A Fourier ptychographic microscopy reconstruction method, device and equipment

Abstract

The present invention discloses a Fourier ptychographic microscopy reconstruction method, device and equipment. The present invention relates to the technical field of microscopy imaging and is used to solve the problems of large data calculation amount and low quality of reconstructed images in the existing reconstruction methods. It includes: obtaining an image dataset that includes Fourier ptychographic microscopy simulation image data, real datasets and is added with Gaussian noise; constructing an attention generative adversarial network including a generative network and a discriminative network based on an attention mechanism; training the attention generative adversarial network with the simulation image data, and adjusting the initial attention generative adversarial network parameters with the real datasets to obtain a target attention generative adversarial network; inputting the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image. It can solve the interference problem caused by the noise of the imaging system to the texture of the Fourier ptychographic microscopy reconstructed image and improve the overall visual quality of the reconstructed image.

Description

Technical Field

[0001] The present invention relates to the field of microscopic imaging technology, and in particular to a Fourier stack microscopic imaging reconstruction method, device and equipment. Background Art

[0002] Fourier ptychographic microscopy (FPM) is an emerging computational imaging technology that can effectively solve the problem of high resolution and large field of view in the field of traditional microscopic imaging. It is based on the microscope platform to collect low-resolution images of samples under LED illumination at different angles. The low-resolution images obtained under illumination at different angles correspond to different spectral information in the frequency domain. Then, the series of low-resolution images are iterated in the frequency domain using the ideas of phase recovery and synthetic aperture, which expands the frequency domain bandwidth and finally restores the high-resolution, large-field-of-view image of the sample. It does not require precise interference devices or precise mechanical scanning drives, and the device is simple and easy to operate. This makes it widely used in the fields of three-dimensional imaging, quantitative phase imaging, adaptive imaging, high-resolution macroscopic imaging, color imaging, etc.

[0003] Large field of view, high resolution and phase imaging have always been the imaging goals pursued. However, in traditional imaging systems, there are insurmountable limitations between high resolution and large field of view. The introduction of Fourier stacking microscopy technology has transferred the imaging challenges of large field of view and high resolution from the field of physical limitations to the field of computing. It has become an important research topic in the field of computational imaging because it takes into account the advantages of large field of view, high resolution and quantitative phase imaging. FPM needs to collect low-resolution images with a certain amount of redundancy. For traditional reconstruction methods, the amount of data calculation is large and the iterative reconstruction time is long. At the same time, the acquisition process has high requirements on the system environment, and the image is extremely susceptible to noise pollution. The reconstruction method based on deep learning in the existing technology simply ignores the existence of noise and integrates all the features from the measurement for prediction, which reduces the quality of the reconstructed image.

[0004] Therefore, there is an urgent need to provide a more reliable Fourier stack microscopy imaging reconstruction solution. Summary of the invention

[0005] The purpose of the present invention is to provide a Fourier stack microscopy imaging reconstruction method, device and equipment to solve the problems of large data calculation amount, long iterative reconstruction time and low reconstructed image quality in the reconstruction method of the prior art.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a Fourier stack microscopy imaging reconstruction method, the method comprising:

[0008] Obtain an image dataset; the image dataset includes simulation image data of Fourier ptychography microscopy and a real dataset, and image noise is added to the image data in the image dataset;

[0009] Construct an attention generative adversarial network based on the attention mechanism; at least a generative network and a discriminative network are included in the attention generative adversarial network; the generative network is a multi-scale residual attention network based on an encoder-decoder structure;

[0010] Train the attention generative adversarial network with the simulation image data to obtain an initial attention generative adversarial network;

[0011] Adjust the parameters of the initial attention generative adversarial network with the real dataset to obtain a target attention generative adversarial network;

[0012] Input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet preset requirements; the image data to be processed contains image noise.

[0013] Compared with the prior art, a Fourier ptychography microscopy reconstruction method provided by the present invention obtains an image dataset including simulation image data of Fourier ptychography microscopy and a real dataset, and the image dataset with Gaussian noise added; constructs an attention generative adversarial network including a generative network and a discriminative network based on the attention mechanism; trains the attention generative adversarial network with the simulation image data of Fourier ptychography microscopy to obtain an initial attention generative adversarial network; adjusts the parameters of the initial attention generative adversarial network with the real dataset to obtain a target attention generative adversarial network; inputs the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image whose resolution, amplitude, and phase meet preset requirements. The attention generative adversarial network can better solve the interference problem caused by the noise of the imaging system to the texture of the Fourier ptychography microscopy reconstruction image, generate real textures, and improve the overall visual quality of the reconstructed image.

[0014] In a second aspect, the present invention provides a Fourier ptychography microscopy reconstruction device, and the device includes:

[0015] An image dataset acquisition module, configured to obtain an image dataset; the image dataset includes simulation image data of Fourier ptychography microscopy and a real dataset, and image noise is added to the image data in the image dataset;

[0016] An attention generative adversarial network construction module, which is used to construct an attention generative adversarial network based on the attention mechanism; at least a generative network and a discriminative network are included in the attention generative adversarial network; the generative network is a multi-scale residual attention network based on an encoder-decoder structure;

[0017] An initial attention generative adversarial network training module, which is used to train the attention generative adversarial network with the simulation image data to obtain an initial attention generative adversarial network;

[0018] A target attention generative adversarial network determination module, which is used to adjust the parameters of the initial attention generative adversarial network with the real data set to obtain a target attention generative adversarial network;

[0019] A target image reconstruction module, which is used to input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet the preset requirements; the image data to be processed contains image noise.

[0020] In a third aspect, the present invention provides a Fourier ptychographic microscopy reconstruction device, and the device includes:

[0021] A communication unit / communication interface, which is used to obtain an image data set; the image data set includes simulation image data and real data sets of Fourier ptychographic microscopy, and image noise is added to the image data in the image data set;

[0022] A processing unit / processor, which is used to construct an attention generative adversarial network based on the attention mechanism; at least a generative network and a discriminative network are included in the attention generative adversarial network; the generative network is a multi-scale residual attention network based on an encoder-decoder structure;

[0023] Train the attention generative adversarial network with the simulation image data to obtain an initial attention generative adversarial network;

[0024] Adjust the parameters of the initial attention generative adversarial network with the real data set to obtain a target attention generative adversarial network;

[0025] Input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet the preset requirements; the image data to be processed contains image noise.

