Image denoising system and method based on multi-scale information extraction

Abstract

The application discloses an image denoising system and method based on multi-scale information extraction, comprising a microscopic image acquisition device and an image denoising device; the microscopic image acquisition device is used for collecting microscopic image data sets for training and debugging, and is also used for collecting cell images of samples to be processed; the microscopic image acquisition device is used for training and debugging a multi-stage denoising network model by using the microscopic image data sets, obtaining a multi-stage denoising network model with optimal parameters, and is also used for processing input cell images of samples to be processed by using the multi-stage denoising network model and outputting denoised images. The application can flexibly adjust the axial distance of the measured sample relative to the objective lens, so that focusing at different depths is realized. On this basis, the multi-stage denoising network model is used for restoring and processing the cell images containing noise, and remarkable effects are achieved in noise removal and texture information preservation.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to an image denoising system and method based on multi-scale information extraction. Background Technology

[0002] Currently, image denoising methods include traditional image denoising methods and deep learning-based denoising methods. Traditional image denoising methods can be further divided into spatial domain pixel feature denoising algorithms and transform domain denoising algorithms; the former performs processing directly in the image space, while the latter performs processing indirectly in the image transform domain.

[0003] Methods based on spatial pixel features obtain new center pixel values ​​by analyzing the direct relationship between the center pixel and its neighboring pixels in grayscale space within a window of a certain size. Therefore, a typical input parameter, the filtering radius, is often used. The filtering radius can be used to calculate pixel similarity within a local window and also serves as the calculation window for some Gaussian or Laplacian operators. Among neighborhood filtering methods, the most representative methods include: arithmetic mean filtering, Gaussian filtering, statistical median filtering, guided filtering, and the NLM (Non-Local Means) algorithm. While these methods can filter out noise in cell microscopy images, they all share a common problem: during noise reduction, they often lose detailed or edge information of the cell microscopy image, leading to image blurring and bokeh.

[0004] The basic idea of ​​image transform domain denoising algorithms is to first perform a transformation to convert the image from the spatial domain to the transform domain. Then, the noise is divided into high, medium, and low frequency noise. This transform domain method can separate noise of different frequencies. Finally, an inverse transform is performed to convert the image from the transform domain back to the original spatial domain, ultimately achieving the goal of removing image noise. There are many methods for converting images from the spatial domain to the transform domain, among which the most representative are Fourier transform, discrete cosine transform, wavelet transform, and multi-scale geometric analysis methods. Although the above methods are superior to methods based on spatial domain pixel features in terms of denoising performance, they still lose detailed and edge information in cell microscopy images. Furthermore, the denoising effect of these methods is highly dependent on the manually selected decomposition transform domain and threshold determination method, and they also have high computational costs and time consumption.

[0005] Current mainstream deep learning denoising methods include Gaussian white noise denoising, Gaussian blind denoising, real-world noisy image denoising, and super-blending noisy image denoising. The DnCNN network model improves image denoising performance by learning the mapping relationship between ideal and noisy images and incorporating batch normalization and residual learning between input and output images. The CBDNet network model uses real-world noisy clean image pairs to train the convolutional blind denoising network and embeds a noise estimation sub-network with asymmetric learning into the network model to suppress the underestimation of noise levels. While these two denoising network models perform adequately on synthetic noise or real-world macroscopic noise datasets, the superposition of various noises during microscopic imaging and the influence of sample surface characteristics limit their effectiveness in restoring noisy cell microscopic images. Although their performance surpasses traditional denoising methods such as Gaussian filtering, median filtering, and BM3D, their effectiveness in restoring microscopic images is limited. Summary of the Invention

[0006] The main objective of this invention is to overcome the aforementioned deficiencies in the prior art and propose an image denoising system and method based on multi-scale information extraction. This system can flexibly adjust the axial distance between the sample under test and the objective lens, thereby enabling focusing at different depths. Based on this, a multi-stage denoising network model is used to restore noisy cell images, which has significant effects on noise removal and texture information preservation.

[0007] The present invention adopts the following technical solution:

[0008] On one hand, an image denoising system based on multi-scale information extraction includes a microscopic image acquisition device and an image denoising device; the microscopic image acquisition device is used to acquire a microscopic image dataset for training and debugging, and also to acquire cell images of the sample to be processed; the image denoising device is used to train and debug a multi-stage denoising network model using the microscopic image dataset to obtain a multi-stage denoising network model with optimal parameters, and also to use the multi-stage denoising network model to process the cell images of the input sample to be processed and output a denoised image.

