Coded imaging reconstruction method and system based on coupling of neural network and genetic algorithm

Abstract

The invention discloses a coding imaging reconstruction method and system based on coupling of a neural network and a genetic algorithm, and relates to the technical field of coding image reconstruction. The method comprises the following steps: inputting a coded image of a scene to be reconstructed into a pre-trained preliminary reconstruction convolutional neural network to obtain an initial source distribution image of the scene to be reconstructed, and disturbing the initial source distribution image to obtain multiple candidate source distribution images; the multiple candidate source distribution images serve as an initial population, the maximum fitness function value serves as a target, optimization is conducted on the source distribution image of the scene to be reconstructed through an improved genetic algorithm, and a target source distribution image of the scene to be reconstructed is obtained; the improved genetic algorithm comprises the step of cooperatively guiding population breeding by adopting residual convolutional neural network guided search and traditional genetic operation; and the residual convolutional neural network is used for learning a residual change rule of the source distribution image in an iterative optimization process. According to the method, the imaging speed can be remarkably improved while the reconstruction precision of the coded image is ensured.

Description

Technical Field

[0001] This invention relates to the field of coded image reconstruction technology, and in particular to a coded imaging reconstruction method and system based on the coupling of neural networks and genetic algorithms. Background Technology

[0002] In recent years, coded imaging technology, with its high signal-to-noise ratio and good spatial resolution, has been widely used in X-ray / gamma-ray astronomical observation, inertial confinement fusion diagnostics, nuclear facility decommissioning, and medical imaging. With the increasing demands for intelligent sensing capabilities across industries, developing high-performance, adaptable novel radiation-coded imaging systems has become an important task. In this context, the performance evaluation of imaging systems largely depends on the quality of the reconstructed images.

[0003] However, existing reconstruction methods often struggle to simultaneously meet the requirements of high precision and high speed, resulting in significant limitations in real-world imaging scenarios. Summary of the Invention

[0004] Therefore, it is necessary to provide a coded image reconstruction method and system based on the coupling of neural networks and genetic algorithms to address the above-mentioned technical problems. This method can simultaneously meet the requirements of high accuracy and high speed in coded image reconstruction.

[0005] The present invention adopts the following technical solution: This invention provides a coding imaging reconstruction method based on the coupling of neural networks and genetic algorithms, comprising: The encoded image of the scene to be reconstructed is input into a pre-trained convolutional neural network for initial reconstruction to obtain the initial source distribution image of the scene to be reconstructed. The initial source distribution image is then perturbed to obtain multiple candidate source distribution images. Multiple candidate source distribution images are used as the initial population, and an improved genetic algorithm is used to optimize the source distribution images of the scene to be reconstructed, with the goal of maximizing the fitness function value. The improved genetic algorithm includes using a residual convolutional neural network to guide the search and traditional genetic operations to guide the population reproduction. The residual convolutional neural network is used to learn the residual change law of the source distribution images during the iterative optimization process. The fitness function represents the difference between the predicted encoded image corresponding to the population of the improved genetic algorithm predicted by the pre-constructed cascaded convolutional neural network and the encoded image to be reconstructed.

[0006] Optionally, the initial source distribution image is perturbed to obtain multiple candidate source distribution images, including: By adding Gaussian noise to the initial source distribution image or performing local random transformation or translation transformation, a variety of candidate source distribution images can be obtained.

[0007] Optionally, multiple candidate source distribution images are used as the initial population, and an improved genetic algorithm is used to optimize the source distribution images of the scene to be reconstructed, with the goal of maximizing the fitness function value, to obtain the target source distribution image of the scene to be reconstructed, including: The source distribution image is randomly initialized, and the randomly initialized source distribution image and multiple candidate source distribution images are used as the initial population. The location of each individual in the initial population is input into a cascaded convolutional neural network to obtain the predicted encoded image of each individual; Based on the predicted encoded image and the encoded image to be reconstructed for each individual, the root mean square error is calculated, and the difference between 1 and the root mean square error is determined as the fitness value. Based on the fitness value, the optimal individual is determined, and individuals with fitness greater than a preset threshold are updated by residual guidance through a residual convolutional neural network, while individuals with fitness less than or equal to the preset threshold are updated by selection, crossover and mutation operations of a traditional genetic algorithm. Perform elite operations on the updated individuals to generate new individual positions; Calculate the fitness value based on the new individual position, and iteratively update the individual position based on the fitness value until the convergence condition is met; The optimal individual position after reaching the convergence condition is determined as the target source distribution image of the scene to be reconstructed.

[0008] Optionally, residual guidance is performed using a residual convolutional neural network to update the individual position, including: The individual's position is input into the residual convolutional neural network to obtain the residual map; the residual map is the residual between the current individual and the better solution after the genetic algorithm iteration; Update the individual positions based on the residual plot.

[0009] Optionally, the formula for updating the individual position based on the residual map is: ; in, For the original individual position, This is the residual map output by the residual convolutional neural network. The location of the new individual generated after mutation guided by the residual convolutional neural network. This is the strength coefficient.

[0010] Optionally, the initial reconstruction of the training process of the convolutional neural network includes: The Monte Carlo simulation generates paired data of multiple sets of sample encoded images and real source distribution images, and performs noise addition and data augmentation operations on the sample encoded images to obtain a training dataset in a standardized format. A preliminary reconstructed convolutional neural network is constructed using the training dataset.

[0011] Optionally, the structure of the initial reconstructed convolutional neural network is in the form of multiple U-Net or ResNet encoders and decoders, which are formed by connecting multiple sub-networks in a cascaded manner; the structure of the cascaded convolutional neural network is the same as that of the initial reconstructed convolutional neural network, and the encoder part of the cascaded convolutional neural network shares weights with the initial reconstructed convolutional neural network.

