Dynamic multi-frame image fusion algorithm and system
By using a dynamic multi-frame image fusion algorithm that combines Langevin equations and B-spline basis functions, the problems of insufficient Brownian motion compensation and resolution limitation in traditional methods are solved, achieving high signal-to-noise ratio and high resolution image fusion effects.
Patent Information
- Application Number
- CN202511017374.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-23
- Publication Date
- 2025-11-07
AI Technical Summary
In nanomaterial analysis and biomedical detection, existing technologies, such as traditional multi-frame image fusion methods, cannot effectively compensate for the randomness and nonlinear displacement of Brownian motion. They are limited in resolution and prone to noise accumulation. Existing improved schemes have high computational complexity and are difficult to achieve displacement compensation with nanometer-level precision.
A noise-aware dynamic weighting model is adopted, and a Brownian motion displacement model is established by combining the Langevin equation. Particle detection and tracking are performed by LoG speckle detection and Kalman filter. Image registration is performed by combining B-spline basis functions. An adaptive weight superposition model is constructed to achieve spatiotemporal adaptive adjustment and high-resolution output.
It significantly improves the signal-to-noise ratio and increases the resolution to the 100nm level, achieving high-precision image fusion. It can effectively compensate for errors caused by Brownian motion and generate high-quality, high-resolution images.
Smart Images
Figure CN120912450A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of image processing, and particularly relates to a dynamic multi-frame image fusion algorithm and system. BACKGROUND
[0002] In the field of nanomaterial analysis and biomedical detection, microscopic imaging technology is faced with the key problems of image blurring and low signal-to-noise ratio caused by Brownian motion of nanoparticles. The traditional multi-frame superposition method only improves the signal-to-noise ratio by simple alignment and averaging, but has the following defects:
[0003] 1) Insufficient motion compensation: Brownian motion is random, and traditional rigid registration cannot compensate for nonlinear displacement and fluid perturbation;
[0004] 2) Resolution is limited: the resolution of a single frame image is limited by the optical diffraction limit (about 1 μm), and the physical resolution cannot be broken after superposition;
[0005] 3) Noise amplification risk: fixed weight superposition is easy to cause noise accumulation in the motion blurred area.
[0006] The existing improved scheme (such as non-rigid registration based on optical flow) has high computational complexity, and does not combine the physical motion model, making it difficult to achieve nanoscale precision displacement compensation. Therefore, there is an urgent need for a dynamic multi-frame fusion method that combines physical models and data-driven methods. SUMMARY
[0007] The purpose of the present application is to overcome the above-mentioned deficiencies, and the purpose of the present application is to provide a dynamic multi-frame image fusion algorithm and system, which uses a noise-aware dynamic weight model, and adjusts the weight in the space-time domain adaptively, significantly improves the signal-to-noise ratio, and is superior to the traditional average superposition; At the same time, by using a physical-data dual driving framework, the Brownian motion physical model is embedded into the image registration process, and the accuracy of the pure data-driven method is improved.
[0008] Technical scheme: In order to achieve the above-mentioned purpose, the present application provides a dynamic multi-frame image fusion algorithm, comprising the following steps:
[0009] S1) input N frames of images;
[0010] S2) precise alignment of multi-time sequence data is realized by estimating and compensating the random drift caused by Brownian motion; Ensure that the images collected at different time points are highly consistent in space, and provide a high-quality data basis for subsequent processing;
[0011] S201) particle motion modeling, a Brownian motion displacement model is established by using Langevin equation, and the formula is as follows:
[0012]
[0013] Where D is the diffusion coefficient, T, η, and r are the real-time measured temperature, viscosity, and estimated particle size, respectively, ξ(t) is a standard normally distributed random variable, and k B The diffusion coefficient D is the Boltzmann constant; it is calculated using real-time measured temperature T, viscosity η, and estimated particle size r; the Langevin equation is introduced to establish a Brownian motion displacement model, which can accurately describe the random motion of particles; the diffusion coefficient D is calculated using real-time measured temperature T, viscosity η, and estimated particle size r, which improves the adaptability and accuracy of the model.
[0014] S202): Multi-frame displacement prediction; Ignoring features below the threshold, the first frame uses the LoG speckle detection algorithm to detect particles, extracts particle coordinates and radii, performs state vector modeling, and combines a Kalman filter for cross-frame prediction; Combining the LoG speckle detection algorithm and the Kalman filter for cross-frame prediction can detect and track particle trajectories with high accuracy.