[0026] In a fourth aspect, the present invention provides a computer storage medium, and instructions are stored in the computer storage medium. When the instructions are run, the above-mentioned Fourier ptychographic microscopy reconstruction method is implemented.

[0027] The technical effects achieved by the device - related solutions provided in the second aspect, the equipment - related solutions provided in the third aspect, and the computer storage medium solutions provided in the fourth aspect are the same as those of the method - related solutions provided in the first aspect, and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings described herein are used to provide a further understanding of the present invention and form a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0029] Figure 1 is a schematic flow chart of a Fourier ptychography microscopy reconstruction method provided by the present invention;

[0030] Figure 2 is a schematic diagram of the overall network structure of an attention - generating adversarial network provided by the present invention;

[0031] Figure 3 is a schematic diagram of the generator network structure provided by the present invention;

[0032] Figure 4 is a schematic diagram of the discriminator network structure provided by the present invention;

[0033] Figure 5 is a schematic diagram of the structure of a Fourier ptychography microscopy reconstruction device provided by the present invention;

[0034] Figure 6 is a schematic diagram of the structure of a Fourier ptychography microscopy reconstruction equipment provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] For the convenience of clearly describing the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, terms such as "first" and "second" are used to distinguish the same items or similar items with basically the same functions and effects. For example, the first threshold and the second threshold are only used to distinguish different thresholds and do not limit their sequence. Those skilled in the art can understand that the terms "first" and "second" do not limit the quantity and execution order, and the terms "first" and "second" do not necessarily mean different.

[0036] It should be noted that in the present invention, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the present invention should not be interpreted as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, using words such as "exemplary" or "for example" aims to present relevant concepts in a specific way.

[0037] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one (piece) of the following" or similar expressions refer to any combination of these items, including any combination of single item (piece) or plural items (pieces). For example, at least one (piece) of a, b, or c can represent: a, b, c, the combination of a and b, the combination of a and c, the combination of b and c, or the combination of a, b, and c, where a, b, and c can be single or multiple.

[0038] Due to the large amount of data calculation in traditional reconstruction methods, the iterative reconstruction time is relatively long. The acquisition process has high requirements for the system environment, and the image is extremely vulnerable to noise pollution. In the prior art, the reconstruction method based on deep learning simply ignores the existence of noise and integrates all features from measurements for prediction, reducing the quality of the reconstructed image.

[0039] Based on this, in order to further improve the performance of the reconstruction method, aiming at the imaging noise and reconstruction detail problems, combined with the attention generative adversarial network, the present invention provides a Fourier ptychographic microscopy imaging reconstruction method. The attention generative adversarial network as a whole includes a generator structure and a discriminator structure. The residual attention network based on encoding and decoding is used as the generator, and the discriminator structure based on the convolutional network is trained on the simulation dataset and the real dataset by combining the pixel loss and the discriminant loss. This network can better solve the interference problem caused by the imaging system noise to the texture of the Fourier ptychographic microscopy reconstructed image, generate real textures, and improve the overall visual quality of the reconstructed image.

[0040] Next, the solution provided in the embodiments of this specification will be described in conjunction with the accompanying drawings:

[0041] As Figure 1 shown, the process may include the following steps:

[0042] Step 110: Obtain an image dataset; the image dataset includes simulation image data of Fourier ptychographic microscopy and a real dataset, and image noise is added to the image data in the image dataset.

[0043] The image dataset may include a large-scale Fourier ptychographic microscopy simulation low-resolution dataset and a small-scale real dataset. The size of the scale can be set according to the index requirements of the actual application scenario. In this solution, the basic data for training the attention generative adversarial network has a wide coverage range, and a large amount of image data is used for training.

[0044] In addition, Gaussian noise is added to the image data for training to simulate the noise environment in the actual application scenario.

[0045] Step 120: Construct an attention generative adversarial network based on the attention mechanism; at least a generative network and a discriminative network are included in the attention generative adversarial network; the generative network is a multi-scale residual attention network based on an encoder-decoder structure.

[0046] The attention generative adversarial network consists of a generative network and a discriminative network. The generative network is responsible for denoising and reconstructing the low-resolution input image, while the discriminative network determines whether the high-resolution input image is the reconstruction of the generative network or a real image. Among them, the generative network part designs the residual attention network into a multi-scale residual attention network based on an encoder-decoder structure.

[0047] Step 130: Train the attention generative adversarial network with the simulation image data to obtain an initial attention generative adversarial network.

[0048] Train the Fourier ptychography reconstruction network with the simulated low-resolution dataset; for example: the simulated dataset can be used, and Gaussian noise with a mean of zero and a standard deviation of 3×10 -4 is added for experiments. The effective reconstruction of the attention generative adversarial network is verified from the experimental reconstruction results and reconstruction result metrics.

[0049] Step 140: Adjust the parameters of the initial attention generative adversarial network with the real dataset to obtain a target attention generative adversarial network.

[0050] Verify the universality of the attention generative adversarial network on real data, and the effectiveness and generality of the method can be further tested on real data. The experiment realizes fast reconstruction by inputting real low-resolution pictures, and the universality and efficiency of the attention generative adversarial network are verified from the real data reconstruction results and reconstruction time.

[0051] Step 150: Input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet the preset requirements; the image data to be processed contains image noise.

[0052] Input the noisy samples in the test set into the trained attention generative adversarial network model, and output the high-resolution amplitude picture and high-resolution phase picture corresponding to the samples. Among them, the input samples can be low-resolution samples or image samples that need to be reconstructed in the actual application scenario.

[0053] Adopt Figure 1The overall structure of the target attention generative adversarial network constructed by the method in can be combined with Figure 2 for illustration:

[0054] In the present invention, the adversarial network Fourier ptychographic microscopy reconstruction method combining the Fourier domain attention mechanism, abbreviated as FPGAN, as Figure 2 shown, the attention generative adversarial network consists of a generative network and a discriminative network. The generative network is responsible for denoising and reconstructing the low-resolution input image, while the discriminative network determines whether the high-resolution input image is the reconstruction of the generative network or a real image. A cross-level connection is introduced. Since features at different levels have different attentions to different frequency information, the cross-level connection can enable the neural network to utilize as much complete frequency information as possible, fuse the features output at different levels, and the network can selectively highlight the deep feature information while retaining the bottom feature information and suppressing the noise channel, reducing the impact of noise input on image reconstruction, so that the model can process images affected by different degrees of noise. The discriminator is built on the basis of a convolutional network and acts as a competitor competing with the generator to better converge the entire network.