[0012] This invention provides a coded imaging reconstruction system based on the coupling of neural networks and genetic algorithms, comprising: The reconstruction processing module is used to input the encoded image of the scene to be reconstructed into a pre-trained preliminary reconstruction convolutional neural network to obtain the initial source distribution image of the scene to be reconstructed. The initial source distribution image is then perturbed to obtain multiple candidate source distribution images. These candidate source distribution images are used as the initial population, and an improved genetic algorithm is employed to optimize the source distribution images of the scene to be reconstructed, aiming to maximize the fitness function value. The improved genetic algorithm includes a residual convolutional neural network guiding the search and traditional genetic operations collaboratively guiding population reproduction. The residual convolutional neural network is used to learn the residual change pattern of the source distribution image during the iterative optimization process. The fitness function represents the difference between the predicted encoded image (predicted by the pre-built cascaded convolutional neural network) and the encoded image to be reconstructed, corresponding to the population of the improved genetic algorithm.

[0013] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms.

[0014] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms.

[0015] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: In this invention, a high-quality starting point for the improved genetic algorithm is first provided by a preliminary reconstruction of the convolutional neural network, which lays a solid foundation for subsequent precise optimization. The genetic algorithm itself has a powerful global search capability. Furthermore, a residual convolutional neural network is used to guide the search. During the evolution process, the most likely modification direction that can improve image quality, i.e., the residual, is predicted. This makes each adjustment more purposeful and can more efficiently approximate the true solution, thereby obtaining higher reconstruction accuracy. The fitness function directly compares the difference between the predicted encoded image and the encoded image to be reconstructed. This can generate the result closest to the observed data (the encoded image to be reconstructed) through the known physical model (cascaded convolutional neural network), which is the optimal solution. This data-driven evaluation criterion is the cornerstone of ensuring the final accuracy. Therefore, this method can improve the accuracy of encoded image reconstruction. Furthermore, this method fully leverages the ability of the neural network to provide output in a short time after training and the global search characteristics of the genetic algorithm. Through a structured coupling mechanism, it can significantly improve the imaging speed while ensuring reconstruction accuracy. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0017] Figure 1 A schematic diagram of the process of a coding imaging reconstruction method based on the coupling of neural networks and genetic algorithms provided by the present invention; Figure 2 A schematic diagram of the preliminary reconstructed convolutional neural network (PRCNN) structure provided for this invention; Figure 3 A schematic diagram of the residual convolutional neural network (RECCNN) structure provided by this invention; Figure 4 This invention provides a specific nesting diagram of a neural network and genetic algorithm coupled reconstruction method; Figure 5 A basic structural diagram of a coding imaging reconstruction system based on the coupling of neural networks and genetic algorithms provided by this invention; Figure 6 A diagram illustrating the composition of a coded imaging hardware system for diagnostics of inertial confinement fusion, provided by this invention. Figure 7 This is a schematic diagram of a computer device for implementing a coding imaging reconstruction method based on the coupling of neural networks and genetic algorithms, as provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Encoded image reconstruction is a typical inverse problem, aiming to recover the source image given the imaging system response (i.e., the coded image) and the point spread function (PSF). Based on the uniqueness of the convergence path, existing reconstruction methods can be divided into deterministic and heuristic methods. Deterministic methods are based on rigorous mathematical derivations and have a unique convergence path given the coded image, thus the reconstruction result is deterministic. Typical algorithms include Maximum Likelihood Expectation Maximization (MLEM), Algebraic Reconstruction Technique (ART), and Simultaneous Algebraic Reconstruction Technique (SART). However, precisely because of its unique convergence path, this method is prone to getting trapped in local optima, and there is still room for improvement in reconstruction accuracy. Heuristic methods, on the other hand, have different convergence paths, and may yield different reconstruction results for the same coded image, possessing the potential to escape local optima and search for the global optimum. These methods are represented by genetic algorithms and neural network methods. Genetic algorithms achieve explicit random search through operations such as selection, crossover, and mutation. Neural networks, on the other hand, may produce different model parameters even when using the same training set due to random factors during the training process (such as weight initialization, mini-batch sampling, Dropout, etc.), resulting in non-unique output images.

[0020] However, genetic algorithms require evaluating individual fitness generation by generation within a large search space, resulting in high computational overhead and long reconstruction time. Neural networks, to achieve high-precision end-to-end reconstruction, heavily rely on training with a large number of real source image-encoded image pairs. Currently, real experimental data is scarce, and data is typically generated using Monte Carlo simulations or analytical models, which may lead to discrepancies between neural network predictions and the real source. Nevertheless, once trained, neural networks can quickly output reconstructed images, a characteristic that complements the global search capability of genetic algorithms. Therefore, coupling genetic algorithms with neural networks to construct an encoded imaging reconstruction method based on this coupling leverages both the global search advantage of genetic algorithms and the acceleration of the reconstruction process using neural networks, thus achieving both high reconstruction accuracy and computational efficiency.

[0021] In existing technologies, a method based on random coding and neural network detection imaging has been proposed, achieving a direct mapping from coded images to source distribution images. However, the reconstruction performance of this method is highly dependent on the completeness and quality of the training data, and its reconstruction capability remains insufficient in complex noisy environments and multi-source scenarios. Another approach proposes a ghost imaging method based on genetic algorithms, which optimizes speckle patterns through genetic operations to improve the quality of reconstructed images. Although this method possesses global search capabilities, the genetic algorithm itself has high computational overhead and slow iterative convergence, making it difficult to meet the requirements of real-time imaging. Yet another approach proposes a gamma-ray radioactivity imaging method based on deep learning, which uses convolutional neural networks to decode and reconstruct coded images and combines this with a deep vision system for image fusion. However, this method still relies on a large amount of simulated data for training and does not introduce sparse priors or optimization algorithms for post-processing, making it prone to artifacts or loss of detail in reconstructed images under low-count or high-noise conditions.