[0015] S3): Overall rigidity compensation; eliminates rigid displacement caused by camera shake or object movement, further improving the accuracy of image alignment;
[0016] S4): Subpixel-level image registration, which is achieved by constructing a local deformation field through B-spline basis function interpolation and optimizing the solution to perform elastic deformation compensation;
[0017] S5): Construct an adaptive weight stacking model to obtain high-resolution output;
[0018] S501): Frame quality assessment;
[0019] S502): Design fuzzy-noise joint optimization weights to achieve dynamic weight allocation, which can automatically optimize weight allocation according to the specific features of the image and improve the quality of the fusion result;
[0020] S503): Design a spatially variable weighted superposition formula to achieve non-uniform fusion and obtain high-resolution output. By using the spatially variable weighted superposition formula to achieve non-uniform fusion, the weights can be dynamically adjusted according to the local image quality to generate high-quality, high-resolution output images.
[0021] Furthermore, the frame quality assessment in S501 includes local contrast, gradient energy, and signal-to-noise ratio estimation;
[0022] The formula for local contrast is as follows:
[0023] C(i,j)=(I max -I min ) / (I max +I min )
[0024] Among them, I maxand I min respectively the maximum and minimum pixel values in the local region of the image;
[0025] The gradient energy is calculated using the Sobel operator; the signal-to-noise ratio estimate is based on wavelet domain threshold denoising residual calculation. By calculating the ratio of the maximum and minimum pixel values in the local region of the image, the local details and texture information of the image can be effectively reflected; at the same time, by combining local contrast, gradient energy and signal-to-noise ratio estimate, the image quality can be comprehensively evaluated from multiple angles.
[0026] Further, the S502 blur-noise joint optimization weight: formula as follows:
[0027]
[0028] Where, α, β, γ are weight coefficients, calibrated by particle swarm optimization algorithm; G n is the gradient energy; is the standard deviation of the noise signal; ε is set to 1×e -6 Prevent division by zero. At the same time, the blur and noise are optimized, and the image quality is comprehensively improved; and by dynamically adjusting the weight coefficients α, β and γ through the particle swarm optimization algorithm, the optimal weight combination can be automatically found according to the characteristics and needs of the specific image, so as to realize adaptive optimization.
[0029] Further, the S503 empty variable weighting superposition formula is as follows:
[0030]
[0031] Where, I fused (x, y) is the high-resolution output, w n (x, y) is the weight of the nth input image at position (x, y), I n ′(x, y) is the registered image. Using the registered image as input ensures that all input images are highly consistent in space, eliminating alignment errors caused by camera shake, object motion or viewpoint changes, thereby improving the quality of the fusion result. At the same time, multi-temporal data fusion effectively integrates image information collected at different time points, generating a more complete and accurate high-resolution image.
[0032] Further, in the S202, the particle coordinates (x i ,y i ) and radius r i are extracted by LoG spot detection algorithm;
[0033] When modeling the state vector, define the particle state vector as follows:
[0034] S i= [x i , y i , vx i , vy i , r i ] T
[0035] wherein x i , y i are the position coordinates of the particle, v is the velocity component of the particle, and r i is the radius of the particle; the coordinates and radius of the particle are extracted by a LoG spot detection algorithm, which can accurately locate and identify the particle; the LoG algorithm combines Gaussian smoothing and Laplace operator, which can effectively suppress noise and detect the center position and size of the particle;
[0036] The cross-frame prediction formula combined with the Kalman filter is as follows:
[0037] S k|k-1 = F·S k-1 + W
[0038] Z k = H·S k + V
[0039] wherein F is a state transition matrix, the state transition matrix F includes a Brownian motion diffusion model; W is a process noise, S k|k-1 is a state prediction value at time k, S k-1 is a state variable at time k-1, Z k is an observation value at time k, H and V are observation matrix and observation noise respectively; the process noise W and the observation noise V are dynamically adjusted according to the inter-frame correlation; the observation matrix H is designed according to the ability of the observation device, and depends on the state variable S k and the observation value Z k .Through the Kalman filter for cross-frame prediction, the state of the particle can be updated in real time, and the accuracy and robustness of tracking can be improved; at the same time, the state transition matrix includes a Brownian motion diffusion model, which can simulate the random motion of the particle and is suitable for dynamic tracking of micro-particles.