[0055] Figure 1 In the method of, an image data set including Fourier ptychographic microscopy simulation image data and a real data set, and an image data set with Gaussian noise added is obtained; an attention generative adversarial network including a generative network and a discriminative network is constructed based on the attention mechanism; the attention generative adversarial network is trained using the Fourier ptychographic microscopy simulation image data to obtain an initial attention generative adversarial network; the parameters of the initial attention generative adversarial network are adjusted using the real data set to obtain a target attention generative adversarial network; the image data to be processed is input into the target attention generative adversarial network to obtain a target image with the resolution, amplitude, and phase of the reconstructed image meeting the preset requirements. The attention generative adversarial network can better solve the interference problem caused by the noise of the imaging system to the texture of the Fourier ptychographic microscopy reconstructed image, generate real textures, and improve the overall visual quality of the reconstructed image.

[0056] Based on Figure 1 the method of, some specific embodiments of this method are also provided in this specification and will be described below.

[0057] Before obtaining the image data set, it may further include:

[0058] Collecting basic image data during the Fourier ptychographic microscopy simulation process; Gaussian noise with a preset mean and standard deviation is added to the basic image data;

[0059] Synthesizing the basic image data in the Fourier domain and converting it into two-channel image data through inverse Fourier transform;

[0060] A set of image data forms an image data set. Each set of image data contains a dual-channel image representing intensity and a dual-channel image representing phase.

[0061] In the present invention, the trained target attention generative adversarial network is the reconstruction network for reconstructing Fourier ptychographic microscopy. This reconstruction network uses dual-channel input of phase and intensity. Therefore, the low-resolution data of Fourier ptychography is synthesized in the Fourier domain and transformed into dual-channel data through inverse Fourier transform, and then used as the input of the network. Similarly, the output of the network is also a dual-channel image of intensity and phase. The large-scale simulation data set is constructed using simulated data with a certain information complexity. The data set can be composed of 25,600 sets of high-resolution image data. Each set contains two images representing the dual channels of intensity and phase respectively. The combination of intensity and phase in each of the 25,600 sets of high-resolution image data is used as the output ground truth of the network. Next, the simulation process is illustrated by examples:

[0062] (1) First, synthesize the complex amplitude data from each set of high-resolution intensity and phase images;

[0063] (2) The synthesized complex amplitude data is simulated through the Fourier ptychography model. Gaussian noise with a mean of 0 and a standard deviation of 3×10 -4 is added during the simulation imaging process to simulate the noise in the actual imaging process, generating a set of Fourier ptychography low-resolution data;

[0064] (3) The Fourier ptychography low-resolution data is synthesized into low-resolution complex amplitude through the traditional Fourier ptychography reconstruction algorithm. Different from the traditional Fourier ptychography reconstruction algorithm, the reconstruction process here only iterates once;

[0065] (4) The synthesized low-resolution complex amplitude is respectively output as intensity and phase, and finally 25,600 sets of intensity and phase simulation inputs are obtained.

[0066] Through the above method, the obtained image data set contains a large amount of simulated image data and real data set. The simulated image data is used to train the attention generative adversarial network, and the real data set is used to adjust the parameters of the initial attention generative adversarial network to adapt to real image data. And Gaussian noise is added to the image data in the image data set, which can simulate the real scene.

[0067] Optionally, an attention generative adversarial network is constructed based on the attention mechanism, which specifically may include:

[0068] An attention generative adversarial network is constructed based on the generative network and the discriminative network based on the attention mechanism; the generative network includes an encoding module, a cross-level connection module, and a decoding module;

[0069] The encoding module stacks residual groups using a nested residual network;

[0070] The decoding module gradually enlarges the features through sub-pixel convolution. The network introduces a Fourier domain attention mechanism at the tail of each encoding module and decoding module, and uses the spectral features of different feature maps in the Fourier domain, enabling the network to adaptively readjust each feature map according to the comprehensive contribution of all frequency components contained in its spectrum;

[0071] The cross-level connection module is used for the fusion of high-level and low-level semantic features, and introduces a cross-level connection of the same scale between the encoding and decoding modules for feature fusion.

[0072] Further, the image data to be processed is input into the target attention generative adversarial network to obtain the reconstructed target image, which specifically may include:

[0073] The generative network extracts and fuses the features in the Fourier ptychography microscopy simulation image data through cross-level connections; the generative network outputs the reconstructed target image by gradually performing the encoding, cross-level connection, and decoding processes;

[0074] The cross-level connection module is used for the fusion of high-level and low-level semantic features, and introduces a cross-level connection method of the same scale between the encoding and decoding modules for feature fusion;

[0075] The decoding module upsamples the hierarchical feature map. The decoding module consists of a splicing layer, an upsampling layer, and an attention layer; the upsampling layer selects sub-pixel convolution operations to enlarge the encoded features, and the sub-pixel convolution layer upsamples by combining the information in the channels;

[0076] The decoding module expands the number of feature channels through convolution or splices and fuses the underlying feature maps of the autoencoder part according to the feature channel dimension;

[0077] The spatial resolution of the fused underlying feature map is doubled and the number of channels is halved through a preset upsampling operation, and the deep feature information is adaptively highlighted through the Fourier domain attention mechanism;

[0078] Based on the deep feature information after splicing and fusion, the target image is obtained through convolution processing.

[0079] Among them, the encoding module consists of a residual attention group and a convolution layer; the first layer of the generative network uses a convolution layer to receive the input Fourier ptychography microscopy simulation image data for preliminary extraction to obtain shallow feature information;

[0080] Extract hierarchical feature maps from the input through the encoding module; each encoding module includes a convolutional layer activated by a rectified linear unit for downsampling and a residual attention group for feature extraction; the residual attention group is connected using a nested residual structure, each residual attention group contains multiple residual blocks, and each residual block consists of a convolution with a preset kernel and a rectified linear unit.