[0022] To address the aforementioned problems, this invention provides a coded imaging reconstruction method and system based on the coupling of neural networks and genetic algorithms. This method employs a preliminary reconstruction convolutional neural network to generate an initial reconstructed image as the initial input to the genetic algorithm. Cascaded convolutional neural networks and residual convolutional neural networks guide the precise calculation of fitness and crossover / mutation operations within the genetic algorithm, respectively, to significantly improve reconstruction efficiency while maintaining accuracy. This method fully leverages the ability of neural networks to provide output quickly after training and the global search characteristics of genetic algorithms. Through a structured coupling mechanism, it can significantly improve imaging speed while ensuring reconstruction accuracy. This provides a good reference for real-time applications such as inertial confinement fusion diagnostics, nuclear facility monitoring, and radioactive source localization, and is of great significance for promoting the practical application of coded imaging technology in multiple cutting-edge fields.

[0023] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Figure 1 This is a schematic diagram of a coding imaging reconstruction method based on the coupling of neural networks and genetic algorithms in this invention, which specifically includes the following steps: S101, the encoded image of the scene to be reconstructed is input into the pre-trained initial reconstruction convolutional neural network to obtain the initial source distribution image of the scene to be reconstructed, and the initial source distribution image is perturbed to obtain multiple candidate source distribution images.

[0025] In one embodiment, the training process for the initial reconstruction of the convolutional neural network includes: generating multiple sets of paired data of sample encoded images and real source distribution images through Monte Carlo simulation, and performing noise addition and data augmentation operations on the sample encoded images to obtain a training dataset in a standardized format; and constructing the initial reconstruction of the convolutional neural network using the training dataset.

[0026] The Preliminary Reconstruction Convolutional Neural Network (PRCNN) has a structure consisting of multiple U-Net or ResNet encoders and decoders, formed by cascading multiple subnetworks. The generated sample encoded images are used as input, and the corresponding source distribution images are used as labels for supervised training to obtain the preliminary reconstructed convolutional neural network.

[0027] In this embodiment, addressing the scarcity of real coded image-source distribution image data in typical application scenarios such as inertial confinement fusion diagnostics, a Monte Carlo method and analytical computation method are employed to accurately simulate the physical processes of the coded imaging system. A large amount of sample coded image-real source distribution image pairing data is generated through simulation, and Gaussian and Poisson noise are artificially introduced to simulate a real noise environment, improving the generalization and robustness of the subsequently trained neural network model. Finally, the simulated data is fused with some available real experimental data to form a sufficiently large and diverse training dataset. This training dataset is used to train a preliminary reconstructed convolutional neural network, ensuring the predictive accuracy of the preliminary reconstructed convolutional neural network.

[0028] Encoded images of the scene to be reconstructed are acquired using a real coded imaging system. These images are then preprocessed and normalized to ensure consistency with the training data format. The processed coded images are then input into a pre-trained convolutional neural network for preliminary reconstruction, outputting an initial source distribution image for preliminary reconstruction. To further increase population diversity, a small perturbation can be applied to the initial source distribution image to generate multiple candidate source distribution images, which serve as the initial population for the genetic algorithm. Preprocessing includes, but is not limited to, dark current correction, flat-field correction, bad pixel repair, and noise filtering. Normalization employs max-min normalization or Z-score standardization to ensure that the input data distribution of the preliminary reconstruction convolutional neural network is consistent with the training data.

[0029] In one embodiment, perturbation of the initial source distribution image to obtain multiple candidate source distribution images includes: adding Gaussian noise to the initial source distribution image or performing local random transformation or translation transformation to obtain multiple candidate source distribution images.

[0030] Specifically, multiple images obtained by adding Gaussian noise to the initial source distribution image or performing local random transformation or translation transformation, as well as the original initial source distribution image, can all be used as candidate source distribution images.

[0031] In one embodiment, the Monte Carlo simulation described above can be implemented using software such as GEANT4, MCNP, or Phits; the imaging system includes a radiation source, an encoding aperture, and a radiation detector; the encoding aperture can be selected from shapes such as pinholes, penumbra apertures, or pinhole arrays, and the material can be heavy metals such as tungsten or lead; the detector can be selected from CCDs, scintillation fiber arrays, or scintillator-coupled photomultiplier tubes; during the simulation, radiation sources with different energy ranges, various encoding plate geometric configurations, and detector response characteristics should be considered to ensure the physical authenticity and diversity of the generated data.

[0032] PRCNN training can incorporate adversarial training or attention mechanisms to improve generalization ability; network depth, number of filters, and number of subnetworks can be adjusted according to the complexity of the actual reconstruction task; early stopping strategy and learning rate decay are used during training to prevent overfitting.

[0033] S102, multiple candidate source distribution images are used as the initial population, and with the goal of maximizing the fitness function value, an improved genetic algorithm is used to optimize the source distribution images of the scene to be reconstructed, thereby obtaining the target source distribution image of the scene to be reconstructed. The improved genetic algorithm includes using a residual convolutional neural network to guide the search and traditional genetic operations to guide population reproduction. The residual convolutional neural network is used to learn the residual change law of the source distribution image during the iterative optimization process. The fitness function represents the difference between the predicted encoded image corresponding to the population of the improved genetic algorithm predicted by the pre-constructed cascaded convolutional neural network and the encoded image to be reconstructed.