[0040] Further, the local deformation field formula for constructing B-spline basis function interpolation in S4 is as follows:
[0041] T(x, y) = å a ij ·B i (x)·B j (y)
[0042] wherein a ij is the weight in the control point matrix, B i (x) and B j(y) are B-spline basis functions in x and y directions respectively; the local deformation field is constructed by B-spline basis functions, complex non-rigid deformation can be flexibly represented, and the local support of B-spline basis functions makes the adjustment of the deformation field in a local area not affect other areas, so that high-precision local deformation modeling is realized;
[0043] The optimization solution aims to minimize the registration error, and is solved by using an L-BFGS algorithm, and the formula is as follows:
[0044]
[0045] Wherein, I n (T(x, y)) is the target image after deformation, I ref (x, y) is the reference image, Reg(T) is a regularization term, and lambda is a regularization parameter. The optimization goal is to minimize the registration error between the target image and the reference image, and the square error term is used to ensure that the target image after deformation is highly consistent with the reference image. At the same time, the regularization term and the regularization parameter are introduced, which can effectively prevent overfitting and ensure the smoothness and stability of the deformation field.
[0046] The application also provides a dynamic multi-frame image fusion system for realizing the dynamic multi-frame image fusion algorithm, which comprises a preprocessing module, an iterative optimization module and an output module; the preprocessing module, the iterative optimization module and the output module are sequentially connected; the preprocessing module realizes real-time input of sensor data (T, eta) and calibration of the diffusion coefficient D at the same time; the iterative optimization module realizes accurate alignment of multi-time sequence data, overall rigid compensation, sub-pixel level image registration and construction of a self-adaptive weight superposition model; and the output module generates a high-resolution image and a particle size distribution statistical report. Each module is responsible for a specific task, facilitating development, debugging and maintenance; meanwhile, a certain module can be independently upgraded or replaced without the need to redesign the whole system, and each module can be independently developed and tested, facilitating reuse in other projects. It can be seen from the above technical solution that the application has the following beneficial effects:
[0047] 1. The dynamic multi-frame image fusion algorithm and system combine global motion prediction of the Langevin equation and local deformation correction of B-spline, and the compensation error is less than 0.2 pixels (equivalent to 40nm).
[0048] 2. The dynamic multi-frame image fusion algorithm and system use a noise perception dynamic weight model, and the weight is adaptively adjusted in space-time domain, so that the signal-to-noise ratio is significantly improved, which is better than that of the traditional average superposition.
[0049] 3. The dynamic multi-frame image fusion algorithm and system use a physical-data dual driving framework, embed a Brownian motion physical model into the image registration process, and improve the accuracy of the pure data driven method. Attached Figure Description
[0050] Figure 1 This is a flowchart illustrating the steps of a dynamic multi-frame image fusion algorithm described in this invention.
[0051] Figure 2 This is a schematic diagram of the overall process of a dynamic multi-frame image fusion algorithm described in this invention;
[0052] Figure 3 This is an example image of the original unit frame in the dynamic multi-frame image fusion algorithm described in this invention;
[0053] Figure 4 This is an example image of a dynamically fused unit frame in a dynamic multi-frame image fusion algorithm described in this invention. Detailed Implementation
[0054] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0055] Example 1
[0056] In this embodiment, as Figure 1 and Figure 2 This invention discloses a dynamic multi-frame image fusion algorithm, comprising the following steps:
[0057] S1): Input N frames of images;
[0058] S2): Achieve accurate alignment of multi-time series data by estimating and compensating for random drift caused by Brownian motion;
[0059] S201): Particle motion modeling, using the Langevin equation to establish a Brownian motion displacement model, the formula is as follows:
[0060]
[0061] Where D is the diffusion coefficient, T, η, and r are the real-time measured temperature, viscosity, and estimated particle size, respectively, ξ(t) is a standard normally distributed random variable, and k B η is the Boltzmann constant; the diffusion coefficient D is calculated using real-time measured temperature T, viscosity η, and estimated particle size r.