[0081] The generation network part designs the residual attention network into a multi-scale residual attention network with an encoder-decoder structure, and outputs the reconstructed intensity and phase images by gradually performing the encoding, cross-level connection, and decoding processes. The encoding module stacks residual groups using a nested residual network, and the decoding module gradually enlarges the features through sub-pixel convolution. Among them, the network introduces a Fourier domain attention mechanism at the end of each encoding and decoding module, and uses the spectral features of different feature maps in the Fourier domain to enable the network to adaptively readjust each feature map according to the comprehensive contribution of all frequency components contained in its spectrum, helping the network focus on more useful channel information. The discriminator based on the convolutional network acts as a competitor competing with the generator during training, uses binary cross-entropy as the method for calculating the discriminant loss, and outputs the predicted probability.

[0082] As Figure 3 shown, in the generation network structure, it includes a convolutional layer, a downsampling module, a residual attention group, a Fourier channel attention, a splicing layer, and an upsampling module. The main parts are the residual attention group and the convolutional layer. Next, combined with Figure 3 take an example to illustrate the network structure of the generation network:

[0083] The first layer of the network can use a 3×3 convolutional layer to receive the input low-resolution image and initially extract shallow features, and then extract hierarchical feature maps from the input through K = 4 consecutive identical modules. Each module includes a convolutional layer activated by a rectified linear unit for downsampling and a residual attention group for feature extraction. A convolutional layer with a stride of 2 and a kernel of 3 is used to replace the pooling layer to compress and remove redundant information from the features. The residual group is connected using a nested residual structure, each residual group contains 4 residual blocks, a Fourier attention mechanism, and a long skip connection. Each residual block consists of a 3×3 convolution and a rectified linear unit (ReLU). With this structure, the feature information and gradient information in the neural network can be transmitted more effectively between convolutional layers.

[0084] The cross-level connection is used for the fusion of high-level and low-level semantic features. A cross-level connection of the same scale is introduced between the encoding and decoding modules to perform feature fusion in a manner similar to a feature pyramid. In this way, more features can be fused without a large amount of additional cost.

[0085] The decoding network performs upsampling on the feature map and consists of a splicing layer, an upsampling layer, and an attention layer. Among them, the upsampling layer selects sub-pixel convolution operation to magnify the encoded features, and the sub-pixel convolution layer performs upsampling by merging the information in the channels. Each module first expands the number of feature channels through convolution or splices and fuses the underlying feature maps from the encoding part according to the feature channel dimension. Then, it doubles the spatial resolution of the input feature map and halves the number of channels through a two-fold upsampling operation. Next, it adaptively highlights the deep feature information through a Fourier domain attention mechanism. Finally, the spliced and fused deep feature information is output as a two-channel complex field through a 3×3 convolution activated by relu and a 3×3 convolution activated by Sigmoid.

[0086] The above-mentioned generation network part designs the residual attention network into a multi-scale residual attention network with an encoder-decoder structure, and better extracts and fuses the information in the low-resolution image through cross-level connection.

[0087] Optionally, the discriminative network is used to identify whether the input image is a real image or whether the input image is an image output by the generation network;

[0088] The discriminative network uses a high-resolution image as the input and passes through multiple cascaded Conv-LReLU modules;

[0089] The output information of the LReLU activation function is input to the global pooling layer, the fully connected layer, the ReLU layer, and the Sigmoid activation function to output the prediction probability, and based on the prediction probability, it is identified whether the input image is a real image or whether the input image is an image output by the generation network.

[0090] The discriminative network is used to identify whether the input is an image output by the generation network or a real high-resolution image. Since the generation network and the discriminative network are in a game relationship, the performance of the discriminative model will also affect the performance phenotype of the generation model. If the discriminative network is not well trained and the model cannot accurately identify the data, then the "fake" samples generated by the generation model will also be considered real by the discriminative network, making the training of the generation network difficult. Only when the discriminative network is correctly trained can it provide effective gradients for training the generation network, and after iterative training, both models are finally optimized. The structure of the discriminative network can be combined with Figure 4 An example is given for illustration:

[0091] During the training process, the discriminator network competes with the generator as a competitor. It can take a high-resolution image as input and start from 5 cascaded Conv-LReLU modules. The number of convolutional filters in each module is 32, 32, 64, 128, and 256 respectively. The output of the last LReLU activation function is sequentially input into a global pooling layer, a fully connected layer, a ReLU layer, another fully connected layer, and a Sigmoid activation function, and then the predicted probability is output.

[0092] In the implementation process of the above scheme, when training the attention generative adversarial network, the loss function and the parameters used during training can be set. More specifically, the objective functions of the generator network G and the discriminator network D are defined separately. The objective function of the generator network is defined as a combination of MSE loss and SSIM loss. The MSE loss ensures pixel accuracy and balances the dynamic range of the prediction, and the SSIM loss enhances the structural similarity of the output. The objective function of the discriminator network D is defined as binary cross-entropy.

[0093] Furthermore, when designing the loss function, it can be implemented based on the following methods:

[0094] The objective function of the generator network is defined as a combination of MSE loss and SSIM loss. The MSE loss ensures pixel accuracy and balances the dynamic range of the prediction, and the SSIM loss enhances the structural similarity of the output. If is defined as the output of the generator, Y is defined as the corresponding GT, and (w, h) is defined as the pixel size of the output image, then the calculation of the objective function can be expressed by formula (1) as formula (1):

[0095]

[0096] Among them, in formula (1), λ is a scalar weight used to balance the relative contributions of SSIM and MSE, and it is set to 0.1 in most cases.

[0097] Objective function: The objective functions of the generator network G and the discriminator network D are defined separately. The objective function of the discriminator network D is defined as binary cross-entropy. The objective function of the generator network G is the sum of two terms: SR

[0098] error (the difference between the output of the generator network G and the real image) and discriminator error L D ,G objective function can be expressed by formula (2):

[0099]

[0100] Among them, X is the input low-resolution image, and Y is the SR target image. β, γ, and λ are scalar weighting factors that balance the corresponding terms, and are empirically set to β = 0.1, γ = 1, and λ = 0.1.