[0009] Preferably, the microscopic image acquisition device includes:

[0010] The system comprises a light source module, a transmission module, and an imaging module. In the direction of optical propagation, the light source module sequentially includes a light source, a beam expander, and a collimating lens group. The transmission module sequentially includes a spatial light modulator, a relay lens group, a total internal reflection prism, a semi-reflective mirror, an objective lens, and an XYZ stage. The imaging module sequentially includes a focusing lens, a detector, and a host computer. The spatial light modulator shapes and outputs the light beam incident from the collimating lens group, and the incident light beam is incident into the spatial light modulator at a specific angle. The beam containing the sample information is transmitted in reverse from the XYZ stage to the semi-reflective mirror, reflected by the semi-reflective mirror, and then reaches the focusing lens, which in turn reaches the detector to obtain coded information. The acquired cell microscopic image is then transmitted to the host computer.

[0011] Preferably, the spatial light modulator is a spatial light modulator with adjustable individual micro-elements.

[0012] Preferably, the imaging debugging method of the microscopic image acquisition device includes:

[0013] The sample is placed on the XYZ stage. The host computer controls all micromirrors of the spatial light modulator to be "on". During the coarse adjustment stage, the stage position is not moved, but the position of the objective lens relative to the XYZ stage is adjusted. During the fine adjustment stage, the objective lens is kept still, but the position of the XYZ stage is adjusted.

[0014] The host computer observes the position, shape, and brightness information of the sample under test, and controls the rotation angle of the individual micro-elements of the spatial light modulator to control the opening or closing of each individual micro-element.

[0015] Preferably, the image denoising device is deployed on a host computer and includes:

[0016] The training dataset construction module receives cell microscopic image data acquired by the microscopic image acquisition device and completes the construction of the dataset required for model training.

[0017] The model building and training module constructs a multi-stage denoising network model and uses data from the model training dataset to complete the construction of model parameters; the multi-stage denoising network model includes a channel association unit, a multi-scale denoising unit, and a fusion compression unit;

[0018] The model saving module observes the optimal number of training epochs for the network model by setting a loss function that gradually decreases with the number of training epochs; it also saves the parameter weights with the best performance by testing on the validation set; finally, the saved model is used to process the input image and output a denoised image.

[0019] The real-time processing module uses the saved model to process the cell images of the sample to be processed acquired by the microscopic image acquisition device and outputs the denoised image.

[0020] Preferably, the channel association unit uses a 10-layer structure to process the input image and concatenates the output result with the input image in the channel domain as the input to the multi-scale denoising unit; wherein the 10-layer network structure uses Conv+ReLU in the first layer, the last layer output uses only Conv, and the middle 8 layers use the same structure, namely Conv+BN+ReLU; in the convolution operation of the first and middle layers of the above channel association unit, the number of feature channels is set to 64, and the convolution kernel size is set to 3x3;

[0021] The multi-scale denoising unit adopts a five-layer parallel feature pyramid design structure to simultaneously capture target information at different scales. First, the graph output by the channel association unit is subjected to parallel pooling operations of 1x1, 2x2, 4x4, 8x8 and 16x16 to obtain feature representations at different scales. Then, these are input into their respective U-Nets for denoising. Finally, the multi-level denoising feature information is fused and concatenated in parallel with the graph output by the channel association unit in the channel domain as the input to the fusion compression unit.

[0022] The fusion compression unit employs four BottleNeck units. Each BottleNeck unit uses two convolutional layers with Batch Normalization (BN) operations, followed by ReLU layers and convolutional layer operations. In each convolutional layer of the BottleNeck unit, the growth rate of the number of feature channels is set to 32, and the kernel size is 3x3.

[0023] Preferably, the multi-stage denoising network model is represented as follows:

[0024]

[0025] Where I1 represents a cell microscopic noise image, i.e., the acquired image; I represents the denoised image restored by the multi-stage denoising network model; I0 is the noise-free cell micrograph corresponding to I1; F(·) represents the multi-stage denoising network model trained on the dataset so that the network model can learn the complex mapping transformation relationship F(·) between I1 and I0.

