bounded in L2loc(R,H), which means there exists a constant K>0 such that
supt∈R∫t+1t|f2(s)|2ds<K, |
and for any t∈R, there also holds
∫t−∞eιs‖f2(s)‖2ds<∞, 0<ι≤νε0, ν=min{ακδ72,1}, | (3.1) |
where κ,δ are the same as parameters in (2.1) and (2.2) respectively.
Construct the infinitesimal generator of right-translation semigroup on MX
Tη=−∂∂sη, |
whose domain is
D(T)={η∈MX: ∂∂sη∈MX, η(0)=0}. |
Given initial datum U(τ)=(u(τ),ητ)∈NV, then (1.1) can be transformed into the following abstract form
{∂∂t(u+αAu)+∫∞0μ(s)Aη(s)ds+B(u,u)+βu=PLfε(t,x), (t,x)∈Ωτ,∂∂tη=Tη+u,divu=0, (t,x)∈Ωτ,u(t,x)=0, (t,x)∈∂Ωτ,u(τ,x)=u(τ), x∈Ω,ητ(s)=∫s0φ(σ)dσ. | (3.2) |
● Global well-posedness of solution
Definition 3.1 A function U(t)=(u(t),ηt): [τ,+∞)→NV is called the weak solution to (3.2), if for any fixed T>τ there hold
(i) U(t)∈C([τ,T];NV), ∂u∂t∈L2(τ,T;V).
(ii) U(τ)=(u(τ),ητ).
(iii) for any w∈C1([τ,T];V) with w(T,x)=0, there holds
−∫Tτ⟨u+αAu,wt⟩dt+∫Tτ∫∞0μ(s)((η(s),w))Vdsdt+∫Tτb(u,u,w)dt+∫Tτ(βu,w)dt=((u(τ),w(τ)))Vα+∫Tτ(PLfε,w)dt. | (3.3) |
Theorem 3.2. Let U(τ)∈NV, and the hypotheses (C1)–(C2) hold. Then the global weak solution U(t,x) to system (3.2) uniquely exists on (τ,T), which generates a strongly continuous process
Sε(t,τ): NV→NV, ∀ t≥τ,0≤ε<ε0 |
and Sε(t,τ)U(τ)=U(t).
Proof. The global well-posedness of solution can be obtained by the Galerkin approximation method, energy estimates and compact scheme. The detailed proof can be found in [15,18] and is omitted here.
● Existence of D-pullback attractors
Theorem 3.3. Assume U(τ)∈NV and the hypotheses (C1)–(C2) hold. Then the process Sε(t,τ): NV→NV generated by the system (3.2) possesses a minimal family of D-pullback attractors Aε={Aε(t)}t∈R in NV.
Proof. See Section 4.2.
When ε=0, the system (3.2) can be reduced to the following autonomous system
{∂∂t(u+αAu)+∫∞0μ(s)Aη(s)ds+B(u,u)+βu=PLf1(x), (t,x)∈Ωτ,∂∂tη=Tη+u,divu=0, (t,x)∈Ωτ,u(t,x)=0, (t,x)∈∂Ωτ,u(τ,x)=u(τ), x∈Ω,ητ(s)=∫s0φ(σ)dσ. | (3.4) |
Remark 3.1. The existence of global attractor A0 in NV can be achieved for the semigroup S0(t−τ) generated by (3.4).
● Robustness: upper semi-continuity of D-pullback attractors
Let ⋁ be a metric space, and {Aλ}λ∈⋁ is a family of subsets in X. Then it is said that {Aλ} has the property of upper semi-continuity as λ→λ0 in X if
limλ→λ0distX(Aλ,Aλ0)=0. |
The upper semi-continuity of attractors and related conclusions can be referred to [1,19,20,22] for more details.
In the following way, we intend to establish some results on the convergence between D-pullback attractors Aε to system (3.2) and global attractor A0 to system (3.4) as ε→0.
Theorem 3.4. Let U(τ)∈NV, Aε is the family of D-pullback attractors of Sε(t,τ) in NV to system (3.2), and A0 is the global attractor of S0(t−τ) in NV to system (3.4). Then the robustness of system is obtained by the following upper semi-continuity
limε→0distNV(Aε,A0)=0. |
Proof. See Section 5.
In this section, we first give the fundamental theory of attractors for dissipative systems, and the related conclusions can be seen in [1,2,3,7].
● Some relevant definitions
Definition 4.1. Assume that P(X) is the family of all nonempty subsets in a metric space X. If D is some nonempty class of families in the form ˆD={D(t):t∈R}⊂P(X), where D(t)⊂X is nonempty and bounded, then D is said to be a universe in P(X).
Definition 4.2. The family ˆD0={D0(t):t∈R}⊂P(X) is D-pullback absorbing for the process S(⋅,⋅) on X if for any t∈R and any ˆD∈D, there exists a τ0(t,ˆD)≤t such that
S(t,τ)D(τ)⊂D0(t), ∀ τ≤τ0(t,ˆD). |
Definition 4.3. A process S(⋅,⋅) on X is said to be D-pullback asymptotically compact if for any t∈R, any ˆD∈D, and any sequences {τn}⊂(−∞,t] and {xn}⊂X satisfying τn→−∞ and xn∈D(τn), the sequence {S(t,τn)xn} is relatively compact in X.