[0101] The training process of the generation network G and the discriminator network D is like a two-player game. Starting from and L D As can be seen, the goal of the generation network G is to fool the discriminator network D and take its output as the ground truth. On the contrary, the goal of the discriminator network D is to determine whether the input image comes from the generation network G or from the ground truth. The generation network G and the discriminator network D are trained in an interleaved scheme, so they compete with each other and finally reach a balanced state.

[0102] During the training process, the Adam optimizer is used, and 16 batches are used according to the network scale. For the GAN-based FPGAN method, the network is randomly initialized and the generation model and the discriminator model are trained with a typical starting learning rate of 2×10 -5 The final model is trained through approximately 50,000 mini-batch iterations, which takes 70 hours. In each iteration, the generator G and the discriminator D are updated three times and two times respectively. The training process is precisely controlled through the mini-batch training method, and the mixed-precision training shortens the training time.

[0103] When the present invention is specifically implemented, it has good technical effects. Next, the technical effects of the present invention will be illustrated with actual experimental implementation data:

[0104] (1) Experimental data

[0105] In order to improve the processing effect of the attention generative adversarial network reconstruction model on real system data, in addition to the simulation data set, the present invention also constructs a small-scale real data set collected by an experimental system.

[0106] During the collection of the real data set, the parameters of the experimental imaging system are first set. The numerical aperture NA of the objective lens is set to 0.13, the illumination wavelength is set to 0.505 μm, the distance between the sample and the 13×13 programmable control light source element LED matrix is 98 mm, the gap between adjacent LEDs is 8 mm, and the system magnification is set to 4 times.

[0107] After the imaging system parameters are set, you can start collecting images. First, place a suitable thin sample on the stage and provide sufficient lighting. Then, focus the microscope with camera imaging. Finally, control the on and off of the LED light board through the serial port to collect and store Fourier stack data. After being collected by the imaging system, the real data sample contains 50 sets of original Fourier stack data, and the results of the traditional iterative reconstruction algorithm are used as the true value. Then these data are randomly cropped to obtain 450 sets of input data and true values, of which 400 sets are training sets and 50 sets are test sets. In order to prevent the training results from being overfitted due to too little real data, 400 sets of data were randomly selected from the simulated data and added to the real training data set to form a small-scale real data set.

[0108] (2) Parameter settings

[0109] The experimental training was conducted on a computer equipped with a Tesla V100 graphics processing card (NVIDIA). During the training process, the Adam optimizer was used and a batch size of 16 was used according to the network size. For the GAN-based FPGAN method, the network was randomly initialized and the number of nodes was 2×10 -5 The generative model and the discriminative model were trained with a typical starting learning rate of . The final model was trained for about 50,000 mini-batch iterations, which took 70 hours. In each iteration, the generator G and the discriminator D were updated three times and twice, respectively. The training process was accurately controlled by the mini-batch training method, and the mixed precision training shortened the training time. Once the network is trained, the model usually takes less than 1 second to reconstruct a 192×192 pixel image. In order to verify the universality of the attention-generated adversarial network on real data, the effectiveness and versatility of the method were further tested on real data. The experiment achieved fast reconstruction by inputting real low-resolution pictures, and verified the universality and efficiency of the attention-generated adversarial network from the reconstruction results and reconstruction time of real data.

[0110] (3) Real data reconstruction results

[0111] Under experimental conditions, it is impossible to obtain the true values ​​of intensity and phase images corresponding to real data for quantitative evaluation. Therefore, the experiment compares the reconstruction results of different reconstruction methods for qualitative evaluation. The reconstructed phase results produced by FPGAN are of higher quality, with more prominent details and clearer cell contours. In particular, the phase image reconstructed by FPGAN has a larger contrast and a smooth background, which verifies its superiority in eliminating real noise.

[0112] (4) Reconstruction time comparison experiment

[0113] In terms of reconstruction time, the results of the FPGAN reconstruction method proposed in the present invention are shown in Table 1.

[0114] Reconstruction time of the method in Table 1

[0115]

[0116] The FPGAN network reconstruction adopted by the present invention constructs a reconstruction network model based on deep learning. Once the training is completed, without the need for the number of iterations, it only takes less than 0.1 second to reconstruct the image in one step. Therefore, they all have the least reconstruction time.

[0117] (5) Ablation experiment of the adversarial network

[0118] To verify the influence of the generative adversarial strategy on network reconstruction after adding the discriminative network, an ablation experiment of the adversarial network was designed. With other variables kept the same, the training and validation metrics of the generative network improved by residual attention before and after adding the discriminative network were compared. In the phase-normalized root mean square error curve, the initial convergence was basically the same, and the error continuously decreased as the training progressed. After a certain number of training cycles, the normalized root mean square of the attention generative adversarial network finally maintained a lower error value. The attention generative adversarial network had a higher peak signal-to-noise ratio than the improved generative network, indicating that the attention generative adversarial network had better performance in the training and validation results.

[0119] (6) Noise robustness experiment

[0120] To test the robustness of the method under different noises, an ablation study was conducted. Different levels of noise were added to the input image for reconstruction. Specifically, Gaussian noises with zero mean and standard deviations of 1×10 -4 , 2×10 -4 , 3×10 -4 were added during the imaging simulation process. Table 2 shows the quantitative results of the reconstruction method.

[0121] Table 2 Quantification indexes of reconstruction results under different noise conditions

[0122]

[0123] From the data in the table, as the noise intensity increased, the mean square error (MSE) of the reconstructed image did not show an obvious decrease, and the peak signal-to-noise ratio (PSNR) decreased slightly but with small fluctuations and still had a relatively stable value, indicating that both methods could stably reconstruct images under different noise conditions and had a certain degree of robustness; under the same noise conditions, the attention generative adversarial network had a higher peak signal-to-noise ratio and a smaller mean square error value than the residual attention network, especially for the phase image. This shows that the attention generative adversarial network reconstruction method can effectively suppress the input noise, thereby improving the reconstruction quality.

