Image denoising neural network architecture search method based on improved multi-target marine predator algorithm

Abstract

The invention relates to the technical field of image denoising, and discloses an image denoising neural network architecture search method based on an improved multi-target marine predator algorithm, which comprises the following steps of: 1, designing a search space of denoising convolutional neural network hyper-parameter optimization (neural architecture search); 2, initializing related parameters of an MPA algorithm, a population and a search space searched by a convolutional neural network architecture, wherein each individual in the population represents a de-noising convolutional neural network; according to the method, an optimal individual (namely an optimal denoising convolutional neural network architecture) is automatically searched by improving a marine predator algorithm, so that image denoising processing is carried out, and the problems that in the prior art, an initialized population is generated through a random initialization method, which may cause the evolution of the population towards a non-optimal direction in an iteration process, and the probability that the population is not optimal in the iteration process is low are solved. And after the individuals of the population are updated, corresponding punishment measures are not taken for the inferior individuals, so that the convergence speed of the algorithm is relatively low.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising, and in particular to an image denoising neural network architecture search method based on an improved multi-target marine predator algorithm. Background Art

[0002] Image denoising is the basis and premise of image processing. As an indispensable part of the image preprocessing stage, image denoising is an important research direction in the field of image processing. Traditional image denoising methods are mainly divided into spatial domain methods and transform domain methods. Commonly used spatial domain denoising methods include mean filtering, median filtering and Gaussian filtering, while transform domain denoising methods mainly involve cosine transform and wavelet domain methods. Traditional denoising methods are simple to operate, but there are problems such as fixed models and inability to adapt. For example, in the patent "Image denoising method, electronic device, and medium based on adaptive value range filtering" applied by Zhijiang Laboratory (patent publication number CN117575941A), the principle of the image denoising method used is relatively simple, and it can better preserve image details in some simple image denoising tasks, but the denoising effect is relatively weak in the case of more complex noise. At present, many deep learning-based methods have achieved excellent performance in the field of image denoising, and many different network structures have been introduced into the field of image denoising, including UNet, which was originally used for medical image segmentation, and DenseNet and Residual Network (ResNet) originally used for image classification. DenseNet and ResNet solve the problem of gradient vanishing through dense connections and residual connections respectively, which helps the network to better learn noise features. These denoising models based on neural networks have good denoising effects on text and images. In 2019, Neji et al. published a paper titled "Adversarial autoencoders for denoising digitized historical documents: the use case of incunabula", which designed an adversarial autoencoder to automatically remove noise from ancient text images. In 2020, Zhang et al. published a paper titled "A novel generative adversarial net for calligraphic tablet images denoising", which introduced a new method based on a generative adversarial network (GAN). GAN is a type of generative adversarial network, which is used as a text image denoiser to generate denoised text images by learning mixed synthetic noise. The First Affiliated Hospital of Nanchang University mentioned "a method for intelligent fusion of CT image data with MRI data" in their patent application (patent application number CN202311650107.6, publication number CN117636116A). Although the denoising model used works well, it is not optimal due to the manually set network parameters. Compared with traditional image denoising methods, these denoising methods based on convolutional neural networks perform better in terms of numerical evaluation indicators and visual effects.However, the above work did not explore the design of network architecture and its impact on denoising performance. Currently, all networks are handcrafted and are limited by available computing resources. Manually designing networks often takes a lot of time to select a suitable network structure. Therefore, compared with manual design methods, Neural Network Search (NAS), which does not rely on expert experience and knowledge, has attracted widespread attention from researchers and has become a research hotspot in the field of artificial intelligence.

[0003] The goal of neural architecture search is to discover and design high-performance neural architectures through automated methods to eliminate the tedious and complex process of manually designing network architectures. Image denoising network architecture search is essentially a multi-objective optimization problem, and the performance of the current denoising convolutional neural network can be judged by indicators such as PSNR and SSIM. At present, evolutionary algorithm is a global optimization algorithm with wide applicability. It has the advantages of self-organization, self-adaptation and self-learning, and is increasingly widely used in the field of neural architecture search. In 2018, Liu et al. used genetic algorithms to optimize neural architectures and network parameters in the article "Hierarchical representations for efficient architecture search". However, the search efficiency of the intelligent algorithms used in these traditional neural architecture search methods is low, and there are still problems such as easy stagnation and easy to fall into local extreme values. In 2020, Faramarzi et al. published the paper "Marine Predators Algorithm: A nature-inspired metaheuristic", proposing a new evolutionary algorithm MPA. Compared with traditional genetic algorithms, MPA has the advantages of fast convergence speed and strong global search capability. However, this method still has some shortcomings. For example, the initialization population is generated by a random initialization method, which may cause the population to evolve in a non-optimal direction during the iteration process. In addition, after the individuals in the population are updated, no corresponding penalty measures are implemented for inferior individuals, resulting in a slow convergence of the algorithm.