[0034] Cascaded Convolutional Neural Networks (CCNNs) serve as surrogate models (SMs) in improved genetic algorithms, used for efficient evaluation of individual fitness in a population. This CCNN takes candidate solutions (i.e., hypothetical source distributions) from the improved genetic algorithm as input and outputs their corresponding predicted encoded images. By calculating the difference between the predicted encoded images and the actual acquired encoded images, it quickly estimates the fitness of individuals, thus avoiding the time-consuming physical forward modeling process. Its network architecture can adopt a structure similar to PRCNN or some of its subnetworks.

[0035] The cascaded convolutional neural network has the same structure as the initial reconstructed convolutional neural network, and the encoder part of the cascaded convolutional neural network shares weights with the initial reconstructed convolutional neural network to reduce computational redundancy and improve training efficiency; fitness calculation uses indicators such as normalized mean square error or structural similarity index.

[0036] Residual Convolutional Neural Networks (RECCNNs) are used to optimize the breeding operation of genetic algorithms. Instead of directly learning the complete source distribution image, RECCNNs learn an optimization increment—that is, "how to iterate from the current individual to approximate a better solution." RECCNNs are trained using a large dataset generated during the initial reconstruction of the convolutional neural network, combined with a genetic algorithm, to learn the residual changes in the source distribution during the iterative optimization process.

[0037] RECCNN learns the residual between a "not-good solution" and a "better solution." RECCNN is trained on a large-scale, diverse training dataset generated through Monte Carlo simulations. This dataset encompasses radioactive sources with varying energies, activities, quantities, and spatial distributions (including point sources, extended sources, and sources with complex shapes). During training, RECCNN learns the residual mapping from the "current candidate source image" to the "better solution" (typically the true source distribution or a better individual obtained through genetic algorithm iterations). This means that RECCNN has seen source images of various shapes and their optimization directions during training, thus learning a general, data-driven optimization prior. After training, it learns the experience: "Seeing an image, know which direction to modify (which pixel to strengthen, which to weaken) to make it better."

[0038] RECCNN networks, through their learned optimization priors, provide a targeted and intelligent mechanism for mutation operations, selecting a subset of high-quality individuals in each generation of the population. Perform RECCNN-guided mutation. Input the selected high-quality individuals into the pre-trained RECCNN model. The network performs forward propagation and outputs a residual map of the same size as the input individuals. The residual plot predicted by RECCNN has a certain intensity (driven by the guiding variability intensity coefficient). (Control) is superimposed on the original individual, thereby generating a new individual that has undergone directed mutation. The formula for updating the individual position based on the residual map is:

[0039] (1); in, For the original individual position, This is the residual map output by the residual convolutional neural network. The location of the new individual generated after mutation guided by the residual convolutional neural network. This is the strength coefficient.

[0040] RECCNN is trained using a large number of source distribution images and their optimization directions to learn the residual mapping from the current solution to a better solution. The network adopts a dense connection structure to promote feature reuse and uses an appropriate loss function to ensure the sparsity and interpretability of the prediction increment.

[0041] In one embodiment, multiple candidate source distribution images are used as an initial population, and an improved genetic algorithm is used to optimize the source distribution images of the scene to be reconstructed, with the goal of maximizing the fitness function value, to obtain the target source distribution image of the scene to be reconstructed. This includes the following steps: S201, Randomly initialize the source distribution image, using the randomly initialized source distribution image and multiple candidate source distribution images as the initial population.

[0042] Among them, images of multiple candidate source distributions can account for 20%-40% of the total initial population.

[0043] S202, the position of each individual in the initial population is input into the cascaded convolutional neural network to obtain the predicted encoded image of each individual. Based on the predicted encoded image of each individual and the encoded image to be reconstructed, the root mean square error is calculated, and the difference between 1 and the root mean square error is determined as the fitness value.

[0044] Fitness evaluation employs a CCNN surrogate model to achieve fast computation and avoid time-consuming forward simulation.

[0045] S203. Based on the fitness value, determine the optimal individual, and update the individual position by using residual convolutional neural network to guide the individual with fitness greater than the preset threshold, and update the individual position by using selection, crossover and mutation operations of traditional genetic algorithm for individuals with fitness less than or equal to the preset threshold.

[0046] The process of updating individual positions using a residual convolutional neural network involves: inputting the individual position into the residual convolutional neural network to obtain a residual map; the residual map represents the residual between the current individual and the better solution obtained after iteration by the genetic algorithm; and updating the individual position based on the residual map. Specifically, individuals with a fitness greater than a preset threshold have their individual positions updated using formula (1).

[0047] The selection operation combines tournament selection and roulette wheel selection; the probabilities of crossover and mutation operations are dynamically determined based on adaptive crossover and mutation factor formulas. On this basis, RECCNN-guided search and traditional genetic operations work together to guide population reproduction: RECCNN guidance is suitable for individuals with high fitness, while traditional genetic operations mainly act on individuals with medium to low fitness, thus achieving a balance between convergence speed and population diversity.

[0048] S204 performs elite operations on the updated individuals to generate new individual positions.

[0049] S205. Calculate the fitness value based on the new individual position, and iteratively update the individual position based on the fitness value until the convergence condition is met.

[0050] S206, the optimal individual position after reaching the convergence condition is determined as the target source distribution image of the scene to be reconstructed.

[0051] The convergence condition is set as either the current best fitness reaches a preset threshold or the total number of iterations reaches a preset number of generations.

[0052] In one embodiment, the coded image reconstruction method based on the coupling of neural networks and genetic algorithms provided by the present invention mainly includes the following steps: imaging process simulation and data generation, neural network model construction and pre-training, actual coded image acquisition, preliminary reconstruction using PRCNN neural network, and optimized reconstruction using the coupling of genetic algorithm and neural network. The core of this method lies in the deep coupling of three neural networks (PRCNN, CCNN, and RECCNN) with genetic algorithms to form a complementary and organic whole, thereby achieving high-precision and high-efficiency coded image reconstruction.