[0062] S202) Multi-frame displacement prediction; ignore features below the threshold, the first frame uses LoG spot detection algorithm for particle detection, extracts particle coordinates and radius, then state vector modeling is performed, and Kalman filter is combined for cross-frame prediction;
[0063] S3) Overall rigid compensation;
[0064] S4) Sub-pixel level image registration, a local deformation field is constructed by constructing a B-spline basis function interpolation, and elastic deformation compensation is optimized and solved;
[0065] S5) Construct an adaptive weight superposition model to obtain a high-resolution output;
[0066] S501) Frame quality evaluation;
[0067] S502) Design of fuzzy-noise joint optimization weight to realize dynamic weight distribution;
[0068] S503) Design of space-variable weighted superposition formula to realize non-uniform fusion and obtain a high-resolution output.
[0069] Specifically, in S1, N frames of images can be read from a sensor or a storage device, while ensuring the time sequence order and timestamp information of the images; At the same time, Gaussian smoothing processing can be performed on the input images to reduce the influence of noise and improve the robustness of the entire process.
[0070] Specifically, in S202, features below the threshold can be set to features below 0.3 times the maximum value for ignoring operation.
[0071] Specifically, in S3, overall rigid compensation is achieved by estimating and compensating overall rigid motion to eliminate rigid displacement caused by camera jitter or object motion; As an option, a global rigid transformation model (such as translation and rotation) can be used to align the images.
[0072] In this embodiment, the frame quality evaluation in S501 includes local contrast, gradient energy and signal-to-noise ratio estimation;
[0073] The local contrast formula is as follows:
[0074] C(i, j) = (I max -I min ) / (I max +I min )
[0075] Where I max and I min are the maximum and minimum pixel values in the local region of the image, respectively;
[0076] The gradient energy is calculated using a Sobel operator; and the SNR estimation is based on wavelet domain threshold denoising residual calculation.
[0077] Specifically, a 3x3 window is selected as a preferred image local region.
[0078] Specifically, the Sobel operator calculation formula is as follows:
[0079]
[0080] wherein, and are the gradients of the image in the horizontal direction (x direction) and the vertical direction (y direction), respectively.
[0081] In this embodiment, the blur-noise joint optimization weight in S502 is as follows:
[0082]
[0083] wherein, α, β, γ are weight coefficients, which are calibrated by a particle swarm optimization algorithm; G n is the gradient energy; is the standard deviation of the noise signal; ε is set to 1×e -6 to prevent division by zero.
[0084] Specifically, the particle swarm optimization algorithm is used to dynamically adjust the weight coefficients α, β and γ, and the steps are as follows:
[0085] (1) Initialize the particle swarm: each particle represents a set of weight coefficients (α, β, γ);
[0086] (2) Define the fitness function as the optimization objective:
[0087] Fitness(α,β,γ)=α·G+β·σ noise +γ·∈
[0088] wherein, G is the gradient energy, is the standard deviation of the noise signal, and ε is set to 1×e -6 to prevent division by zero.
[0089] (3) Particle update: update the position and velocity of each particle, and evaluate the performance of each particle according to the fitness function;
[0090] (4) Update the global optimal solution and individual optimal solution;
[0091] (5) Iterative optimization: repeat the above steps until a predetermined number of iterations is reached or the fitness is no longer significantly improved.
[0092] In this embodiment, the formula for the empty variable weighting superposition in S503 is as follows:
[0093]
[0094] where I fused (x,y) is the high resolution output, w n (x,y) is the weight of the n-th input image at position (x,y), I n ′(x,y) is the registered image.
[0095] Specifically, the weight map spatial resolution is set to 64x64 partitions as a preferred.
[0096] In this embodiment, the particle coordinates (x i ,y i ) and radius r i are extracted by the LoG spot detection algorithm in S202.
[0097] When modeling the state vector, the particle state vector is defined as follows:
[0098] S i =[x i ,y i ,vx i ,vy i ,r i ] T
[0099] where x i , y i are the position coordinates of the particle, v is the velocity component of the particle, and r i is the radius of the particle.