[0004] In summary, how to design an efficient denoising convolutional neural network neural architecture search method is a key issue that needs to be solved urgently.

[0005] To address this problem, an improved marine predator algorithm is proposed to automatically search for the optimal neural network architecture, and ultimately an image denoising network model with excellent denoising performance is obtained. Summary of the invention

[0006] The purpose of the present invention is to solve the problems existing in the prior art and to propose an image denoising neural network architecture search method based on an improved multi-objective marine predator algorithm.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: An image denoising neural network architecture search method based on an improved multi-objective marine predator algorithm comprises the following steps: Step 1: Design the search space for denoising convolutional neural network hyperparameter optimization (neural architecture search); Step 2: Initialize the relevant parameters of the MPA algorithm, the population, and the search space of the convolutional neural network architecture search, where each individual in the population represents a denoising convolutional neural network; Step 3: Use the training set to train the convolutional neural network model corresponding to the individuals in the population, and use the PSNR and SSIM indicators to evaluate the quality of each individual in the population; Step 4: Use the non-dominated quick sort algorithm to rank the individuals, where level 0 is the highest level, that is, the approximate Pareto optimal individual of the current population, and use the improved MPA to update each individual; Step 5: Repeat steps 3 and 4. When the maximum number of iterations is reached, the approximate Pareto optimal individual obtained is the current optimal denoising convolutional neural network architecture. Step 6: Use the optimal denoising convolutional neural network architecture to denoise the image.

[0008] Preferably, in step 1, the parameters for hyperparameter optimization include network structure type, number of network blocks, number of convolutional layers, convolution kernel size, number of convolution kernels, activation function, optimizer, batch normalization processing, learning rate, loss function, and batch size.

[0009] Preferably, the specific process of step 2 is: based on the value space of the parameters of hyperparameter optimization, a hybrid initialization method is used to perform evolutionary initialization on the population: that is, first a random initialization method is used to generate a population, then a quasi-reflective learning mechanism is used to update the individuals in the population, and finally the convolutional neural network model corresponding to each individual is trained using the training set, and the fitness of the individual is evaluated using the two indicators of PSNR and SSIM, and the individuals with better performance between the individuals generated by random initialization and the individuals updated by quasi-reflective learning are selected to construct the initial population.

[0010] Preferably, the specific process of step 3 is: training the denoising convolutional neural network corresponding to the individuals in the population in the training set, and calculating the values ​​of the two indicators PSNR and SSIM obtained by the trained network on the test set.

[0011] Preferably, the specific process of step 4 is: Step 4.1: According to the fitness value of the individual, the non-dominated quick sorting method is used to sort the individuals in the population, where the individuals at level 0 have the highest level and are used as individuals in the elite population; Step 4.2: Use the improved marine predator algorithm to update the population.

[0012] Preferably, in step 4.1, the fitness value of each individual has two evaluation indicators, namely PSNR and SSIM, and the specific steps of ranking the individuals in the population using the non-dominated quick sorting method are: If an individual in the population has one or two fitness values ​​that are better than other individuals in the population, then the number of dominated solutions of this individual is 0, that is, it is an excellent individual with a level of 0 in the population.