[0026]

[0027] Where f1, f2, and f3 represent the functional forms of the complex mapping transformation relationship F(·) between I1 and I0 in the first, second, and third stages, respectively. This means that after the input image I1 is processed through different stages f1, f2, and f3, a skip connection is made to achieve a channel-domain connection with the result of the previous stage; wherein, the first stage is the processing stage of the channel association unit, the second stage is the processing stage of the multi-scale denoising unit, and the third stage is the processing stage of the fusion compression unit.

[0028] The parameter adjustment method for the multi-stage denoising network model is as follows: Let

[0029]

[0030] Where, arg min f Loss (·) represents the value of f that minimizes the loss of the network model. Loss The values ​​of the variables in (·) This represents the values ​​of the weight parameters at each stage of the network model. The network model calculates its output using the loss function. The distance between the loss value and the true target value I0 is used to adjust the weight parameters of the network model using the loss value as a feedback signal through the optimizer.

[0031] Preferably, the construction of the dataset required for model training includes:

[0032] Cell microscopic image data is obtained by taking multiple photos of the same scene and averaging the images from these multiple photos.

[0033] Preferably, the image denoising system based on multi-scale information extraction further includes:

[0034] The data preprocessing module performs data augmentation, image scaling, image rotation, or random segmentation on the images in the model training pre-dataset.

[0035] On the other hand, an image denoising method based on multi-scale information extraction includes:

[0036] Training dataset construction: Receive cell microscopic image data acquired by the microscopic image acquisition device and complete the construction of the dataset required for model training;

[0037] Model building and training: A multi-stage denoising network model is constructed, and the model parameters are constructed using data from the model training dataset; the multi-stage denoising network model includes a channel association unit, a multi-scale denoising unit, and a fusion compression unit;

[0038] The model is saved by setting a loss function that gradually decreases with the number of training epochs to observe the optimal number of training epochs for the network model; the parameter weights with the best performance are saved by testing on the validation set; finally, the saved model is used to process the input image and output a denoised image.

[0039] Real-time processing: The saved model is used to process the cell images of the sample to be processed acquired by the microscopic image acquisition device and output the denoised image.

[0040] The channel association unit uses a 10-layer structure to process the input image and concatenates the output with the input image in the channel domain as the input to the multi-scale denoising unit. The 10-layer network structure uses Conv+ReLU in the first layer, Conv only in the last layer, and the middle 8 layers use the same structure, namely Conv+BN+ReLU. In the convolution operations of the first and middle layers of the channel association unit, the number of feature channels is set to 64, and the kernel size is set to 3x3.

[0041] The multi-scale denoising unit adopts a five-layer parallel feature pyramid design structure to simultaneously capture target information at different scales. First, the graph output by the channel association unit is subjected to parallel pooling operations of 1x1, 2x2, 4x4, 8x8 and 16x16 to obtain feature representations at different scales. Then, these are input into their respective U-Nets for denoising. Finally, the multi-level denoising feature information is fused and concatenated in parallel with the graph output by the channel association unit in the channel domain as the input to the fusion compression unit.

[0042] The fusion compression unit employs four BottleNeck units. Each BottleNeck unit uses two convolutional layers with Batch Normalization (BN) operations, followed by ReLU layers and convolutional layer operations. In each convolutional layer of the BottleNeck unit, the growth rate of the number of feature channels is set to 32, and the kernel size is 3x.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] (1) The present invention uses an XYZ stage and a single micro-element adjustable spatial light modulator during image acquisition, which can flexibly adjust the axial distance between the sample under test and the objective lens, thereby achieving focusing at different depths and obtaining an image that makes the sample under test achieve the best imaging conditions, and achieving certain noise reduction in hardware.