The D-pullback asymptotic compactness can be characterized by the Kuratowski measure of noncompactness κ(B) (B⊂X), relating definition and properties can be referred to [25,26], and the definition of D−κ-pullback contraction will be given as follows.
Definition 4.4. For any t∈R and ε>0, a process S(t,τ) on X is said to be D−κ-pullback contracting if there exists a constant TD(t,ε)>0 such that
κ(S(t,t−τ)D(t−τ))≤ε, ∀ τ≥TD(t,ε). |
Definition 4.5. A family A(t)={A(t)}t∈R is called the D-pullback attractors of process S(t,τ), if for any t∈R and any {D(t)}∈D, the following properties hold.
(i) A(t) is compact in X.
(ii) S(t,τ)A(τ)=A(t), t≥τ.
(iii) limτ→−∞distX(S(t,τ)D(τ),A(t))=0.
In addition, D-pullback attractor A is said to be minimal if whenever ˆC is another D-attracting family of closed sets, then A(t)⊂C(t) for all t∈R.
● Some conclusions
Theorem 4.6. ([1,27]) Let S(⋅,⋅):R2d×X→X be a continuous process, where R2d={(t,τ)∈R2|t≥τ}, D is a universe in P(X), and a family ˆD0={D0(t):t∈R}⊂P(X) is D-pullback absorbing for S(⋅,⋅), which is D-pullback asymptotically compact. Then, the family of D-pullback attractors AD={AD(t):t∈R} exists and
AD(t)=⋂s≥0¯⋃τ≥sS(t,t−τ)D(t−τ)X, t∈R. |
Remark 4.1. If ˆD0∈D, then AD is minimal family of closed subsets attracting pullback to D. It is said to be unique provided that ˆD0∈D, D0(t) is closed for any t∈R, and D is inclusion closed.
Theorem 4.7. ([21]) Assume that ˜D={ˆ˜D(t)} is a family of sets in X, S(⋅,⋅) is continuous, and, for any t∈R, there exists a constant T(t,D,˜D) such that
S(t,t−τ)D(t−τ)⊂˜D(t), ∀ τ≥T(t,D,˜D). |
If S(⋅,⋅) is D-pullback absorbing and ˆD−κ-pullback contracting, then the D-pullback attractors AD={AD(t):t∈R} exist for S(⋅,⋅).
Lemma 4.8. ([28]) Assume that S(⋅,⋅)=S1(⋅,⋅)+S2(⋅,⋅), ˜D={ˆ˜D(t)} is a family of subsets in X, and for any t∈R and any τ∈R+ there hold
(i) For any u(t−τ)∈˜D(t−τ),
‖S1(t,t−τ)u(t−τ)‖X≤Φ(t,τ)→0 (τ→+∞). |
(ii) For any T≥τ, ∪0≤τ≤TS2(t,t−τ)˜D(t−τ) is bounded, and S2(t,t−τ)˜D(t−τ) is relatively compact in X.
Then S(⋅,⋅) is ˆD−κ-pullback contracting in X.
From Theorems 3.2, we know that the system (3.2) generates a continuous process Sε(t,τ) in NV. To obtain the D-pullback attractors, we need to establish the existence of D-pullback absorbing set and the D-pullback asymptotic compactness of Sε(t,τ).
● Existence of D-pullback absorbing set in NV
Let D denote a family of all {D(t)}t∈R⊂P(NV) satisfying
limτ→−∞eωτsupU(τ)∈D(τ)J(2|U(τ)|2)=0, |
where ω=ω3/4,1,4,3,fε>0 and J(⋅)=J3/4,1,4,3,fε(⋅). Next, we establish the existence of D-pullback absorbing set.