[0013] Preferably, the step 4.2 comprises the following steps: Step 4.2.1: The population update strategy depends on the ratio of the current evolutionary generation t to the maximum iteration number T. The population adopts different evolution mechanisms at different evolutionary stages. The specific update mechanism is: when When the individual (prey) moves faster than the predator (elite individual), the speed of the prey and predator is relatively high, and the individual mainly explores in the search space. In this stage, the individual (prey) uses Brownian motion to move towards the optimal individual, where the Brownian motion step size is a Gaussian random variable with a mean of 0 and unit variance; when When , the speed of prey and predator is similar, and individuals in the population conduct global exploration and local exploitation at the same time. That is, half of the individuals in the population are updated using Lévy motion to conduct local exploitation in the search space, and the remaining individuals are updated using Brownian motion to conduct global exploration in the entire search space; Step 4.2.2: Calculate the fitness of the updated new individual. The updated new individual will be compared with the old individual. If both meet the requirements > and > , then the old individuals in the population are replaced by new individuals; In each generation of the evolutionary process, after all individuals in the population are updated through the above-mentioned update mechanism, the population is non-dominated and sorted, and the lowest-ranked individuals are punished to redirect the poor individuals toward the optimal individuals. At the same time, the eddy current and fish aggregating devices (FADs) mechanism is used to continue updating the population and non-dominated sorting is performed on the updated population to prevent the population from falling into a local optimum.

[0014] Preferably, in step 5, when the maximum number of iterations is reached, the individual with a ranking level of 0 is the current optimal denoising convolutional neural network architecture.

[0015] Preferably, in step 5, the optimal denoising convolutional neural network architecture obtained includes multiple ones, whose values ​​depend on the individuals in the approximate Pareto front contained in the currently updated population when the maximum number of iterations is reached, wherein the approximate Pareto front refers to those individuals with the highest ranks in the population after the individuals in the population are ranked using the fast non-dominated sorting algorithm.

[0016] Preferably, in step 6, the optimal denoising convolutional neural network is determined by judging the quality of the image denoising effect according to the fitness value PSNR and SSIM index.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention can automatically search for the optimal individual (i.e., the optimal denoising convolutional neural network architecture) by improving the marine predator algorithm, thereby performing image denoising processing, thereby solving the problem that the initialization population in the prior art is generated by a random initialization method, which may cause the population to evolve in a non-optimal direction during the iteration process, and after the individuals in the population are updated, no corresponding penalty measures are implemented for inferior individuals, thereby resulting in a slow convergence speed of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of an image denoising method based on an improved marine predator neural architecture search provided by an example of the present invention; Figure 2 A flowchart of performing non-dominated sorting on a population to perform individual level classification on the population provided by an example of the present invention; Figure 3 The PNSR and SSIM index sizes of the network structure represented by the final Pareto optimal set (optimal individual set) and non-optimal individuals in the validation set provided in the example of the present invention; Figure 4 The performance of the optimal denoising convolutional neural network obtained for denoising the noisy image 1 in the public MRI dataset (the left is the original image, the middle is the image after adding noise, and the right is the image after denoising); Figure 5 The performance of the optimal denoising convolutional neural network obtained for denoising the noisy image 2 in the public MRI dataset (the left is the original image, the middle is the image after adding noise, and the right is the denoised image); Figure 6The optimal denoising convolutional neural network obtained is used to denoise the noisy image 3 in the public MRI dataset (the left is the original image, the middle is the image with added noise, and the right is the denoised image). DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0020] like Figures 1 to 6 As shown: In this implementation case, an image denoising neural network architecture search method based on an improved multi-objective marine predator algorithm is used. Figure 1 The pseudo code of the overall process is shown in Table 1: Table 1 Convolutional neural network model method based on improved marine predator algorithm optimization Convolutional neural network model method based on improved marine predator algorithm optimization Input: population size P, maximum number of iterations - T, public medical image training set -, public medical image test set -, public medical image verification set; Output: optimal denoising convolutional neural network architecture 1: Start 2: Initialize the population - generate a population of N individuals according to the quasi-reverse learning mechanism, and perform model training according to, to generate the fitness value of each individual; sort the population individuals according to the non-dominated quick sorting algorithm, and put the individuals with the highest ranking into the optimal network architecture set; 3: for i = 1 to T do / / Perform iterative operation 4: forj = 1 to P do / / New population generation operation 5: if i <= T / 3: 6: Use formulas (5) and (6) to update the population individuals; 7: else if T / 3 < i <= 2 / 3T: 8: Use formulas (7), (8), (9), and (10) to update the population individuals; 9: else if i > 2 / 3T: 10: Use formulas (15) and (16) to update the population individuals; 11: end if 12: If the fitness value of the new individual is better than that of the old individual, update the old individual; 13: endfor14: Use non-dominated quick sorting for the new population, and put the highest-ranked individuals into the optimal network architecture set; 15: Implement a penalty mechanism for individuals with poor performance, and generate new individuals to replace these individuals; 16: Generate the highest-ranked individuals in the new population in a loop as individuals in the Elite population (predator population); 17: end for 18: Rank the obtained optimal network architecture set using the non-dominated quick sorting algorithm; 19: return the optimal denoising network architecture (Pareto front). 20: Use the optimal denoising network to denoise the image, and verify the network performance in. 21: End The method specifically comprises the following steps: Step 1: Design a search space for hyperparameter optimization of denoising convolutional neural network; The 11 parameters to be optimized are network structure type, number of network blocks, number of convolutional layers, convolution kernel size, number of convolution kernels, activation function, optimizer, batch normalization, learning rate, loss function, batch size. The parameters of the convolutional neural network to be optimized and their value space are shown in Table 2: Table 2 Types, ranges and parameter subscripts of convolutional neural network parameter models Hyperparameters Hyperparameter selection range Subscript selection Network structure type Redisual, Dense TOB=[0,1] Number of network blocks 2,3,4,5,6,7 NOB=[0,1,2,3,4,5] Number of convolutional layers 1,2,3,4 NOC=[0,1,2,3] Convolution kernel size 3*3,5*5,7*7 FS=[0,1,2] Number of convolution kernels 8,16,32,64,128 NOF=[0,1,2,3,4] Activation Function ReLU, Tanh, Leaky ReLU ACT=[0,1,2] Optimizer Adam,Adadelta,Adamax,SGD,Adagrad OPT=[0,1,2,3,4] Batch Normalization False,True BN=[0,1] Learning Rate 0.1,0.01,0.001 LR=[0,1,2] Loss Function MAE,MSE LF=[0,1] Batch size 8,16,32,64,96 BS=[0,1,2,3,4] At this time, if an individual in the population is represented as [0,1,2,1,0,0,1,1,0,1,2], it means that a network block type is Redisual, i.e. TOB[0], the number of network blocks is 3, i.e. NOB[1], the number of convolutional layers is 3, i.e. NOC[2], the convolution kernel size is 5*5, i.e. FS[1], the number of convolution kernels is 8, i.e. NOF[0], the activation function is ReLU, i.e. ACT[0], the optimizer is Adadelta, i.e. OPT[0], the batch normalization is True, i.e. BN[1], the learning rate is 0.1, i.e. LR[0], the loss function is MSE, i.e. LF[1], and the batch size is 32, i.e. BS[2], is represented by an 11-dimensional instance vector.