[0100] The cross-frame prediction formula combined with the Kalman filter is as follows:
[0101] S k|k-1 =F·S k-1 +W
[0102] Z k =H·S k +V
[0103] where F is a state transition matrix, the state transition matrix F includes a Brownian motion diffusion model; W is a process noise, S k|k-1 is a state prediction value at time k, S k-1 is a state variable at time k-1, Z k is an observation value at time k, H and V are an observation matrix and an observation noise, respectively; the process noise W and the observation noise V are dynamically adjusted according to the inter-frame correlation; the observation matrix H is designed according to the ability of the observation device and depends on the state variable S k and the observation value Z k .
[0104] Specifically, the following is an exemplary scenario to derive the observation matrix H, the particle state vector is as follows:
[0105] S i =[x i ,y i ,vx i ,vy i ,r i ] T
[0106] Assume that the observation device can only measure the position (x i ,y i ) and radius r i of the particle, so the observation vector Z k is:
[0107] Z k =[x i ,y i ,r i ] T
[0108] The role of the observation matrix H is to extract the observation vector from the state vector. According to the definition of the observation vector, the observation matrix H should extract x i ,y i and r i from the state vector, and ignore the velocity components vx i and vy i , at this time, the observation matrix H is:
[0109]
[0110] In this embodiment, the local deformation field formula for constructing B-spline basis function interpolation in S4 is as follows:
[0111] T(x, y) = ∑α ij ·B i (x)·B j (y)
[0112] Wherein, α ij is the weight in the control point matrix, B i (x) and B j (y) are B-spline basis functions in x direction and y direction respectively;
[0113] The optimization solution aims to minimize the registration error, and uses L-BFGS algorithm for solution, the formula is as follows:
[0114]
[0115] Wherein, I n(T(x, y)) is the target image after deformation, I ref (x, y) is the reference image, Reg(T) is a regularization term, and λ is a regularization parameter.
[0116] Specifically, the regularization term Reg(T) is used to constrain the smoothness of the deformation field, and the regularization parameter λ can be adaptively adjusted according to the local gradient.
[0117] Specifically, the regularization term selects gradient penalty as an option, which constrains the gradient of the deformation field to make it change smoothly, and the formula is as follows:
[0118]
[0119] Embodiment 2
[0120] In this embodiment, on the basis of embodiment 1, the application further discloses a dynamic multi-frame image fusion system for implementing the dynamic multi-frame image fusion algorithm, which comprises a preprocessing module, an iterative optimization module and an output module; the preprocessing module, the iterative optimization module and the output module are connected in sequence; the preprocessing module realizes real-time input of sensor data (T, η) and simultaneously calibrates the diffusion coefficient D; the iterative optimization module realizes accurate alignment of multi-time sequence data, overall rigid compensation, sub-pixel level image registration and construction of an adaptive weight superposition model; and the output module generates a high-resolution image and a particle size distribution statistical report.
[0121] Specifically, the system uses GPU parallel processing technology, and each thread block processes 256 particles; at the same time, the OpenCL technology is used to realize pipeline processing of image registration, and through block processing (Tile Size = 512x512), the system can efficiently process large-scale image data. In order to ensure efficient transmission of data between different modules, the real-time transmission adopts a PCIe 4.0x16 interface, and combines with zero-copy memory transmission technology to directly process data in memory, avoiding unnecessary data copying and further improving the efficiency of data transmission.
[0122] In embodiments 1 and 2, the application discloses a dynamic multi-frame image fusion algorithm and system, and comparison of the original unit frame and the dynamic fused unit frame can obtain (as shown in Figure 3 and Figure 4 The following beneficial effects are achieved:
[0123] 1) Resolution improvement: the equivalent resolution of the output image reaches 100nm level;
[0124] 2) Real-time: the 4K image processing speed can meet the real-time requirement;
[0125] 3) Quantitative accuracy: particle size distribution statistical error and edge sharpness can meet the actual requirements;
[0126] 4) Extending applicability: can be migrated to live cell observation, microfluidic chip detection, etc.
[0127] The above only describes the preferred embodiments of the present application, and it should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements can be made, and these improvements should also be considered as the protection scope of the present application.