[0124] It should be noted that in the present invention, the generation network introduces a Fourier-domain attention mechanism at the end of each encoding and decoding module. By utilizing the spectral features of different feature maps in the Fourier domain, the network can adaptively readjust each feature map according to the comprehensive contribution of all frequency components contained in its spectrum, helping the network focus on more useful channel information. The cross-level connections in the generation network are used for the fusion of high- and low-level semantic features. Cross-level connections of the same scale are introduced between the encoding and decoding modules for feature fusion in a manner similar to a feature pyramid, improving the utilization rate of semantic features at different levels. The attention generative adversarial network adopts dual-channel input of phase and intensity. The low-resolution data of Fourier ptychography is synthesized in the Fourier domain and transformed into dual-channel data through inverse Fourier transform, and then used as the input of the network. Similarly, the output of the network is also a dual-channel image of intensity and phase. The experimental results of simulation and real data show that the attention generative adversarial network can better solve the interference problem caused by the noise of the imaging system to the texture of the Fourier ptychographic microscopy reconstruction image, generate real textures, and improve the overall visual quality of the reconstructed image.

[0125] Based on the same idea, the present invention also provides a Fourier ptychographic microscopy reconstruction device, as Figure 5 shown, the device may include:

[0126] An image dataset acquisition module 510, configured to acquire an image dataset; the image dataset includes simulation image data and real datasets of Fourier ptychographic microscopy, and image noise is added to the image data in the image dataset;

[0127] An attention generative adversarial network construction module 520, configured to construct an attention generative adversarial network based on an attention mechanism; at least a generation network and a discriminator network are included in the attention generative adversarial network; the generation network is a multi-scale residual attention network based on an encoder-decoder structure;

[0128] An initial attention generative adversarial network training module 530, configured to train the attention generative adversarial network using the simulation image data to obtain an initial attention generative adversarial network;

[0129] A target attention generative adversarial network determination module 540, configured to adjust the parameters of the initial attention generative adversarial network using the real dataset to obtain a target attention generative adversarial network;

[0130] A target image reconstruction module 550, configured to input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet preset requirements; image noise is included in the image data to be processed.

[0131] Based onFigure 5 The device in

[0132] Optionally, the attention generation adversarial network construction module 520 can specifically be used for:

[0133] Constructing an attention generation adversarial network based on a generation network and a discriminative network based on an attention mechanism; the generation network includes an encoding module, a cross-level connection module, and a decoding module;

[0134] The encoding module stacks residual groups using a nested residual network;

[0135] The decoding module gradually magnifies the features through sub-pixel convolution, and the network introduces a Fourier domain attention mechanism at the tail of each encoding module and decoding module, and uses the spectral features of different feature maps in the Fourier domain to enable the network to adaptively readjust each feature map according to the comprehensive contribution of all frequency components contained in its spectrum;

[0136] The cross-level connection module is used for the fusion of high-level and low-level semantic features, and introduces a same-scale cross-level connection between the encoding and decoding modules for feature fusion;

[0137] The discriminative network calculates the discriminative loss using binary cross-entropy and outputs a prediction probability.

[0138] Optionally, the device may further include:

[0139] A basic image data acquisition module, configured to acquire basic image data during the Fourier ptychography microscopy simulation process; Gaussian noise with a preset mean and standard deviation is added to the basic image data;

[0140] A synthesis module, configured to synthesize the basic image data in the Fourier domain and convert it into two-channel image data through inverse Fourier transform;

[0141] An image data set determination module, configured to form an image data set from multiple groups of image data, and each group of image data includes a two-channel image representing intensity and a two-channel image representing phase.

[0142] Optionally, the device further includes:

[0143] A loss function setting module, configured to set the loss function during network learning and training; the objective function of the discriminative network is defined as binary cross-entropy; the objective function of the generation network is a combination of MSE loss and SSIM loss;

[0144] A parameter setting module, configured to set the parameter settings during network learning and training; the parameters at least include the optimizer type, network initialization method, number of iterative training times, and training time during the training process.

[0145] Optionally, the target image reconstruction module 550 can specifically be used for:

[0146] The generation network extracts and fuses features in the Fourier ptychographic microscopy simulation image data through cross-level connections; the generation network outputs the reconstructed target image by gradually performing an encoding, cross-level connection, and decoding process.

[0147] The cross-level connection module is used for the fusion of high-level and low-level semantic features, and introduces a same-scale cross-level connection method between the encoding and decoding modules for feature fusion.

[0148] The decoding module performs upsampling on the hierarchical feature maps. The decoding module consists of a splicing layer, an upsampling layer, and an attention layer; the upsampling layer selects sub-pixel convolution operations to magnify the encoded features, and the sub-pixel convolution layer performs upsampling by combining information in the channels.

[0149] The decoding module expands the number of feature channels through convolution or splices and fuses the underlying feature maps of the auto-encoding part according to the feature channel dimension.

[0150] The spatial resolution of the fused underlying feature maps is doubled and the number of channels is halved through a preset upsampling operation, and the deep feature information is adaptively highlighted through the Fourier domain attention mechanism.

[0151] Based on the deep feature information after splicing and fusion, the target image is obtained through convolution processing.

[0152] Optionally, the encoding module consists of a residual attention group and a convolution layer; the first layer of the generation network uses a convolution layer to receive the input Fourier ptychographic microscopy simulation image data for preliminary extraction to obtain shallow feature information.

[0153] Hierarchical feature maps are extracted from the input through the encoding module; each encoding module includes a convolution layer activated by a rectified linear unit for downsampling and a residual attention group for feature extraction; the residual attention group is connected using a nested residual structure, and each residual attention group contains multiple residual blocks, and each residual block consists of a convolution with a preset kernel and a rectified linear unit.

[0154] Optionally, the discriminative network is used to identify whether the input image is a real image or to identify whether the input image is an image output by the generation network.

[0155] The discriminative network uses a high-resolution image as the input and passes through multiple cascaded Conv-LReLU modules.

[0156] The output information of the LReLU activation function is input into a global pooling layer, a fully connected layer, a ReLU layer, and a sigmoid activation function to output a prediction probability, and based on the prediction probability, it is identified whether the input image is a real image or whether the input image is an image output by a generative network.