[0045] (2) When performing software denoising on the acquired cell images on the host computer, the present invention adopts a multi-stage denoising network model including a channel association module, a multi-scale denoising module and a fusion compression module, which has significant effects on noise removal and texture information preservation. Attached Figure Description

[0046] Figure 1 This is a structural block diagram of an image denoising system based on multi-scale information extraction according to an embodiment of the present invention;

[0047] Figure 2 This is a structural block diagram of the microscopic image acquisition device according to an embodiment of the present invention;

[0048] Figure 3 This is a structural block diagram of the image denoising device according to an embodiment of the present invention;

[0049] Figure 4 This is a network structure diagram of the multi-stage denoising network model according to an embodiment of the present invention;

[0050] Figure 5 This is a simplified flowchart of the image denoising method based on multi-scale information extraction according to an embodiment of the present invention;

[0051] Figure 6 This is a detailed flowchart of the image denoising method based on multi-scale information extraction according to an embodiment of the present invention;

[0052] Figure 7 The images show a comparison of the visual quality of the denoising process of the method in this embodiment of the invention and that of existing methods. In this comparison, (a) represents a denoising comparison image of a cell image and (b) represents a denoising comparison image of another cell image. Detailed Implementation

[0053] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0054] See Figure 1 As shown in the figure, this embodiment discloses an image denoising system based on multi-scale information extraction, including a microscopic image acquisition device 10 and an image denoising device 20. The microscopic image acquisition device 10 is used to acquire a microscopic image dataset for training and debugging, and also to acquire cell images of the sample to be processed. The image denoising device 20 is used to train and debug a multi-stage denoising network model using the microscopic image dataset to obtain a multi-stage denoising network model with optimal parameters, and also to use the multi-stage denoising network model to process the cell images of the input sample to be processed and output a denoised image.

[0055] See Figure 2 As shown, the microscopic image acquisition device 10 includes:

[0056] The system comprises a light source module, a transmission module, and an imaging module. In the direction of light propagation, the light source module sequentially includes a light source 101, a beam expander 102, and a collimating lens group 103. The transmission module sequentially includes a spatial light modulator 104, a relay lens group 105, a total internal reflection prism 106, a semi-reflective mirror 107, an objective lens 108, and an XYZ stage 109. The imaging module sequentially includes a focusing lens 110, a detector 111, and a host computer 112. The spatial light modulator 104 shapes and outputs the light beam incident from the collimating lens group 103, and the incident light beam is incident into the spatial light modulator 104 at a specific angle. The beam containing the sample information is transmitted in reverse from the XYZ stage 109 to the semi-reflective mirror 107, and after reflection by the semi-reflective mirror 107, it reaches the focusing lens 110, then the detector 111 to obtain the encoded information, and the acquired cell microscopic image is transmitted to the host computer 112.

[0057] Specifically, the light source module includes, in sequence according to the direction of light propagation: light source 101, beam expander 102, and collimating lens group 103; wherein the beam expander 102 is used to expand the beam emitted by the light source 101, and the collimating lens group 103 is used to convert the emitted beam into a parallel beam.

[0058] The transmission module, in accordance with the optical path propagation direction, sequentially includes: a spatial light modulator 104, a relay lens group 105, a total internal reflection prism 106, a semi-reflective mirror 107, an objective lens 108, and an XYZ stage 109. The spatial light modulator 104 shapes and exits the light beam incident from the collimating lens group 103, and the incident light beam enters the spatial light modulator 104 at a specific angle, then sequentially passes through the relay lens group 105, the total internal reflection prism 106, the semi-reflective mirror 107, and the objective lens 108 before reaching the stage. The relay lens group 105 extends the beam transmission distance; the total internal reflection prism 106 reflects the beam at a 90-degree angle to the semi-reflective mirror 107; the objective lens 108 increases the incident flux; and the XYZ stage 109 achieves precise axial focusing.

[0059] The imaging module, arranged according to the propagation direction of the light path, sequentially includes: a focusing lens 110, a detector 111, and a host computer 112. The beam of light carrying the sample information originates from the XYZ stage 109 and is transmitted in the reverse direction to the semi-reflective mirror 107. After reflection by the semi-reflective mirror 107, it reaches the focusing lens 110 and then the detector 111 to acquire the encoded information. The focusing lens 110 serves to homogenize the non-uniform illumination on the photosensitive surface of the detector 111. Finally, a multi-stage denoising network model of the image denoising device 20 deployed on the host computer 112 is used to process and display the acquired cell microscopic images.

[0060] The spatial light modulator 104 can be any single micro-element adjustable spatial light modulator 104, such as a DMD or a transmissive LCD; and the angle at which the light beam enters the spatial light modulator 104 is 24°.