Lemma 4.9. Let (u(τ),ητ)∈NV, then the process {Sε(t,τ)} to system (3.2) possesses a D-pullback absorbing set ˆDε0(t)={Dε0(t)}t∈R in NV, where
Dε0(t)=ˉBNV(0,ρεNV(t)), |
with radius
ρεNV(t)=√2√Λ3/4,1,4,3,fε(2C(|f1|2+εK)+1). | (4.1) |
Proof. Multiplying (3.2) by u, we have
12ddt‖u‖2Vα+∫∞0μ(s)Aη(s)u(t)ds+β|u|2=(PLfε,u), | (4.2) |
that is
12ddt‖u‖2Vα+∫∞0μ(s)Aη(s)(∂tη(s)+∂sη(s))ds+β|u|2=12ddt(‖u‖2Vα+‖η‖2MV)+12∫∞0μ(s)dds‖η‖2ds+β|u|2≤|(fε,u)|. | (4.3) |
Multiplying (3.2) by ut, we have
‖ut‖2Vα+((η,ut))MV+12βddt|u|2+b(u,u,ut)=(PLfε,ut). | (4.4) |
Then, the interpolation inequality and Young inequality lead to
βddt|u|2+2‖ut‖2Vα≤2|((η,ut))MV|+2|(fε,ut)|+2|b(u,u,ut)|≤2|((η,ut))MV|+2|(fε,ut)|+C‖u‖L3‖u‖‖u‖L6≤2|((η,ut))MV|+2|(fε,ut)|+C|u|1/2‖u‖1/2‖u‖‖u‖L6≤α‖ut‖2+C‖η‖2MV+C|u|‖u‖3+C|fε|2. | (4.5) |
To estimate the term ∫∞0μ(s)dds‖η‖2ds in (4.3) and avoid the possible singularity of μ at zero, we refer to [18] and construct the following new function
˜μ(s)={μ(˜s),0<s≤˜s,μ(s),s>˜s |
where ˜s is fixed such that ∫˜s0μ(s)ds≤κ/2. Also, if we set
Φ(t)=−4κ∫∞0˜μ(s)((η(s),u(t)))ds, |
then differentiating in t leads to
ddtΦ(t)+‖u‖2≤4μ(˜s)κ2∫∞0μ(s)dds‖η‖2ds+4ακε‖η‖2MV+αε‖ut‖2. | (4.6) |
We use the technique in [18] and set
yε(t)=E(t)+νεΦ(t)+ε2Ψ(t), |
where
E(t)=12(‖u‖2Vα+‖η‖2MV), Ψ(t)=2β|u|2. |
For sufficient small ε, it leads to
E(t)≤2yε(t)≤2(E(t)+1), |
where we choose ε0 satisfying
νε0supt∈[τ,T]Φ(t)+ε20supt∈[τ,T]Ψ(t)=1, |
and 0≤ε<ε0. Then, there holds
ddtyε(t)+Cεyε(t)≤Cε4yε(t)3+C|fε|2, | (4.7) |
and
yε(t)≤yε(τ)e−ε(t−τ)+Cε4∫tτe−ε(t−s)⋅1⋅yε(s)3ds+Csupt∈R∫t+1t|fε|2ε−1ds≤yε(τ)e−ε(t−τ)+Cε4∫tτe−ε(t−s)⋅1⋅yε(s)3ds+C(|f1|2+εK)ε−1. | (4.8) |
Then by Lemma 2.1, there exist
ω=ω3/4,1,4,3,fε>0, Λ=Λ3/4,1,4,3,fε>0, |
and an increasing function
J(⋅)=J3/4,1,4,3,fε(⋅): R+→R+ |
such that
E(t)≤J(2E(τ))e−ω(t−τ)+Λ(2C(|f1|2+εK)+1), |
which implies the conclusion holds.
Remark 4.2. For the semigroup S0(t−τ), it has the global absorbing set D00 in NV, where
D00={U∈NV; ‖U‖NV≤ρ0NV=√2√Λ3/4,1,4,3,f1(2C|f1|2+1)} | (4.9) |
and
lim supε→0ρεNV(t)=ρ0NV. | (4.10) |
● D−κ-pullback contraction of Sε(t,τ) in NV
To verify the pullback contraction of Sε(t,τ), we decompose Sε(t,τ) as follows
Sε(t,τ)U(τ)=Sε1(t−τ)U1(τ)+Sε2(t,τ)U2(τ)=:U1(t)+U2(t), |
which solve the following two problems respectively
{∂∂t(u1+αAu1)+∫∞0μ(s)Aη1(s)ds+B(u,u1)=0, (t,x)∈Ωτ,∂∂tη1=Tη1+u1,divu1=0, (t,x)∈Ωτ,u1(t,x)=0, (t,x)∈∂Ωτ,u1(τ,x)=u(τ), x∈Ω,ητ1(s)=∫s0φ(σ)dσ, | (4.11) |
and
{∂∂t(u2+αAu2)+∫∞0μ(s)Aη2(s)ds+B(u,u2)+βu2=PLfε−βu1, (t,x)∈Ωτ,∂∂tη2=Tη2+u2,divu2=0, (t,x)∈Ωτ,u2(t,x)=0, (t,x)∈∂Ωτ,u2(τ,x)=0, x∈Ω,ητ2(s)=0. | (4.12) |
Lemma 4.10. Let U(τ)∈Dε0(τ), then the solution Sε1(t−τ)U(τ) to the system (4.11) satisfies
‖Sε1(t−τ)U(τ)‖NV≤J(2E(τ))e−ω(t−τ)→0 (τ→−∞). |
Proof. Multiplying (4.11) by u1 and ∂∂tu1 respectively, and repeating the reasonings as shown as in Lemma 4.9, in which β=0 and fε=0, we can derive the conclusion finally. The parameter ω is dependent on ε and the increasing function J(∗) is different from the one in Lemma 4.9. Despite all this, these parameters can be unified in same representation, and the concrete details are omitted here.