[0157] Based on the same idea, an embodiment of this specification also provides a Fourier ptychographic microscopy reconstruction device. As Figure 6 shown, it may include:

[0158] A communication unit / communication interface for acquiring an image dataset; the image dataset includes simulation image data and a real dataset of Fourier ptychographic microscopy, and image noise is added to the image data in the image dataset;

[0159] A processing unit / processor for constructing an attention generative adversarial network based on an attention mechanism; at least a generative network and a discriminative network are included in the attention generative adversarial network; the generative network is a multi-scale residual attention network based on an encoder-decoder structure;

[0160] The attention generative adversarial network is trained using the simulation image data to obtain an initial attention generative adversarial network;

[0161] The parameters of the initial attention generative adversarial network are adjusted using the real dataset to obtain a target attention generative adversarial network;

[0162] The image data to be processed is input into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet preset requirements; the image data to be processed contains image noise.

[0163] As Figure 6 shown, the above terminal device may further include a communication line. The communication line may include a path for transmitting information between the above components.

[0164] Optionally, as Figure 6 shown, the terminal device may further include a memory. The memory is used to store computer execution instructions for implementing the solution of the present invention and is controlled by the processor for execution. The processor is used to execute the computer execution instructions stored in the memory, thereby implementing the method provided by the embodiment of the present invention.

[0165] In a specific implementation, as an embodiment, as Figure 6 shown, the processor may include one or more CPUs, such as Figure 6 CPU0 and CPU1 in

[0166] In a specific implementation, as an embodiment, asFigure 6 As shown, the terminal device may include multiple processors, such as Figure 6 the processors in . Each of these processors can be a single-core processor or a multi-core processor.

[0167] The above mainly introduces the solution provided by the embodiments of the present invention from the perspective of the interaction between various modules. It can be understood that, in order to implement the above functions, each module includes the corresponding hardware structure and / or software unit for executing each function. Those skilled in the art should easily realize that, combining the units and algorithm steps of each example described in the embodiments disclosed in this article, the present invention can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed in the way of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described function for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0168] The embodiments of the present invention can divide functional modules according to the above method examples. For example, each functional module can be divided corresponding to each function, or two or more functions can be integrated into one processing module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module. It should be noted that the division of modules in the embodiments of the present invention is illustrative, only a logical function division, and there can be other division methods in actual implementation.

[0169] The processor in this specification can also have the function of a memory. The memory is used to store the computer execution instructions for executing the solution of the present invention and is controlled by the processor to execute. The processor is used to execute the computer execution instructions stored in the memory, thereby implementing the method provided by the embodiments of the present invention.

[0170] The memory can be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or it can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory can exist independently and be connected to the processor through a communication line. The memory can also be integrated with the processor.

[0171] Optionally, the computer-executable instructions in the embodiments of the present invention can also be referred to as application program code, and the embodiments of the present invention do not make specific limitations thereto.

[0172] The methods disclosed in the above embodiments of the present invention can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, the steps of the above methods can be completed through the integrated logic circuit in the hardware of the processor or instructions in software form. The above-mentioned processor can be a general-purpose processor, a digital signal processor (DSP), an ASIC, a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the methods disclosed in combination with the embodiments of the present invention can be directly embodied as being executed and completed by a hardware decoding processor, or by a combination of the hardware and software modules in the decoding processor. The software modules can be located in mature storage media in the art such as random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers, etc. This storage media is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above methods.

[0173] In one possible implementation, a computer-readable storage medium is provided, and instructions are stored in the computer-readable storage medium. When the instructions are run, they are used to implement the method in the above embodiments.

[0174] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present invention are executed in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a terminal, a user device, or other programmable devices. The computer program or instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer program or instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired or wireless manner. The computer-readable storage medium can be any available medium that the computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; it can also be an optical medium, such as a digital video disc (DVD); it can also be a semiconductor medium, such as a solid state drive (SSD).

[0175] Although the present invention has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed invention, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality. A single processor or other unit can implement several functions recited in the claims. Certain measures are recited in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0176] Although the present invention has been described in connection with specific features and their embodiments, it will be apparent that various modifications and combinations can be made without departing from the spirit and scope of the invention. Accordingly, this specification and the drawings are merely exemplary illustrations of the invention defined by the appended claims and are considered to cover any and all modifications, variations, combinations or equivalents within the scope of the invention. Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A Fourier ptychographic microscopy reconstruction method, characterized in that The method includes: Obtaining an image dataset; the image dataset includes simulated image data of Fourier ptychography microscopy and a real dataset, and image noise is added to the image data in the image dataset; Constructing an attention generative adversarial network based on an attention mechanism; The attention generative adversarial network includes at least a generative network and a discriminative network; the generative network is a multi-scale residual attention network based on an encoder-decoder structure; Training the attention generative adversarial network with the simulated image data to obtain an initial attention generative adversarial network; Adjusting the parameters of the initial attention generative adversarial network with the real dataset to obtain a target attention generative adversarial network; Inputting the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet preset requirements; The image data to be processed contains image noise; Constructing an attention generative adversarial network based on an attention mechanism specifically includes: Constructing an attention generative adversarial network based on a generative network and a discriminative network based on an attention mechanism; the generative network includes an encoding module, a cross-level connection module, and a decoding module; The encoding module stacks residual groups using a nested residual network; The decoding module gradually enlarges features through sub-pixel convolution. The network introduces a Fourier domain attention mechanism at the end of each encoding module and decoding module, and uses the spectral features of different feature maps in the Fourier domain to enable the network to adaptively readjust each feature map according to the comprehensive contribution of all frequency components contained in its spectrum; The cross-level connection module is used for the fusion of high-level and low-level semantic features, and introduces a same-scale cross-level connection between the encoding and decoding modules for feature fusion; The discriminative network calculates the discriminative loss using binary cross-entropy and outputs a prediction probability.

2. The Fourier ptychographic microscopy reconstruction method according to claim 1, wherein Before obtaining the image dataset, it further includes: Collecting basic image data during the simulation process of Fourier ptychography microscopy; Gaussian noise with a preset mean and standard deviation is added to the basic image data; Synthesizing the basic image data in the Fourier domain and converting it into two-channel image data through inverse Fourier transform; Multiple groups of image data form an image dataset, and each group of image data contains a two-channel image representing intensity and a two-channel image representing phase.

3. The Fourier ptychographic microscopy reconstruction method according to claim 1, wherein Before constructing an attention generative adversarial network based on an attention mechanism, it further includes: Setting the loss function during network learning and training; the objective function of the discriminative network is defined as binary cross-entropy; the objective function of the generative network is a combination of MSE loss and SSIM loss; Setting the parameter settings during network learning and training; the parameters at least include the optimizer type, network initialization method, number of iterative training times, and training time during the training process.