[0061] Furthermore, the imaging adjustment method of the microscopic image acquisition device 10 includes:

[0062] (1) First, place the sample on the XYZ stage 109. The host computer 112 controls all micromirrors of the spatial light modulator 104 to be "on". Then, in the coarse adjustment stage, the stage position is not moved, but the position of the objective lens 108 relative to the XYZ stage 109 is adjusted. In the fine adjustment stage, the objective lens 108 is kept still, but the position of the XYZ stage 109 is adjusted so that the detector 111 can achieve the best image clarity.

[0063] (2) According to (1), the host computer 112 can observe the position information, shape information and brightness information of the sample under test, and control the rotation angle of the independent micro-element of the spatial light modulator 104 through the host computer 112 to control the opening or closing of the individual micro-element, so that the sample under test can achieve the best imaging conditions.

[0064] It should be noted that the optimal imaging conditions mentioned here refer to the conditions that allow for proper exposure of the sample across the entire field of view without affecting the operator's observation.

[0065] After processing according to the following imaging debugging method, the hardware debugging of the device has been completed and the optimal imaging conditions have been achieved. Then, the image denoising device 20 can be deployed on the host computer 112, and the multi-stage denoising network model can be trained and debugged in the cell microscopic image dataset. The model parameters that perform best on the validation set during training are saved (during training, the model parameters of the current training stage are used to test the data of the validation set, and the loss value and accuracy are remembered. The program is designed to compare the values ​​of the current training stage with those of the previous training stage and retain the model parameters of the stage with better values). Then, the model is used to perform denoising processing on the test sample image to be processed.

[0066] Specifically, the image denoising device 20 includes:

[0067] The training dataset construction module 201 receives cell microscopic image data acquired by the microscopic image acquisition device 10 and completes the construction of the dataset required for model training.

[0068] Data preprocessing module 202 performs data augmentation, image scaling, image rotation, or random segmentation on the images in the model training pre-dataset;

[0069] The model building and training module 203 constructs a multi-stage denoising network model and uses data from the model training dataset to complete the construction of model parameters; the multi-stage denoising network model includes a channel association unit, a multi-scale denoising unit, and a fusion compression unit.

[0070] The model saving module 204 observes the optimal number of training epochs for the network model by setting a loss function that gradually decreases with the number of training epochs; it also saves the parameter weights with the best performance by testing on the validation set; finally, the saved model is used to process the input image and output a denoised image.

[0071] The real-time processing module 205 uses the saved model to process the cell images of the sample to be processed acquired by the microscopic image acquisition device 10 and outputs the denoised image.

[0072] Specifically, the construction of the dataset required for model training includes:

[0073] Cell microscopic image data is obtained by taking multiple photos of the same scene and averaging the images from these multiple photos.

[0074] See Figure 4 As shown, the channel association unit uses a 10-layer structure to process the input image and concatenates the output result with the input image in the channel domain as the input to the multi-scale denoising unit; the 10-layer network structure uses Conv+ReLU in the first layer, the last layer output uses only Conv, and the middle 8 layers use the same structure, namely Conv+BN+ReLU; in the convolution operation of the first and middle layers of the above channel association unit, the number of feature channels is set to 64, and the convolution kernel size is set to 3x3;

[0075] The multi-scale denoising unit adopts a five-layer parallel feature pyramid design structure to simultaneously capture target information at different scales. First, the graph output by the channel association unit is subjected to parallel pooling operations of 1x1, 2x2, 4x4, 8x8 and 16x16 to obtain feature representations at different scales. Then, these are input into their respective U-Nets for denoising. Finally, the multi-level denoising feature information is fused and concatenated in parallel with the graph output by the channel association unit in the channel domain as the input to the fusion compression unit.

[0076] The fusion compression unit employs four BottleNeck units. Each BottleNeck unit uses two convolutional layers with Batch Normalization (BN) operations, followed by ReLU layers and convolutional layer operations. In each convolutional layer of the BottleNeck unit, the growth rate of the number of feature channels is set to 32, and the kernel size is 3x3.

[0077] Furthermore, the multi-stage denoising network model is represented as follows:

[0078]

[0079] Where I1 represents a cell microscopic noise image, i.e., the acquired image; I represents the denoised image restored by the multi-stage denoising network model; I0 is the noise-free cell micrograph corresponding to I1; F(·) represents the multi-stage denoising network model trained on the dataset so that the network model can learn the complex mapping transformation relationship F(·) between I1 and I0.