Lemma 4.11. Let U(τ)∈Dε0(τ), then for any t∈R, there exists Cε(t)>0 such that the solution Sε2(t,τ)U(τ) to the system (4.12) satisfies
‖Sε2(t,τ)U(τ)‖NW≤Cε(t). |
Proof. Multiplying (4.12) by Au2, we have
12ddt(‖u2‖2+α|Au2|2)+∫∞0μ(s)Aη2(s)Au2(t)ds+β‖u2‖2+b(u,u2,Au2)=(PLfε−βu1,Au2), | (4.13) |
from the existence of pullback absorbing set, Lemma 4.10, the interpolation inequality and Young inequality, we have
12ddt(‖u2‖2+α|Au2|2+‖η2‖2MW)+12∫∞0μ(s)dds|Aη2|2ds+β‖u2‖2≤|(fε,Au2)|+|(βu1,Au2)|+‖u‖L6‖∇u2‖L3|Au2|≤νε4|Au2|2+C‖u‖‖u2‖1/2|Au2|1/2|Au2|+C|fε|2≤νε2|Au2|2+C|fε|2+C. | (4.14) |
Multiplying (4.12) by A∂tu2, we have
‖∂tu2‖2+α|A∂tu2|2+((η2,∂tu2))MW+12βddt‖u2‖2+b(u,u2,∂tu2)=(PLfε−βu1,A∂tu2). | (4.15) |
By the existence of pullback absorbing set, Lemma 4.10 and Young inequality, one has
βddt‖u2‖2+2‖∂tu2‖2+2α|A∂tu2|2≤2|((η2,∂tu2))MW|+2|(PLfε−βu1,A∂tu2)|+2|b(u,u2,∂tu2)|≤2|((η2,∂tu2))MW|+2|(PLfε−βu1,A∂tu2)|+C|Au2||A∂tu2|≤α|A∂tu2|2+C‖η2‖2MW+C|Au2|2+C|fε|2. | (4.16) |
To estimate the term ∫∞0μ(s)dds|Aη2|2ds in (4.14), we set
Φ2(t)=−6κ∫∞0˜μ(s)(Aη2(s),Au2(t))ds, |
and differentiating in t leads to
ddtΦ2(t)+6κ∫∞0˜μ(s)(Au2(t),Au2(t))ds≤−6κ∫∞0˜μ(s)(−A∂sη2,Au2(t))ds−6κ∫∞0˜μ(s)(Aη2(s),A∂tu2(t))ds, | (4.17) |
where
6κ∫∞0˜μ(s)(Au2(t),Au2(t))ds≥6κ∫∞˜s˜μ(s)ds⋅|Au2(t)|2, | (4.18) |
and
−6κ∫∞0˜μ(s)(−A∂sη2,Au2(t))ds=−6κ∫∞˜sμ′(s)(Aη2,Au2(t))ds≤6κ∫∞˜s−μ′(s)|Aη2||Au2(t)|ds≤6κ(∫∞˜s−μ′(s)|Aη2|2ds)1/2(∫∞˜s−μ′(s)|Au2(t)|2ds)1/2≤6κ(∫∞˜sμ(s)dds|Aη2|2ds)1/2(2μ(˜s))1/2|Au2(t)|≤|Au2(t)|2+18μ(˜s)κ2∫∞˜sμ(s)dds|Aη2|2ds, | (4.19) |
and
−6κ∫∞0˜μ(s)(Aη2(s),A∂tu2(t))ds≤6κ∫∞0μ(s)|Aη2(s)||A∂tu2(t)|ds≤αε|A∂tu2(t)|2+9ακ2ε‖η‖2MW. | (4.20) |
Also, from the fact that μ′(s)+δμ(s)≤0, we have
∫∞0μ(s)dds|Aη2|2ds≥∫∞0δμ(s)|Aη2|2ds=δ‖η2‖2MW. | (4.21) |
Thus
\begin{eqnarray} &&\frac{d}{dt}\Phi_2(t)+2|Au_2|^2\leq \frac{18\mu(\tilde{s})}{\kappa^2}\int_{0}^{\infty}\mu(s)\frac{d}{ds}|A\eta_2|^2ds+\frac{9}{\alpha\kappa\varepsilon}\|\eta_2\|_{M_W}^{2}+\alpha\varepsilon|A\partial_t u_2|^2. \end{eqnarray} | (4.22) |
We use the technique in [18] and set
z_{\varepsilon}(t) = E_2(t)+\nu\varepsilon\Phi_2(t)+\varepsilon^2\Psi_2(t), |
where
E_2(t) = \frac{1}{2}(\|u_2\|^{2}+\alpha|Au_2|^{2}+\|\eta_2\|_{M_W}^2), \ \Psi_2(t) = \beta\|u\|^2. |
For sufficient small enough \varepsilon , it leads to
E_2(t)\leq 2z_{\varepsilon}(t)\leq 2(E_2(t)+1), |
and there holds
\begin{eqnarray} \frac{d}{dt}z_{\varepsilon}(t)+\nu\varepsilon z_{\varepsilon}(t)\leq C+C|f_\varepsilon|^2, \end{eqnarray} | (4.23) |
it follows from the Gronwall lemma that
\begin{eqnarray} E_2(t)&\leq& Ce^{-\nu\varepsilon(t-\tau)}E_2(\tau)+C\varepsilon\int_{\tau}^{t}e^{\nu\varepsilon(s-t)}|f_2(s, x)|^2ds+C|f_1|^2+C\\ &\leq& C\varepsilon e^{-\nu\varepsilon t }\int_{\tau}^{t}e^{\nu\varepsilon_0 s}|f_2(s, x)|^2ds+C|f_1|^2+C, \end{eqnarray} | (4.24) |
which means the conclusion holds.
Above all, Lemmas 4.10, 4.11 and 4.8 lead to
Lemma 4.12. Let U(\tau)\in N_V , then the process S^{\varepsilon}(t, \tau):\ N_V\rightarrow N_V generated by the system (3.2) is \mathcal{D}-\kappa -pullback contracting in N_V .