[0021] Step 2: Initialize the relevant parameters of the MPA algorithm, the population, and the search space of the convolutional neural network architecture search, where each individual in the population represents a denoising convolutional neural network; As a further improvement of the present invention, the step 2 specifically comprises: Initialize the population (prey matrix) :

[0022] in, It is the first predator species Individuals, Represents the individual dimension, Indicates the population size.

[0023] Step 2.1, based on the value space of 11 parameters such as network structure type, number of network blocks, number of convolutional layers, etc., a hybrid initialization method is used to initialize the population for evolution. That is, firstly, a random initialization method is used to generate the population, and then the quasi-reflective learning mechanism is used to update the individuals in the population. Finally, the convolutional neural network model corresponding to each individual is trained using the training set, and the fitness of the individual is evaluated using the two indicators of PSNR and SSIM. The individuals with better performance between the individuals generated by random initialization and the individuals updated by quasi-reflective learning are selected to construct the initial population. It can be seen that the population is optimized in the initialization stage to ensure that the individuals in the initial population have diversity and convergence, thereby reducing the possibility of the population falling into local extremes in the subsequent evolution process.

[0024] As a further improvement of the present invention, the step 2.1 specifically includes: Assume that the individuals in the population are randomly initialized and labeled as , the individuals of the new population are generated by the reverse learning mechanism, and the newly generated individuals are , as shown in formulas (1) and (2):

[0025]

[0026] in It is an 11-dimensional vector containing the maximum index value that each hyperparameter can take. Contains the minimum index value that each hyperparameter can take.