[0053] (1) Simulation of the imaging process and data generation: To address the challenge of scarce paired data for real-world coded images and source distributions, this invention employs Monte Carlo and analytical computation methods to accurately simulate the physical process of coded imaging, generating large-scale, diverse training datasets. Specifically, based on the different coded imaging system structures used in different scenarios, corresponding physical models are established in simulation software such as GEANT4, MCNP, or Phits. This model must accurately describe the entire process of rays emitted from the source region, modulated by coded apertures of specific materials (such as tungsten or lead) and shapes (such as pinholes, penumbra apertures, or pinhole arrays), and forming a coded image on a radiation detector (such as a CCD, scintillation fiber array, or scintillator-coupled photomultiplier tube). In analytical computation, this forward imaging process can be mathematically represented as the convolution of the source distribution and the system's point spread function (PSF). Its discrete form is:

[0054] (2) in, Indicates the first in the source image i Intensity of each pixel Indicates the first digit in the encoded image j The measurement value of each pixel. Let be the element of the system response matrix (i.e., the PSF matrix), representing the th element. i The source point pairs with the first j The contribution of each detector pixel, This represents noise. Once the imaging system is determined, its response matrix A is also determined. Therefore, by assuming a large number of source distributions f with different morphologies, paired simulated coded image data can be generated by calculating g = Af + n. To improve the realism and diversity of the data, the simulation needs to cover radioactive sources with different energies, activities, quantities, and spatial distributions, and consider various system geometric configurations. Subsequently, Gaussian and Poisson noise are artificially introduced into the simulated coded images to approximate the real experimental environment. Finally, the generated simulated data (including Monte Carlo simulations and analytical calculations) are fused with some available real experimental data to form a sufficiently large and diverse training dataset, laying the foundation for the generalization ability and robustness of the subsequent neural network.

[0055] (2) Neural network model construction and pre-training: (2.1) The core function of PRCNN is to quickly map the encoded image into a high-quality preliminary source distribution estimate. Its overall architecture is as follows: Figure 2As shown. The core function of PRCNN is to quickly map the preprocessed actual encoded image into a high-quality preliminary source distribution estimate, providing a high-quality initial population for subsequent genetic algorithms. PRCNN takes the preprocessed and normalized actual encoded image as input, which is a two-dimensional array whose dimensions depend on the detector resolution and preprocessing strategy, typically a single-channel grayscale image [H, W, 1]. The network output is the preliminary reconstructed source distribution image, whose size is consistent with the target reconstructed source region, for example [H′, W′, 1]. The network adopts an encoder-decoder symmetric architecture and embeds a skip connection mechanism to integrate multi-scale features, taking into account both global context modeling and local detail recovery. The encoder consists of multiple cascaded sub-network modules, each containing two consecutive 3×3 convolutional layers (followed by a ReLU activation function), followed by 2×2 max pooling downsampling to progressively extract multi-scale abstract features and compress redundant information. The decoder also consists of multiple modules. Each module first performs upsampling using transposed convolutions, then concatenates the feature maps passed from the corresponding encoder layers via skip connections. The features are then further fused through two 3×3 convolutional layers (containing ReLU activation). This skip connection mechanism effectively facilitates the transmission of low-level detail information, improves the reconstruction quality of edges and fine structures, and alleviates the gradient vanishing problem. Finally, the network integrates the feature map channels to 1 using a 1×1 convolutional layer, generating a preliminary reconstructed single-channel source distribution image.

[0056] (2.2) CCNN acts as a surrogate model in the genetic algorithm to accelerate fitness evaluation. Its input is a candidate individual in the genetic algorithm (i.e., a hypothetical source distribution image), and its output is the predicted encoded image corresponding to the hypothetical source distribution. CCNN can also adopt an encoder-decoder structure, and its network architecture adopts a structure similar to PRCNN or some of its sub-networks. In the iteration of the genetic algorithm, traditional methods require time-consuming physical forward modeling simulation for each individual to calculate its fitness. This invention uses a trained CCNN to replace this process: the candidate individual is input into the CCNN to quickly obtain its predicted encoded image, and its fitness (Fitness = 1 - NMSE) is efficiently estimated by calculating the difference between the predicted encoded image and the actual acquired encoded image (such as the normalized mean square error NMSE), thereby significantly reducing the computational overhead.

[0057] (2.3) RECCNN is used for the intelligently guided genetic algorithm's reproduction operations (crossover and mutation), and its overall architecture is as follows: Figure 3As shown, this architecture can adaptively adjust to the complexity of the learning task. Instead of directly learning the complete source distribution, the network learns an optimization increment (residual). Its input is the individual image of the current generation, and the learning objective is the residual between that individual and a better solution after genetic iteration. RECCNN can employ structures such as dense connections to facilitate feature reuse. After training, RECCNN receives candidate individuals from the current population as input. Its output is a residual image, which encodes the optimization direction and magnitude learned by the network. A weighted summation operation is performed during the mutation operation of the genetic algorithm. ,in For the original individual, The residual image output by RECCNN. This refers to the new individuals generated after mutation guided by RECCNN. This process transforms learned prior knowledge into specific, targeted genomic changes, replacing the completely random mutations in traditional genetic algorithms, thereby improving search efficiency and convergence speed.

[0058] (3) Actual coded image acquisition: The coded image of the target region to be reconstructed is acquired through a real coded imaging system. The data acquisition and control unit controls the acquisition time, performs energy spectrum screening, and completes background measurement. The raw coded data acquired by the detector needs to undergo a series of preprocessing operations, including dark current correction, flat field correction, bad pixel repair, and noise filtering, to eliminate system errors. Subsequently, the preprocessed image is normalized using the maximum-minimum normalization or Z-score normalization method to ensure that its data distribution is consistent with the training data generated in step (1) before it can be input into the subsequent neural network model.