Consequently, from Theorem 4.7, we can finish the proof of Theorem 3.3.
By the definition of upper semi-continuity, the following lemmas can be used to obtain the robustness of pullback attractors for evolutionary systems.
Lemma 5.1. ([20]) Let \varepsilon\in (0, \varepsilon_0] , \{S^{\varepsilon}(t, \tau)\} is the process of evolutionary system with non-autonomous term (depending on \varepsilon ), which is obtained by perturbing the semigroup S^0(\tau) of system without \varepsilon , and, for any t\in \mathbb{R} , there also hold that
(i) S^{\varepsilon}(t, \tau) has the pullback attractors \mathcal{A}^{\varepsilon}(t) , and \mathcal{A}^{0} is the global attractor for S^0(\tau) .
(ii) For any \tau\in \mathbb{R}^+ and any u\in X , there holds uniformly that
\lim\limits_{\varepsilon\rightarrow 0}\mathit{\mbox{d}}_X(S^{\varepsilon}(t, t-\tau)u, S^{0}(\tau)u) = 0. |
(iii) There exists a compact subset G\subset X such that
\lim\limits_{\varepsilon\rightarrow 0}\mathit{\mbox{dist}}_X(A^{\varepsilon}(t), G) = 0. |
Then, for any t\in \mathbb{R} , there holds
\lim\limits_{\varepsilon\rightarrow 0}dist_{X}(\mathcal{A}^{\varepsilon}, \mathcal{A}^{0}) = 0. |
Lemma 5.2. ([21]) For any t\in \mathbb{R} , \tau\in \mathbb{R}^+ , and \varepsilon\in (0, \varepsilon_0] , \widehat{D}^\varepsilon_{0}(t) = \{D^\varepsilon_{0}(t):t\in\mathbb{R}\} is the pullback absorbing set for S^{\varepsilon}(t, \tau) , and \widehat{C}_{0}^\varepsilon(t) = \{C_0^\varepsilon(t):t\in\mathbb{R}\} is a family of compact subsets in X . Assume that S^\varepsilon(\cdot, \cdot) = S^\varepsilon_1(\cdot, \cdot)+S^\varepsilon_2(\cdot, \cdot) , and there hold
(i) For any u_{t-\tau}\in D^\varepsilon_{0}(t-\tau) ,
\|S^\varepsilon_1(t, t-\tau)u_{t-\tau}\|_X\leq \Phi(t, \tau)\rightarrow 0\ (\tau\rightarrow \infty). |
(ii) For any T\geq \tau , \cup_{0\leq \tau\leq T} S^\varepsilon_2(t, t-\tau)D^\varepsilon_{0}(t-\tau) is bounded, and there exists a constant T_{\mathcal{D}^\varepsilon_{0}}(t) , independent of \varepsilon , such that
S^\varepsilon_2(t, t-\tau)D^\varepsilon_{0}(t-\tau)\subset C^\varepsilon_{0}(t), \ \forall\ \tau > T_{\mathcal{D}^\varepsilon_{0}}(t). |
(iii) There is a compact subset G\subset X such that
\lim\limits_{\varepsilon\rightarrow 0}\mathit{\mbox{dist}}_X(C_0^{\varepsilon}(t), G) = 0. |
Then, the process S^{\varepsilon}(t, \tau) has the pullback attractors \mathcal{A}^{\varepsilon}(t) , and
\lim\limits_{\varepsilon\rightarrow 0}dist_{X}(\mathcal{A}^{\varepsilon}, G) = 0. |
We give the following procedure to verify Theorem 3.4.
Lemma 5.3. Let (u^\varepsilon, \eta^\varepsilon) = S^{\varepsilon}(t, \tau)U(\tau) be the solution to system (3.2), and (u, \eta) = S^{0}(t-\tau)U(\tau) is the solution to system (3.4), then, for any bounded subset B\subset N_V , there holds
\lim\limits_{\varepsilon\rightarrow 0}\sup\limits_{U(\tau)\in B}d_{N_V}(S^{\varepsilon}(t, \tau)U(\tau), S^{0}(t-\tau)U(\tau)) = 0. |
Proof. We know
\begin{eqnarray} \left\{\begin{array}{ll} \frac{\partial}{\partial t}(u^\varepsilon+\alpha Au^\varepsilon)+\int_{0}^{\infty}\mu(s)A\eta^\varepsilon(s)ds+B(u^\varepsilon, u^\varepsilon)+\beta u^\varepsilon = P_L f_\varepsilon(t, x), \ (t, x)\in \Omega_{\tau}, \\ \frac{\partial}{\partial t}\eta^\varepsilon = T\eta^\varepsilon+u^\varepsilon, \end{array}\right. \end{eqnarray} | (5.1) |