4. The Fourier ptychographic microscopy reconstruction method according to claim 1, wherein Inputting the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image specifically includes: The generation network extracts and fuses features in the Fourier ptychography microscopy simulation image data through cross-level connections; the generation network outputs a reconstructed target image by gradually performing an encoding, cross-level connection, and decoding process; The cross-level connection module is used for the fusion of high-level and low-level semantic features, and a cross-level connection method of the same scale is introduced between the encoding and decoding modules for feature fusion; The decoding module performs upsampling on the hierarchical feature map. The decoding module consists of a splicing layer, an upsampling layer, and an attention layer; the upsampling layer selects sub-pixel convolution operation to magnify the encoded features, and the sub-pixel convolution layer performs upsampling by combining information in the channels; The decoding module expands the number of feature channels through convolution or splices and fuses the underlying feature maps of the autoencoder part according to the feature channel dimension; The spatial resolution of the fused underlying feature map is doubled and the number of channels is halved through a preset upsampling operation, and the deep feature information is adaptively highlighted through the Fourier domain attention mechanism; Based on the deep feature information after splicing and fusion, the target image is obtained through convolution processing.

5. The Fourier ptychographic microscopy reconstruction method according to claim 4, wherein The encoding module consists of a residual attention group and a convolutional layer; the first layer of the generation network uses a convolutional layer to receive the input Fourier ptychography microscopy simulation image data for preliminary extraction to obtain shallow feature information; Hierarchical feature maps are extracted from the input through the encoding module; each encoding module includes a convolutional layer activated by a rectified linear unit function for downsampling and a residual attention group for feature extraction; the residual attention group is connected using a nested residual structure, and each residual attention group contains multiple residual blocks, and each residual block consists of a convolution with a preset kernel and a rectified linear unit.

6. The Fourier ptychographic microscopy reconstruction method according to claim 1, wherein The discriminative network is used to identify whether the input image is a real image or to identify whether the input image is an image output by the generation network; The discriminative network uses a high-resolution image as the input and passes through multiple cascaded Conv-LReLU modules; The output information of the LReLU activation function is input to a global pooling layer, a fully connected layer, a ReLU layer, and a Sigmoid activation function to output a prediction probability, and based on the prediction probability, it is identified whether the input image is a real image or whether the input image is an image output by the generation network.

7. A Fourier ptychographic microscopy reconstruction device, characterized in that, The device is applied to the Fourier ptychography microscopy reconstruction method according to any one of claims 1-6. The device includes: An image dataset acquisition module for acquiring an image dataset; the image dataset includes Fourier ptychography microscopy simulation image data and a real dataset, and image noise is added to the image data in the image dataset; An attention generative adversarial network construction module for constructing an attention generative adversarial network based on the attention mechanism; at least a generation network and a discriminative network are included in the attention generative adversarial network; the generation network is a multi-scale residual attention network based on an encoding-decoding structure; An initial attention generative adversarial network training module for training the attention generative adversarial network using the simulation image data to obtain an initial attention generative adversarial network; A target attention generative adversarial network determination module, configured to adjust parameters of the initial attention generative adversarial network by using the real data set to obtain a target attention generative adversarial network; A target image reconstruction module, configured to input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet preset requirements; the image data to be processed contains image noise; The attention generative adversarial network construction module is specifically configured to: Construct an attention generative adversarial network based on a generative network and a discriminative network based on an attention mechanism; the generative network includes an encoding module, a cross-level connection module, and a decoding module; The encoding module stacks residual groups by using a nested residual network; The decoding module gradually amplifies features through sub-pixel convolution. The network introduces a Fourier domain attention mechanism at the tail of each encoding module and decoding module, and uses the spectral features of different feature maps in the Fourier domain to enable the network to adaptively readjust each feature map according to the comprehensive contribution of all frequency components included in its spectrum; The cross-level connection module is used for the fusion of high-level and low-level semantic features, and introduces a same-scale cross-level connection between the encoding and decoding modules for feature fusion; The discriminative network calculates the discriminative loss by using binary cross entropy and outputs a prediction probability.

8. A Fourier ptychographic microscopy reconstruction device, characterized in that, The device is applied to the Fourier ptychographic microscopy reconstruction method according to any one of claims 1-6. The device includes: A communication unit / communication interface, configured to obtain an image data set; the image data set includes simulation image data and real data set of Fourier ptychographic microscopy, and image noise is added to the image data in the image data set; A processing unit / processor, configured to construct an attention generative adversarial network based on an attention mechanism; the attention generative adversarial network includes at least a generative network and a discriminative network; the generative network is a multi-scale residual attention network based on an encoding-decoding structure; Train the attention generative adversarial network by using the simulation image data to obtain an initial attention generative adversarial network; Adjust parameters of the initial attention generative adversarial network by using the real data set to obtain a target attention generative adversarial network; Input the image data to be processed into the target attention generative adversarial network to obtain a reconstructed target image; the target image is an image whose resolution, amplitude, and phase meet preset requirements; the image data to be processed contains image noise; Constructing an attention generative adversarial network based on an attention mechanism specifically includes: Construct an attention generative adversarial network based on a generative network and a discriminative network based on an attention mechanism; the generative network includes an encoding module, a cross-level connection module, and a decoding module; The encoding module stacks residual groups by using a nested residual network; The decoding module gradually magnifies the features through sub-pixel convolution, where the network introduces a Fourier-domain attention mechanism at the tail of each encoding module and decoding module, and uses the spectral features of different feature maps in the Fourier domain to enable the network to adaptively readjust each feature map according to the comprehensive contribution of all frequency components contained in its spectrum; The cross-level connection module is used for the fusion of high- and low-level semantic features, and a same-scale cross-level connection is introduced between the encoding and decoding modules for feature fusion; The discriminant network calculates the discriminant loss using binary cross-entropy and outputs the prediction probability.

9. A computer storage medium, characterized in that, Instructions are stored in the computer storage medium, and when the instructions are run by the processor, the Fourier ptychography microscopy imaging reconstruction method according to any one of claims 1 to 6 is implemented.

Images