[0080]

[0081] Where f1, f2, and f3 represent the functional forms of the complex mapping transformation relationship F(·) between I1 and I0 in the first, second, and third stages, respectively. This means that after the input image I1 is processed through different stages f1, f2, and f3, a skip connection is made to achieve a channel-domain connection with the result of the previous stage; wherein, the first stage is the processing stage of the channel association unit, the second stage is the processing stage of the multi-scale denoising unit, and the third stage is the processing stage of the fusion compression unit.

[0082] The parameter adjustment method for the multi-stage denoising network model is as follows: Let

[0083]

[0084] Where, argminf Loss (·) represents the value of f that minimizes the loss of the network model. Loss The values ​​of the variables in (·) This represents the values ​​of the weight parameters at each stage of the network model. The network model calculates its output using the loss function. The distance between the loss value and the true target value I0 is used to adjust the weight parameters of the network model using the loss value as a feedback signal through the optimizer.

[0085] See Figure 5 and Figure 6 As shown, according to another aspect of the present invention, this embodiment of an image denoising method based on multi-scale information extraction includes:

[0086] S501, Training Dataset Construction: Receives cell microscopic image data acquired by the microscopic image acquisition device and completes the construction of the dataset required for model training.

[0087] S502, Data preprocessing, performs data augmentation, image scaling, image rotation, or random segmentation on the images in the model training pre-dataset;

[0088] S503, Model Building and Training: Construct a multi-stage denoising network model and use data from the model training dataset to build the model parameters; the multi-stage denoising network model includes a channel association unit, a multi-scale denoising unit, and a fusion compression unit.

[0089] S504, Model Saving: By setting a loss function that gradually decreases with the number of training epochs, the optimal number of training epochs for the network model is observed; the best-performing parameter weights are saved through testing on the validation set; finally, the saved model is used to process the input image and output a denoised image.

[0090] S505, real-time processing, uses the saved model to process the cell images of the sample to be processed acquired by the microscopic image acquisition device and outputs the denoised image.

[0091] The channel association unit uses a 10-layer structure to process the input image and concatenates the output with the input image in the channel domain as the input to the multi-scale denoising unit. The 10-layer network structure uses Conv+ReLU in the first layer, Conv only in the last layer, and the middle 8 layers use the same structure, namely Conv+BN+ReLU. In the convolution operations of the first and middle layers of the channel association unit, the number of feature channels is set to 64, and the kernel size is set to 3x3.

[0092] The multi-scale denoising unit adopts a five-layer parallel feature pyramid design structure to simultaneously capture target information at different scales. First, the graph output by the channel association unit is subjected to parallel pooling operations of 1x1, 2x2, 4x4, 8x8 and 16x16 to obtain feature representations at different scales. Then, these are input into their respective U-Nets for denoising. Finally, the multi-level denoising feature information is fused and concatenated in parallel with the graph output by the channel association unit in the channel domain as the input to the fusion compression unit.

[0093] The fusion compression unit employs four BottleNeck units. Each BottleNeck unit uses two convolutional layers with Batch Normalization (BN) operations, followed by ReLU layers and convolutional layer operations. In each convolutional layer of the BottleNeck unit, the growth rate of the number of feature channels is set to 32, and the kernel size is 3x.

[0094] See Figure 7The figure shows a comparison of the visual quality of the denoising process after processing by the method of the present invention and existing methods. As can be seen from the figure, the image denoising method based on multi-scale information extraction of the present invention performs better in noise removal and texture information preservation for the restoration of noisy cell microscopic images compared with other methods (Gaussian filtering, BM3D, DnCNN, etc.).

[0095] Other implementations of the image denoising method based on multi-scale information extraction in this embodiment will not be repeated in this embodiment.