and
\begin{eqnarray} \left\{\begin{array}{ll} \frac{\partial}{\partial t}(u+\alpha Au)+\int_{0}^{\infty}\mu(s)A\eta(s)ds+B(u, u)+\beta u = P_L f_1(x), \ (t, x)\in \Omega_{\tau}, \\ \frac{\partial}{\partial t}\eta = T\eta+u.\end{array}\right. \end{eqnarray} | (5.2) |
Let w^\varepsilon = u^\varepsilon-u and \xi^\varepsilon = \eta^\varepsilon-\eta , we can derive
\begin{eqnarray} \frac{\partial}{\partial t}(w^\varepsilon+\alpha Aw^\varepsilon)+\int_{0}^{\infty}\mu(s)A\xi^\varepsilon(s)ds+B(u^\varepsilon, w^\varepsilon)+B(w^\varepsilon, u)+\beta w^\varepsilon = \varepsilon P_L f_2(t, x), \end{eqnarray} | (5.3) |
and multiplying it by w^\varepsilon leads to
\begin{eqnarray} \frac{1}{2}\frac{d}{d t}\|w^\varepsilon\|_{\alpha}^{2}+\int_{0}^{\infty}\mu(s)(A\xi^\varepsilon(s), w^\varepsilon(t))ds+b(w^\varepsilon, u, w^\varepsilon)+\beta |w^\varepsilon|^2 = \varepsilon (P_L f_2(t, x), w^\varepsilon), \end{eqnarray} | (5.4) |
it follows that
\begin{eqnarray} &&\quad\frac{1}{2}\frac{d}{d t}\|w^\varepsilon\|_{\alpha}^{2}+\int_{0}^{\infty}\mu(s)((\xi^\varepsilon(s), w^\varepsilon))ds\leq |b(w^\varepsilon, u, w^\varepsilon)|+\varepsilon|(f_2(t, x), w^\varepsilon)|. \end{eqnarray} | (5.5) |
Integrating (5.5) over [\tau, t] , from Lemma 2.2 we derive that
\begin{eqnarray} &&\quad\|w^\varepsilon(t)\|_{\alpha}^{2}+\|\xi^\varepsilon(t-s)\|_{M_V}^{2}\\ &&\leq \|w^\varepsilon(\tau)\|_{\alpha}^{2}+\|\xi^{\varepsilon}(\tau-s)\|_{M_V}^{2}+ 2\int_{\tau}^{t}|b(w^\varepsilon, u, w^\varepsilon)|ds+2\varepsilon\int_{\tau}^{t}|(f_2(t, x), w^\varepsilon)|ds\\ &&\leq \|(w^\varepsilon, \xi^\varepsilon)|_\tau\|_{N_V}^{2}+ C\int_{\tau}^{t}\|u\|\|w^\varepsilon\|^2ds+2\varepsilon\int_{\tau}^{t}|(f_2(t, x), w^\varepsilon)|ds\\ &&\leq \|(w^\varepsilon, \xi^\varepsilon)|_\tau\|_{N_V}^{2}+\varepsilon^2\int_{\tau}^{t}|f_2|^2ds+C\int_{\tau}^{t}\|w^\varepsilon\|^2ds, \end{eqnarray} | (5.6) |
that is
\begin{eqnarray} &&\|(w^\varepsilon, \xi^\varepsilon)|_t\|_{N_V}^{2}\leq \|(w^\varepsilon, \xi^\varepsilon)|_\tau\|_{N_V}^{2}+\varepsilon^2\int_{\tau}^{t}|f_2|^2ds+C\int_{\tau}^{t}\|(w^\varepsilon, \xi^\varepsilon)|_s\|_{N_V}^{2}ds, \end{eqnarray} | (5.7) |
and the Gronwall inequality leads to
\begin{eqnarray} &&\|(w^\varepsilon, \xi^\varepsilon)|_t\|_{N_V}^{2}\leq C(\|(w^\varepsilon, \xi^\varepsilon)|_\tau\|_{N_V}^{2}+\varepsilon^2\int_{\tau}^{t}|f_2|^2ds)\rightarrow 0\ (\varepsilon\rightarrow 0), \end{eqnarray} | (5.8) |
which means that the conclusion is finished.
Proof of Theorem 3.4. From (4.24) and the fact that W\hookrightarrow V is compact, we know that there exists a compact subset G\subset N_V such that
\begin{eqnarray} \lim\limits_{\varepsilon\rightarrow 0}\mbox{dist}_X(C_0^{\varepsilon}(t), G) = 0. \end{eqnarray} | (5.9) |
Combining Lemma 4.10, Lemma 5.2 and (5.9), we have
\lim\limits_{\varepsilon\rightarrow 0}dist_{X}(\mathcal{A}^{\varepsilon}, G) = 0. |
In addition, the confirmation of condition (ii) in Lemma 5.1 is finished from Lemma 5.3, and we have
\lim\limits_{\varepsilon\rightarrow 0}dist_{X}(\mathcal{A}^{\varepsilon}, \mathcal{A}^{0}) = 0. |
Rong Yang was partially supported by the Science and Technology Project of Beijing Municipal Education Commission (No. KM202210005011).
The authors declare there is no conflict of interest.