[0027] As a further improvement of the present invention, the step 2.1 specifically includes: The fitness indicators are PSNR and SSIM, which are mainly used to evaluate the image denoising ability of the convolutional neural network architecture.

[0028] PSNR is an indicator to measure image quality. Its calculation method is shown in formula (3):

[0029] Where MSE is the mean square error between the original image and the denoised image, and n represents the number of image sampling points. The PSNR value range is 20-40dB. The larger the PSNR value, the better the image quality.

[0030] SSIM measures image similarity from three aspects: brightness, contrast, and structure. The mean is used as an estimate of brightness, the standard deviation is used as an estimate of contrast, and the covariance is used as a measure of structural similarity. The SSIM value range is [0,1]. The larger the value, the smaller the image distortion and the more similar it is to the original image. The calculation method of SSIM is shown in formula (4):

[0031] in and is a constant. and are the means of images x and y respectively, and are the variances of images x and y, respectively. is the covariance of images x and y.

[0032] Step 3: Use the non-dominated quick sort algorithm to rank the individuals, where level 0 is the highest level, that is, the approximate Pareto optimal individual of the current population. At the same time, use the improved marine predator algorithm to perform the optimization update operation on the individuals and complete a complete iteration; As a further improvement of the present invention, the step 3 specifically includes: The specific steps of the non-dominated quick sort algorithm are: Assume that there are two individuals x and y. The corresponding fitness values ​​are , and , The two individuals are initially dominated by and As shown in formula (5):

[0033] The other cases specifically mean that if one and only one of the two fitness indicators of individual x is better than that of individual y, then the two individuals do not dominate each other.

[0034] Step 3.1: Use the training set to train the convolutional neural network model. and The two indicators evaluate the quality of each individual in the population; the denoising convolutional neural network corresponding to the individuals in the population is trained in the training set, and the trained network is calculated on the test set. and The values ​​of the two indicators.

[0035] According to the fitness value of the individual, the non-dominated quick sorting method is used to sort the individuals in the population, among which the individuals with level 0 have the highest level and are regarded as individuals in the elite population; As a further improvement of the present invention, the step 3 specifically includes: The elite matrix (i.e., the predator population) guides the movement direction of individuals in the population, allowing the individuals in the population to evolve towards the optimal individuals. The individuals in the elite matrix are composed of those individuals with a level of 0 after the non-dominated quick sorting of each generation of the population (i.e., the prey population). The elite matrix is ​​shown below:

[0036] in, Is the elite population Individuals, Represents the individual dimension, surface Step 3.2: Use the improved marine predator algorithm to update the population.

[0037] Step 3.2.1: The population update strategy depends on the ratio of the current evolutionary generation t to the maximum iteration number T. The population adopts different evolution mechanisms at different evolutionary stages. The specific update mechanism is: Step 3.2.1.1: When the individual (prey) moves faster than the predator (elite individual), the speed of the prey and predator is relatively high, and the individual mainly explores in the search space. In this stage, the individual (prey) uses Brownian motion to move towards the optimal individual, where the Brownian motion step size is a Gaussian random variable with a mean of 0 and unit variance; As a further improvement of the present invention, the step 3.2.1.1 specifically includes: when When , it is in the high speed ratio stage, at this time the prey's movement speed is faster than the predator, and the speed ratio between the prey and the predator is high. This stage is the exploration stage, and the update mechanism of the population individual (prey) is shown in formulas (6) and (7):

[0038]

[0039] The standard Brownian motion is a random process whose step size comes from the probability density function defined by a Gaussian distribution with mean 0 and unit variance. is a random number vector simulating Brownian motion; is a term-by-term multiplication operator; P = 0.5 is a constant, and R is a random number vector in the range [0,1]. and Represent the prey matrix and Elite Matrix The i-th row of .