[0059] (4) Preliminary reconstruction of the PRCNN neural network: The preprocessed and normalized actual encoded image from step (3) is input into the trained PRCNN model for forward propagation inference, quickly outputting a preliminary reconstructed source distribution image. This preliminary reconstruction result inherits the speed advantage of neural networks and can capture the main structure and global features of the source distribution. To further increase the diversity of the initial population of the genetic algorithm and avoid premature convergence, small perturbations (such as adding a small amount of Gaussian noise, performing random rotations or translations at small angles) can be applied to this preliminary reconstruction result to generate multiple variants. Finally, the direct output of PRCNN and its perturbation variants are combined to form a high-quality initial population of the genetic algorithm, accounting for about 20%-40% of the total population, providing a superior starting point for the subsequent global optimization of the genetic algorithm.

[0060] (5) Reconstruction by coupling genetic algorithm with neural network: Based on the initial population generated by the aforementioned steps, the neural network and genetic algorithm are coupled and reconstructed. Figure 4 This is a specific nested diagram of the neural network and genetic algorithm coupled reconstruction method.

[0061] (5.1) Fitness Evaluation: For each individual (candidate source image) in the population, instead of using traditional, time-consuming physical forward modeling to calculate its fitness, the pre-trained CCNN surrogate model from step (2.2) is invoked. The individual is input into the CCNN to quickly obtain its predicted encoded image, and its fitness value is defined by calculating the difference between the predicted image and the actual acquired encoded image (e.g., 1-NMSE). This significantly reduces computational overhead.

[0062] (5.2) Selection, Crossover, and Mutation: A strategy combining tournament selection and roulette wheel selection is adopted to select superior individuals for the next generation based on their fitness. In the crossover and mutation operations, this invention innovatively introduces RECCNN for intelligent guidance: for superior individuals with high fitness, the crossover and mutation operations are dominated by the residual increments predicted by RECCNN, resulting in more deterministic directional optimization; for individuals with medium to low fitness, the random crossover and mutation operations in traditional genetic algorithms are mainly used to maintain population diversity. The crossover and mutation probabilities can be adaptively adjusted according to the fitness value.

[0063] (5.3) Termination judgment: Repeat the iterative process of selection, reproduction (crossover and mutation), and evaluation. When the fitness value of the current best individual reaches the preset threshold, or the total number of iterations reaches the preset maximum number of generations, the algorithm terminates and outputs the current best individual as the final reconstructed source image.

[0064] This invention, through the organic coupling of neural networks and genetic algorithms, not only overcomes the slow convergence of genetic algorithms but also improves the reconstruction stability of neural networks under conditions of few samples and high noise, demonstrating promising practical applications. Compared to existing technologies, this invention offers the following technical advantages:

[0065] (A) This invention creatively constructs a novel reconstruction framework that complements the advantages of neural networks by deeply coupling the rapid reconstruction capability of neural networks with the global optimization capability of genetic algorithms.

[0066] (B) This invention organically combines Reconstruction Convolutional Neural Network (PRCNN) with genetic algorithm, giving full play to the ability of neural network to quickly generate initial solutions and the global optimization characteristics of genetic algorithm. It effectively overcomes the problems of slow convergence speed and large computational overhead of traditional genetic algorithm, and significantly improves imaging efficiency while ensuring reconstruction accuracy, which can meet the needs of real-time imaging applications.

[0067] (C) This invention innovatively introduces a cascaded convolutional neural network (CCNN) as a surrogate model in the genetic algorithm to efficiently and accurately evaluate the fitness of individuals in the population, avoids the time-consuming forward simulation process in traditional methods, significantly reduces the consumption of computing resources, and improves the overall efficiency of the reconstruction process.

[0068] (D) This invention uses a residual convolutional neural network (RECCNN) to explicitly learn the residual change pattern of the source distribution during the iteration process, which can effectively guide the crossover, mutation and other reproduction operations in the genetic algorithm, enhance the algorithm's ability to escape local optima, improve the quality of the solution and the convergence stability, and is especially suitable for complex noise environments and multi-source scenarios.

[0069] (E) The method proposed in this invention has strong versatility and scalability. It obtains diverse training data through Monte Carlo simulation, analytical calculation and experimentation, and combines deep learning strategies such as adversarial training and attention mechanism to improve the model's adaptability to different imaging scenarios, noise types and source distributions. It can still maintain excellent reconstruction performance under conditions of few samples and low count.

[0070] When applying the coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms provided by this invention, it is not necessary to... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0071] The above describes a coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding coded imaging reconstruction system based on the coupling of neural networks and genetic algorithms, comprising: The reconstruction processing module is used to input the encoded image of the scene to be reconstructed into a pre-trained preliminary reconstruction convolutional neural network to obtain the initial source distribution image of the scene to be reconstructed. The initial source distribution image is then perturbed to obtain multiple candidate source distribution images. These candidate source distribution images are used as the initial population, and an improved genetic algorithm is employed to optimize the source distribution images of the scene to be reconstructed, aiming to maximize the fitness function value. The improved genetic algorithm includes a residual convolutional neural network guiding the search and traditional genetic operations collaboratively guiding population reproduction. The residual convolutional neural network is used to learn the residual change pattern of the source distribution image during the iterative optimization process. The fitness function represents the difference between the predicted encoded image (predicted by the pre-built cascaded convolutional neural network) and the encoded image to be reconstructed, corresponding to the population of the improved genetic algorithm.