[1] | World Health Organization, World Report on Vision, 2019. Available from: https://www.who.int/publications/i/item/world-report-on-vision. |
[2] | K. Han, Y. Liu, F. Yuan, The tactile graphics representation and its design application for the visually impaired, Design, 4 (2015), 18–19. |
[3] | D. Wu, J. Gan, Assistant haptic interaction technology for blind Internet user, J. Eng. Design, 17 (2010), 128–133. |
[4] | X. Li, Reading resources for the blind and the present situation of reading promotion in China, New Century Lib., 5 (2013), 19–22. |
[5] |
S. Spooner, "What page, Miss?" enhancing text accessibility with DAISY (Digital Accessible Information System), J. Visual Impairment Blindness, 108 (2014), 201–211. https://doi.org/10.1177/0145482X1410800304 doi: 10.1177/0145482X1410800304
![]() |
[6] |
S. Suksakulchai, S. Chaisanit, A daisy book production system over the internet, J. Commun. Comput., 8 (2011), 744–750. https://doi.org/10.1038/sj.bdj.4800904 doi: 10.1038/sj.bdj.4800904
![]() |
[7] | R. Zhang, Digital accessible information system and its development trend, Discovering Value, 6 (2010), 304. |
[8] |
A. Quint, Scalable vector graphics, IEEE Multimedia, 10 (2003), 99–102. https://doi.org/10.1109/MMUL.2003.1218261 doi: 10.1109/MMUL.2003.1218261
![]() |
[9] | G. Browne, Ebook indexes, EPUB and the international digital publishing forum, Online Curr., 26 (2012), 127–130. |
[10] | P. P. Rege, A. C. Chandrakar, Text-image separation in document images using boundary/perimeter detection, ACEEE Int. J. Signal Image Process., 3 (2012), 10–14. https://doi.org/01.IJSIP.03.01.70 |
[11] | P. Lyu, M. Liao, C. Yao, W. Wu, X. Bai, Mask textspotter: An end-to-end trainable neural network for spotting text with arbitrary shapes, in Proceedings of the European Conference on Computer Vision (ECCV), (2018), 67–83. https://doi.org/10.48550/arXiv.1807.02242 |
[12] | Y. Baek, B. Lee, D. Han, S. Yun, H. Lee, Character region awareness for text detection, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2019), 9365–9374. https://doi.org/10.1109/CVPR.2019.00959 |
[13] |
J. Memon, M. Sami, R. A. Khan, M. Uddin, Handwritten optical character recognition (OCR): A comprehensive systematic literature review (SLR), IEEE Access, 8 (2020), 142642–142668. https://doi.org/10.1109/ACCESS.2020.3012542 doi: 10.1109/ACCESS.2020.3012542
![]() |
[14] |
T. Hegghammer, OCR with Tesseract, Amazon Textract, and Google Document AI: a benchmarking experiment, J. Comput. Social Sci., 5 (2022), 861–882. https://doi.org/10.1007/s42001-021-00149-1 doi: 10.1007/s42001-021-00149-1
![]() |
[15] | X. Zhou, C. Yao, H. Wen, Y. Wang, S. Zhou, W. He, et al., East: an efficient and accurate scene text detector, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2017), 5551–5560. https://doi.org/10.1109/CVPR.2017.283 |
[16] | M. Suzuki, T. Kanahori, N. Ohtake, K. Yamaguchi, An integrated OCR software for mathematical documents and its output with accessibility, in International Conference on Computers for Handicapped Persons, (2004), 648–655. https://doi.org/10.1007/978-3-540-27817-7_97 |
[17] | M. Suzuki, Pen-Based Mathematical Computing Organizers, Stephen Watt and Clare So, (2005), 60. |
[18] |
Z. Wang, J. C. Liu, Translating math formula images to LaTeX sequences using deep neural networks with sequence-level training, Int. J. Doc. Anal. Recogn., 24 (2021), 63–75. https://doi.org/10.1007/s10032-020-00360-2 doi: 10.1007/s10032-020-00360-2
![]() |
[19] | O. Vinyals, A. Toshev, S. Bengio, D. Erhan, Show and tell: A neural image caption generator, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2015), 3156–3164. https://doi.org/10.1109/CVPR.2015.7298935 |
[20] | Y. Wang, J. Xu, Y. Sun, End-to-end transformer based model for image captioning, in Proceedings of the AAAI Conference on Artificial Intelligence, 36 (2022), 2585–2594. https://doi.org/10.1609/aaai.v36i3.20160 |
[21] |
T. Fuda, S. Omachi, H. Aso, Recognition of line graph images in documents by tracing connected components, Syst. Comput. Jpn., 38 (2007), 103–114. https://doi.org/10.1002/scj.10615 doi: 10.1002/scj.10615
![]() |
[22] |
S. Shimada, S. Kakumoto, M. Ejiri, A recognition algorithm of dashed and chained lines for automatic inputting of drawings, IEICE Trans. Inf. Syst., 18 (1986), 759–770. https://doi.org/10.1002/scj.4690180603 doi: 10.1002/scj.4690180603
![]() |
[23] | N. Yokokura, T. Watanabe, Recognition of business bar-graphs using layout structure knowledge, J. Inf. Process., 40 (1999), 2954–2966. |
[24] | M. Yo, T. Kawahara, R. Fukuda, Handwriting graphics input system for high school mathematics, IEICE Tech. Rep., 291 (2004), 43–48. |
[25] | A. Balaji, T. Ramanathan, V. Sonathi, Chart-text: A fully automated chart image descriptor, preprint, arXiv: 1812.10636. https://doi.org/10.48550/arXiv.1812.10636 |
[26] |
J. Chen, N. Takagi, Development of a method for extracting and recognizing graph elements in mathematical graphs for automating translation of tactile graphics, J. Jpn. Soc. Fuzzy Theory Intell. Inf., 26 (2014), 593–605. https://doi.org/10.3156/jsoft.26.593 doi: 10.3156/jsoft.26.593