Claims
1. A dynamic multi-frame image fusion algorithm, characterized in that: It comprises the following steps: S1): inputting N frames of images; S2): realizing accurate alignment of multi-time sequence data by estimating and compensating random drift caused by Brownian motion; S201): modeling particle motion, establishing a Brownian motion displacement model by using a Langevin equation, and the formula is as follows: where D is a diffusion coefficient, T, η, r are respectively a real-time measured temperature, viscosity and estimated particle size, ξ(t) is a standard normal distribution random variable, k B is a Boltzmann constant; the diffusion coefficient D is calculated by a real-time measured temperature T, viscosity η and estimated particle size r; S202): multi-frame displacement prediction; ignoring features below a threshold, using a LoG spot detection algorithm for particle detection in the first frame, extracting particle coordinates and radius, then modeling a state vector, and combining a Kalman filter for cross-frame prediction; S3): overall rigidity compensation; S4): sub-pixel level image registration, constructing a local deformation field by B-spline basis function interpolation, and optimizing and solving for elastic deformation compensation; S5): constructing an adaptive weight superposition model to obtain high-resolution output; S501): frame quality evaluation; S502): designing a blur-noise joint optimization weight to realize dynamic weight distribution; S503): designing a space-variable weighted superposition formula to realize non-uniform fusion and obtain high-resolution output.
2. The dynamic multi-frame image fusion algorithm of claim 1, wherein: The frame quality evaluation in S501 comprises local contrast, gradient energy and signal-to-noise ratio estimation; The local contrast formula is as follows: C(i,j) = (I max - I min ) / (I max + I min ) where I max and I min are the maximum and minimum pixel values in the local region of the image, respectively. The gradient energy is calculated by using a Sobel operator; and the signal-to-noise ratio estimation is based on wavelet domain threshold denoising residual error calculation.
3. The dynamic multi-frame image fusion algorithm of claim 1, wherein: The blur-noise joint optimization weight in S502 is as follows: wherein a, β, γ are weight coefficients, calibrated by particle swarm optimization algorithm; G n is the gradient energy; is the standard deviation of the noise signal; ε is set to 1 x e -6 Prevent division by zero.
4. The dynamic multi-frame image fusion algorithm of claim 1, wherein: The space-variable weighted superposition formula in S503 is as follows: where I fused (x, y) is the high resolution output, w n (x, y) is the weight of the n-th input image at position (x, y), I n ′(x, y) is the registered image.
5. The dynamic multi-frame image fusion algorithm of claim 1, wherein: The particle coordinates (x i ,y i ) and radius r i are extracted by the LoG spot detection algorithm in S202. i i i When modeling the state vector, the particle state vector is defined as follows: S i = [x i , y i , vx i , vy i , r i ] T where x i , y i are the position coordinates of the particle, v is the velocity component of the particle, and r i is the radius of the particle. The cross-frame prediction formula combining the Kalman filter is as follows: S k|k-1＝ F·S k-1 +W Z k = H · S k + V Wherein, F is a state transition matrix, the state transition matrix F contains a Brownian motion diffusion model; W is a process noise, S k|k-1 is a state prediction value at time k, S k-1 is a state variable at time k-1, Z k is an observation value at time k, H and V are an observation matrix and an observation noise respectively; the process noise W and the observation noise V are dynamically adjusted according to inter-frame correlation; the observation matrix H is designed according to the capability of an observation device, and depends on the state variable S k and the observation value Z k .
6. The dynamic multi-frame image fusion algorithm of claim 1, wherein: The formula for constructing the B-spline basis function interpolation local deformation field in S4 is as follows: T(x, y) = ∑a ij • B i (x) • B j (y) where a ij are the weights in the control point matrix, B i (x) and B j (y) are the B-spline basis functions in the x and y directions, respectively. The optimization solving takes minimizing registration error as the target, and uses an L-BFGS algorithm for solving, and the formula is as follows: where I n (T(x, y)) is the deformed target image, I ref (x, y) is the reference image, Reg(T) is the regularization term, and λ is the regularization parameter.
7. A dynamic multi-frame image fusion system for implementing the dynamic multi-frame image fusion algorithm of claims 1-6, characterized in that: It comprises a preprocessing module, an iterative optimization module and an output module; The preprocessing module, the iterative optimization module and the output module are connected in sequence; the preprocessing module realizes real-time input of sensor data (T, η) and simultaneously calibrates the diffusion coefficient D; The iterative optimization module realizes accurate alignment of multi-time sequence data, overall rigidity compensation, sub-pixel level image registration and construction of an adaptive weight superposition model; and the output module generates a high-resolution image and a particle size distribution statistical report.