[0072] The data simulation module is used to generate paired data of multiple sets of sample encoded images and real source distribution images through Monte Carlo simulation, and to perform noise addition and data augmentation operations on the sample encoded images to obtain a standardized training dataset; the training dataset is used to build a preliminary reconstructed convolutional neural network.

[0073] The network training module is used for the initial reconstruction, construction, training, and validation of convolutional neural networks, residual convolutional neural networks, and cascaded convolutional neural networks.

[0074] The image acquisition module is used to acquire coded images of the scene to be reconstructed.

[0075] The human-computer interaction interface is used to display the network's training status, the reconstruction process during network training, and multiple views of the reconstruction results.

[0076] Specifically, the overall structure of the system is as follows: Figure 5 As shown, the system integrates five major functional modules—data simulation, network training, image acquisition, reconstruction processing, and human-computer interaction—through hardware and software collaboration. The system not only completes the entire process from imaging simulation to reconstruction output, but also effectively improves the efficiency and accuracy of reconstruction through deep coupling of neural networks and genetic algorithms.

[0077] The data simulation module is used for Monte Carlo simulation of the imaging process and generation of training datasets. This module automatically performs batch simulation tasks, using Monte Carlo methods (such as GEANT4 and MCNP) to perform forward modeling of the coded imaging process, generating paired coded image-source distribution data. By establishing the system response matrix A and introducing noise models (Gaussian and Poisson noise) and data augmentation, this module generates standardized training datasets in batches, providing a data foundation for network training.

[0078] The network training module is responsible for building, training, and validating PRCNN, CCNN, and RECCNN. It provides network architecture selection (U-Net, ResNet, etc.), hyperparameter configuration, and training process visualization. The trained model weights are available for use by the reconstruction module. Specifically, this module provides a network architecture configuration interface (supporting architectures such as U-Net and ResNet), hyperparameter settings (learning rate, batch size, loss function), and training strategy management (learning rate decay, early stopping, cross-validation). The module supports GPU-accelerated training and provides training process visualization (loss curves, accuracy curves, reconstruction effect comparison). The trained model weights are automatically saved to the model library for use by the reconstruction module.

[0079] Image acquisition module: Composed of a coded imaging hardware system and a data acquisition control unit, it completes the acquisition and preprocessing of actual coded images. The hardware includes a radiation source, coding aperture, radiation detectors (CCD, scintillator + photomultiplier tube, semiconductor detector, and scintillator fiber array, etc.) and related auxiliary equipment. The data acquisition control unit is responsible for acquisition control and raw data correction, and outputs standardized coded images. Specifically, the data acquisition control unit controls the acquisition time, background measurement, energy spectrum screening, and performs preliminary preprocessing on the raw data (dark current correction, flat field correction, bad pixel repair), outputting coded images that meet the format requirements.

[0080] The reconstruction processing module is the core computing unit, integrating neural network forward inference and genetic algorithm optimization processes. This module loads pre-trained models (PRCNN, CCNN, RECCNN) and executes the following process: First, PRCNN is used to quickly reconstruct the encoded image and generate a high-quality initial population. Then, a genetic algorithm is used for iterative optimization, where fitness evaluation is accelerated by a CCNN surrogate model, and the breeding operation combines RECCNN's directional guidance with traditional genetic operators. Finally, the optimal reconstructed image is output based on the termination condition. The module supports multi-threaded / GPU parallel computing and provides customizable reconstruction parameter settings (population size, number of iterations, crossover / mutation rate, etc.).

[0081] Human-Computer Interface: A graphical user interface used for system parameter settings, algorithm flow monitoring, and result visualization. The interface provides panels for data simulation parameter configuration, network training control, image acquisition control, and reconstruction parameter settings; it displays training status and reconstruction progress in real time (current iteration count, fitness curve, population diversity index); and supports multi-view display of reconstruction results (2D / 3D display, profile analysis, quantitative evaluation) and data export functionality.

[0082] To illustrate the effectiveness of the method and system provided by this invention, in one embodiment, Figure 6 This diagram illustrates the hardware system composition for coded imaging used in inertial confinement fusion diagnostics. In this system, neutrons are spatially modulated through coded apertures (such as a gadolinium-doped tungsten pinhole array). The modulated neutron image is received by a detector (such as a scintillation fiber array) and converted into a visible light signal. This light signal is reflected 90° by a mirror (this design prevents direct irradiation by neutrons and associated radiation, which could damage subsequent optoelectronic devices), then focused by a lens group and enhanced by an image intensifier, and finally acquired by a CCD camera to obtain a coded image of the neutron source region. In the system, the optical principal axes before and after reflection by the mirror must be 90° and in the same plane to ensure that the coded image recorded by the CCD accurately reflects the neutron distribution without geometric distortion. Simultaneously, shielding materials (such as lead or polyethylene) are placed beside the CCD and image intensifier to enhance radiation protection.

[0083] The acquired coded images are sent to the reconstruction processing module via a data transmission system. This module integrates the method described in this invention, first calling PRCNN for rapid preliminary reconstruction, then initiating a genetic algorithm enhanced by CCNN and RECCNN for iterative optimization, and finally outputting a high-fidelity two-dimensional distribution image of the fusion neutron source region through a human-computer interaction interface. This system, through the efficient coupling of neural networks and genetic algorithms, inherits the global optimization capabilities of genetic algorithms and utilizes neural networks to accelerate the optimization process, providing a stable and effective diagnostic scheme for inertial confinement fusion diagnosis.

[0084] Specific limitations regarding the coded imaging reconstruction system based on the coupling of neural networks and genetic algorithms can be found in the limitations of the coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms mentioned above, and will not be repeated here. Each module in the aforementioned coded imaging reconstruction system based on the coupling of neural networks and genetic algorithms can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0085] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method is a coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms.