![]() |
[27] |
J. Staker, K. Marshall, R. Abel, C. M. McQuaw, Molecular structure extraction from documents using deep learning, J. Chem. Inf. Model., 59 (2019), 1017–1029. https://doi.org/10.1021/acs.jcim.8b00669 doi: 10.1021/acs.jcim.8b00669
![]() |
[28] |
M. Oldenhof, A. Arany, Y. Moreau, J. Simm, ChemGrapher: optical graph recognition of chemical compounds by deep learning, J. Chem. Inf. Model., 60 (2020), 4506–4517. https://doi.org/10.1021/acs.jcim.0c00459 doi: 10.1021/acs.jcim.0c00459
![]() |
[29] | A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, et al., Attention is all you need, in Advances in Neural Information Processing Systems, 30 (2017), 5998–6008. https://doi.org/10.48550/arXiv.1706.03762 |
[30] | A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, et al., An image is worth 16x16 words: Transformers for image recognition at scale, preprint, arXiv: 2010.11929. https://doi.org/10.48550/arXiv.2010.11929 |
[31] | Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, et al., Swin transformer: Hierarchical vision transformer using shifted windows, in Proceedings of the IEEE/CVF International Conference on Computer Vision, (2021), 10012–10022. https://doi.org/10.1109/ICCV48922.2021.00986 |
[32] | Z. Liu, H. Mao, C. Wu, C. Feichtenhofer, T. Darrell, S. Xie, A convnet for the 2020s, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2022), 11976–11986. https://doi.org/10.1109/CVPR52688.2022.01167 |
[33] | M. Tan, R. Pang, Q. V. Le, Efficientdet: Scalable and efficient object detection, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2020), 10781–10790. https://doi.org/10.1109/CVPR42600.2020.01079 |
[34] | K. He, G. Gkioxari, P. Dollár, R. Girshick, Mask R-CNN, in Proceedings of the IEEE International Conference on Computer Vision, 42 (2017), 2961–2969. https://doi.org/10.1109/TPAMI.2018.2844175 |
[35] | L. Chen, Y. Zhu, G. Papandreou, F. Schroff, H. Adam, Encoder-decoder with atrous separable convolution for semantic image segmentation, in Proceedings of the European Conference on Computer Vision (ECCV), 40 (2018), 801–818. https://doi.org/10.1109/TPAMI.2017.2699184 |
[36] | J. Liu, W. Zhang, Y. Tang, J. Tang, G. Wu, Residual feature aggregation network for image super-resolution, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2020), 2359–2368. https://doi.org/10.1109/CVPR42600.2020.00243 |
[37] | S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: Towards real-time object detection with region proposal networks, in Advances in Neural Information Processing Systems, (2015), 28. https://doi.org/10.48550/arXiv.1506.01497 |
[38] | K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2016), 770–778. https://doi.org/10.1109/CVPR.2016.90 |
[39] | T. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature pyramid networks for object detection, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2017), 936–944. https://doi.org/10.1109/CVPR.2017.106 |
[40] | A. Kirillov, Y. Wu, K. He, R. Girshick, PointRend: Image segmentation as rendering, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2020), 9799–9808. https://doi.org/10.1109/CVPR42600.2020.00982 |
[41] | C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2016), 2818–2826. https://doi.org/10.1109/CVPR.2016.308 |
[42] | Cambridge International Examinations, IGCSE Guide Additional Mathematics Examples & Practice Book 1, Curriculum Planning and Development, 2021. |
[43] |
D. Bolya, C. Zhou, F. Xiao, Y. J. Lee, Yolact++: Better real-time instance segmentation, IEEE Trans. Pattern Anal. Mach. Intell., 44 (2020). https://doi.org/10.1109/TPAMI.2020.3014297 doi: 10.1109/TPAMI.2020.3014297
![]() |
[44] | S. Liu, L. Qi, H. Qin, J. Shi, J. Jia, Path aggregation network for instance segmentation, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2018), 8759–8768. https://doi.org/10.1109/CVPR.2018.00913 |
[45] | H. Chen, K. Sun, Z. Tian, C. Shen, Y. Huang, Y. Yan, Blendmask: Top-down meets bottom-up for instance segmentation, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2020), 8573–8581. https://doi.org/10.1109/CVPR42600.2020.00860 |
[46] | F. Li, H. Zhang, H. Xu, S. Liu, L. Zhang, L. M. Ni, et al., Mask dino: Towards a unified transformer-based framework for object detection and segmentation, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2022), 3041–3050. https://doi.org/10.48550/arXiv.2206.02777 |
[47] | R. R. Selvaraju, M. Cogswell, A. Das, R. V edantam, D. Parikh, D. Batra, Grad-cam: Visual explanations from deep networks via gradient-based localization, in Proceedings of the IEEE International Conference on Computer Vision, (2017), 618–626. https://doi.org/10.1007/s11263-019-01228-7 |
[48] | X. Zhu, H. Hu, S. Lin, J. Dai, Deformable convnets v2: More deformable, better results, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2019), 9308–9316. https://doi.org/10.1109/CVPR.2019.00953 |