[0040] Step 3.2.1.2: When , the speed of prey and predator is similar, and individuals in the population conduct global exploration and local exploitation at the same time. That is, half of the individuals in the population are updated using Lévy motion to conduct local exploitation in the search space, and the remaining individuals are updated using Brownian motion to conduct global exploration in the entire search space; As a further improvement of the present invention, the step 3.2.1.2 specifically includes: when , the speed of the predator and the prey are similar. This stage occurs in the middle of evolution. The prey individuals and predator individuals follow the Levy run and Brownian motion respectively. The prey population in the first half and the prey population in the second half are updated using formulas (8), (9), (10), and (11) respectively:

[0041]

[0042]

[0043]

[0044] in is a random number vector. CF is an adaptive parameter used to control the step length of individual movement in the population, as shown in formula (12):

[0045] is obtained by Levy movement, using Levy ( ) is expressed as shown in formula (13):

[0046] x and y are two normally distributed variables with standard deviations of and , where x and y are as shown in formulas (14) and (15):

[0047]

[0048] Step 3.2.1.3: When the predator improves the development efficiency of the algorithm through the Levy movement strategy, it further makes the individuals (prey) in the population evolve towards the optimal individual.

[0049] As a further improvement of the present invention, the step 3.2.1.3 specifically includes: when It occurs in the late stage of population individual evolution. In this stage, the prey improves the development ability of the algorithm through the Levy movement strategy. The update process of the prey is shown in formulas (16) and (17):

[0050]

[0051] The new individual obtained after the prey individual is updated in the above stage , will be with the old individual Compare the fitness value indicators. If the PSNR and SSIM of the new individual are better than those of the old individual, the original old individuals in the population are updated and replaced by the new individuals.

[0052] Step 3.2.1.4: In each generation of the evolutionary process, after the individuals in the population are updated through the above-mentioned update mechanism, the population is sorted in a non-dominated manner. The lowest-ranked individuals are penalized, and the poor individuals are redirected to approach the optimal individuals. At the same time, the eddy current and fish aggregation effect FADs mechanism is adopted to prevent the individuals in the population from falling into the local optimum. At the same time, after all individuals in the population are updated, the worst individuals (those with the largest number of dominated solutions after non-dominated quick sorting) will be penalized to cause mutations. As shown in formula (18):

[0053] in It represents individuals with poor current performance. and are any two different individuals in the population except the worst ones, Represents the penalty factor, which is 0.5. The current individual is guided to evolve towards the optimal individual through other excellent individuals.

[0054] The FADs optimization process is shown in formula (19):

[0055] Among them, U is a binary vector. FADs represents the probability of affecting the optimization process and is usually set to 0.2. and are two random different individuals in the prey matrix (not the current individual ). r is a random decimal in the range [0,1]. When r <= FADs, the prey will jump longer in different dimensions to find other possible optimal individual solution spaces, thereby achieving the possibility of jumping out of the local optimum. When r > FADs, the predator will move in the current prey space.

[0056] Step 4: Repeat step 3. When the maximum number of iterations is reached, the approximate Pareto optimal individual obtained is the current optimal denoising convolutional neural network architecture. Step 4.1, the approximate Pareto front refers to the individuals with the highest ranking in the population according to the non-dominated quick sort; Step 5: Use the optimal denoising convolutional neural network architecture to perform image denoising; Step 5.1: Use the optimal denoising convolutional neural network architecture to perform image denoising. The quality of image denoising is judged based on the PSNR and SSIM indicators.

[0057] Figure 3 The fitness value indicators of the denoising convolutional neural network structure represented in the final approximate Pareto optimal individual are evaluated in the public dataset.

[0058] The present invention improves the marine predator algorithm to automatically search for the optimal individual (i.e., the optimal denoising convolutional neural network architecture) to perform image denoising, thereby solving the problem that the initialization population in the prior art is generated by a random initialization method, which may cause the population to evolve in a non-optimal direction during the iteration process, and after the individuals of the population are updated, no corresponding penalty measures are implemented for inferior individuals, resulting in a slow convergence speed of the algorithm.

[0059] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A neural network architecture search method for image denoising based on an improved multi-objective marine predator algorithm, characterized by: The following steps are involved: Step 1: Design the search space for denoising convolutional neural network hyperparameter optimization (neural architecture search); Step 2: Initialize the relevant parameters of the MPA algorithm, the population, and the search space of the convolutional neural network architecture search, where each individual in the population represents a denoising convolutional neural network; Step 3: Use the training set to train the convolutional neural network model corresponding to the individuals in the population, and use the PSNR and SSIM indicators to evaluate the quality of each individual in the population; Step 4: Use the non-dominated quick sort algorithm to rank the individuals, where level 0 is the highest level, that is, the approximate Pareto optimal individual of the current population, and use the improved MPA to update each individual; Step 5: Repeat steps 3 and 4. When the maximum number of iterations is reached, the approximate Pareto optimal individual obtained is the current optimal denoising convolutional neural network architecture. Step 6: Use the optimal denoising convolutional neural network architecture to denoise the image.

2. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 1 is characterized by: In step 1, the parameters for hyperparameter optimization include network structure type, number of network blocks, number of convolutional layers, convolution kernel size, number of convolution kernels, activation function, optimizer, batch normalization, learning rate, loss function, and batch size.

3. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 2 is characterized by: The specific process of step 2 is: based on the value space of the parameters of hyperparameter optimization, a hybrid initialization method is used to perform evolutionary initialization on the population: that is, firstly, a random initialization method is used to generate the population, then a quasi-reflective learning mechanism is used to update the individuals in the population, and finally, the convolutional neural network model corresponding to each individual is trained using the training set, and the fitness of the individual is evaluated using the two indicators of PSNR and SSIM, and the individuals with better performance between the individuals generated by random initialization and the individuals updated by quasi-reflective learning are selected to construct the initial population.

4. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 3 is characterized by: The specific process of step 3 is: training the denoising convolutional neural network corresponding to the individuals in the population in the training set, and calculating the values ​​of the two indicators PSNR and SSIM obtained by the trained network on the test set.

5. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 4, characterized in that: The specific process of step 4 is as follows: Step 4.1: According to the fitness value of the individual, the non-dominated quick sorting method is used to sort the individuals in the population, where the individuals at level 0 have the highest level and are used as individuals in the elite population; Step 4.2: Use the improved marine predator algorithm to update the population.

6. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 5, characterized in that: In step 4.1, the fitness value of each individual has two evaluation indicators, namely PSNR and SSIM. The specific steps of using the non-dominated quick sorting method to rank the individuals in the population are as follows: If an individual in the population has one or two fitness values ​​that are better than other individuals in the population, then the number of dominated solutions of this individual is 0, that is, it is an excellent individual with a level of 0 in the population.

7. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 6, characterized in that: The step 4.2 comprises the following steps: Step 4.2.1: The population update strategy depends on the ratio of the current evolutionary generation t to the maximum iteration number T. The population adopts different evolution mechanisms at different evolutionary stages. The specific update mechanism is: when When the individual (prey) moves faster than the predator (elite individual), the speed of the prey and predator is relatively high, and the individual mainly explores in the search space. In this stage, the individual (prey) uses Brownian motion to move towards the optimal individual, where the Brownian motion step size is a Gaussian random variable with a mean of 0 and unit variance; when When , the speed of prey and predator is similar, and individuals in the population conduct global exploration and local exploitation simultaneously, that is, half of the individuals in the population are updated by Lévy motion to conduct local exploitation in the search space, and the remaining individuals are updated by Brownian motion to conduct global exploration in the entire search space; Step 4.2.2: Calculate the fitness of the updated new individual. The updated new individual will be compared with the old individual. If both meet the requirements > and > , then the old individuals in the population are replaced by new individuals; In each generation of the evolutionary process, after all individuals in the population are updated through the above-mentioned update mechanism, the population is non-dominated and sorted, and the lowest-ranked individuals are punished to redirect the poor individuals toward the optimal individuals. At the same time, the eddy current and fish aggregating devices (FADs) mechanism is used to continue updating the population and non-dominated sorting is performed on the updated population to prevent the population from falling into a local optimum.

8. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 7, characterized in that: In step 5, when the maximum number of iterations is reached, the individual with a ranking level of 0 is the current optimal denoising convolutional neural network architecture.

9. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 8, characterized in that: In step 5, the optimal denoising convolutional neural network architecture obtained includes multiple ones, and the value of the architecture depends on the individuals in the approximate Pareto front contained in the currently updated population when the maximum number of iterations is reached, wherein the approximate Pareto front refers to the individuals with the highest rank in the population after the individuals in the population are ranked using the fast non-dominated sorting algorithm.

10. The image denoising neural network architecture search method based on the improved multi-objective marine predator algorithm according to claim 6, characterized in that: In step 6, the optimal denoising convolutional neural network is determined by judging the quality of the image denoising effect according to the fitness value PSNR and SSIM index.

Images