[0086] The present invention also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The provided method is a coded imaging reconstruction method based on the coupling of neural networks and genetic algorithms.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0088] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for coded imaging reconstruction based on coupling of neural network and genetic algorithm, characterized in that, The application relates to a method for reconstructing a source distribution image of a scene to be reconstructed, comprising the following steps: encoding images of the scene to be reconstructed are input into a preliminarily trained preliminary reconstruction convolutional neural network to obtain an initial source distribution image of the scene to be reconstructed, and the initial source distribution image is disturbed to obtain multiple candidate source distribution images; the multiple candidate source distribution images are taken as an initial population, and a maximum fitness function value is taken as a target, and an improved genetic algorithm is used to optimize the source distribution image of the scene to be reconstructed to obtain a target source distribution image of the scene to be reconstructed; the improved genetic algorithm comprises using a residual convolutional neural network to guide search and traditional genetic operations to cooperatively guide population breeding; the residual convolutional neural network is used to learn a residual change rule of the source distribution image in an iterative optimization process; the fitness function represents a difference between a predicted encoding image of the population corresponding to the improved genetic algorithm and a to-be-reconstructed encoding image, which is predicted by a preliminarily constructed cascade convolutional neural network.

2. The method of claim 1, wherein, The initial source distribution image is disturbed to obtain the multiple candidate source distribution images, comprising the following steps: Gaussian noise is added to the initial source distribution image or local random transformation or translation transformation is performed on the initial source distribution image to obtain the multiple candidate source distribution images.

3. The method of claim 1, wherein, The multiple candidate source distribution images are taken as the initial population, and the maximum fitness function value is taken as the target, and the improved genetic algorithm is used to optimize the source distribution image of the scene to be reconstructed to obtain the target source distribution image of the scene to be reconstructed, comprising the following steps: the source distribution image is randomly initialized, and the randomly initialized source distribution image and the multiple candidate source distribution images are taken as the initial population; each individual position of the initial population is input into a cascade convolutional neural network to obtain a predicted encoding image of each individual; a root mean square difference is calculated according to the predicted encoding image of each individual and the to-be-reconstructed encoding image, and a difference value between 1 and the root mean square difference is determined as a fitness value; an optimal individual is determined according to the fitness value, and an individual position is updated through residual guidance of the residual convolutional neural network for an individual whose fitness value is greater than a preset threshold value, and the individual position is updated through selection, crossover and mutation operations of the traditional genetic algorithm for an individual whose fitness value is less than or equal to the preset threshold value; an elite operation is performed on the updated individual to generate a new individual position; the fitness value is calculated according to the new individual position, and the individual position is iteratively updated according to the fitness value until a convergence condition is reached; the optimal individual position after the convergence condition is reached is determined as the target source distribution image of the scene to be reconstructed.

4. The method of claim 3, wherein, The individual position is input into the residual convolutional neural network to obtain a residual image; the residual image is a residual between a current individual and a better solution after the genetic algorithm is iterated; the individual position is updated according to the residual image. A calculation formula for updating the individual position according to the residual image is:

5. The method of claim 4, wherein, A training process of the preliminary reconstruction convolutional neural network comprises the following steps: ； wherein, is the original individual position, is the residual map output by the residual convolutional neural network, is the new individual position generated after guiding the variation via the residual convolutional neural network, is the intensity coefficient.

6. The method of claim 1, wherein, a plurality of sets of paired data of sample encoding images and real source distribution images are generated through Monte Carlo simulation, and noise adding and data enhancement operations are performed on the sample encoding images to obtain a standardized training data set; the preliminary reconstruction convolutional neural network is constructed through the training data set. ​ 7. The method of claim 6, wherein, The structure of the preliminary reconstruction convolutional neural network is in the form of multiple U-Net or ResNet encoders and decoders, which are connected by cascading multiple sub-networks; the cascaded convolutional neural network has the same structure as the preliminary reconstruction convolutional neural network, and the encoder part of the cascaded convolutional neural network shares weights with the preliminary reconstruction convolutional neural network.

8. A coding imaging reconstruction system based on coupling of neural network and genetic algorithm, characterized in that, The system comprises: The reconstruction processing module is configured to input the encoded image of the scene to be reconstructed into the preliminary reconstruction convolutional neural network trained in advance to obtain an initial source distribution image of the scene to be reconstructed, and to disturb the initial source distribution image to obtain multiple candidate source distribution images; the multiple candidate source distribution images are taken as an initial population, and a target source distribution image of the scene to be reconstructed is obtained by optimizing the source distribution image of the scene to be reconstructed through an improved genetic algorithm with the maximum fitness function value as the target; the improved genetic algorithm comprises residual convolutional neural network guided search and traditional genetic operation for guiding population reproduction; the residual convolutional neural network is configured to learn the residual change rule of the source distribution image in the iterative optimization process; and the fitness function represents the difference between the predicted encoded image corresponding to the population of the improved genetic algorithm predicted by the preliminary constructed cascaded convolutional neural network and the encoded image to be reconstructed.

9. The system of claim 8, wherein, The system further comprises: The data simulation module is configured to generate paired data of multiple sets of sample encoded images and real source distribution images through Monte Carlo simulation, and to perform noise addition and data enhancement operations on the sample encoded images to obtain a standardized training data set; and the training data set is configured to construct the preliminary reconstruction convolutional neural network; The network training module is configured to construct, train and verify the preliminary reconstruction convolutional neural network, the residual convolutional neural network and the cascaded convolutional neural network; The image acquisition module is configured to acquire the encoded image of the scene to be reconstructed; The human-computer interaction interface is configured to display the training state of the network, the reconstruction progress in the network training process, and the multi-view display of the reconstruction result.

Images