[0096] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. An image denoising system based on multi-scale information extraction, characterized in that, The system includes a microscopic image acquisition device and an image denoising device. The microscopic image acquisition device is used to acquire a microscopic image dataset for training and debugging, and also to acquire cell images of the sample to be processed. The image denoising device is used to train and debug a multi-stage denoising network model using the microscopic image dataset to obtain a multi-stage denoising network model with optimal parameters, and also to use the multi-stage denoising network model to process the cell images of the input sample to be processed and output a denoised image. The image denoising device is deployed on a host computer and includes: The training dataset construction module receives cell microscopic image data acquired by the microscopic image acquisition device and completes the construction of the dataset required for model training. The model building and training module constructs a multi-stage denoising network model and uses data from the model training dataset to complete the construction of model parameters; the multi-stage denoising network model includes a channel association unit, a multi-scale denoising unit, and a fusion compression unit; The model saving module observes the optimal number of training epochs for the network model by setting a loss function that gradually decreases with the number of training epochs; it also saves the parameter weights with the best performance by testing on the validation set; finally, the saved model is used to process the input image and output a denoised image. The real-time processing module uses the saved model to process the cell images of the sample to be processed acquired by the microscopic image acquisition device and outputs the denoised image. The channel association unit uses a 10-layer structure to process the input image and concatenates the output with the input image in the channel domain as the input to the multi-scale denoising unit. The 10-layer network structure uses Conv+ReLU in the first layer, Conv only in the last layer, and the middle 8 layers use the same structure, namely Conv+BN+ReLU. In the convolution operations of the first and middle layers of the channel association unit, the number of feature channels is set to 64, and the kernel size is set to 3x3. The multi-scale denoising unit adopts a five-layer parallel feature pyramid design structure to simultaneously capture target information at different scales. First, the graph output by the channel association unit is subjected to parallel pooling operations of 1x1, 2x2, 4x4, 8x8 and 16x16 to obtain feature representations at different scales. Then, these are input into their respective U-Nets for denoising. Finally, the multi-level denoising feature information is fused and concatenated in parallel with the graph output by the channel association unit in the channel domain as the input to the fusion compression unit. The fusion compression unit employs four BottleNeck units. Each BottleNeck unit uses two convolutional layers with Batch Normalization (BN) operations, followed by ReLU layers and convolutional layer operations. In each convolutional layer of the BottleNeck unit, the growth rate of the number of feature channels is set to 32, and the kernel size is 3x3.

2. The image denoising system based on multi-scale information extraction according to claim 1, characterized in that, The microscopic image acquisition device includes: The system comprises a light source module, a transmission module, and an imaging module. In the direction of optical propagation, the light source module sequentially includes a light source, a beam expander, and a collimating lens group. The transmission module sequentially includes a spatial light modulator, a relay lens group, a total internal reflection prism, a semi-reflective mirror, an objective lens, and an XYZ stage. The imaging module sequentially includes a focusing lens, a detector, and a host computer. The spatial light modulator shapes and outputs the light beam incident from the collimating lens group, and the incident light beam is incident into the spatial light modulator at a specific angle. The beam containing the sample information is transmitted in reverse from the XYZ stage to the semi-reflective mirror, reflected by the semi-reflective mirror, and then reaches the focusing lens, which in turn reaches the detector to obtain coded information. The acquired cell microscopic image is then transmitted to the host computer.

3. The image denoising system based on multi-scale information extraction according to claim 2, characterized in that, The spatial light modulator is a spatial light modulator with adjustable individual micro-elements.

4. The image denoising system based on multi-scale information extraction according to claim 2, characterized in that, The imaging debugging method of the microscopic image acquisition device includes: The sample is placed on the XYZ stage. The host computer controls all micromirrors of the spatial light modulator to be "on". During the coarse adjustment stage, the stage position is not moved, but the position of the objective lens relative to the XYZ stage is adjusted. During the fine adjustment stage, the objective lens is kept still, but the position of the XYZ stage is adjusted. The host computer observes the position, shape, and brightness information of the sample under test, and controls the rotation angle of the individual micro-elements of the spatial light modulator to control the opening or closing of each individual micro-element.

5. The image denoising system based on multi-scale information extraction according to claim 1, characterized in that, The multi-stage denoising network model is represented as follows: ； in, This represents a cell microscopic image with noise, i.e., the acquired image; This represents the denoised image restored after being processed by a multi-stage denoising network model. Is with Corresponding noise-free cell micrographs; This indicates that the multi-stage denoising network model is trained on the dataset so that the network model can learn... arrive Complex mapping and transformation relationships between them ; ； in, They represent arrive Complex mapping and transformation relationships between them The functional forms in the first, second, and third stages, Indicates the input image go through After processing at different stages, a skip connection is made to achieve a channel domain connection with the result of the previous stage; wherein, the first stage is the channel association unit processing stage, the second stage is the multi-scale denoising unit processing stage, and the third stage is the fusion compression unit processing stage. The parameter adjustment method for the multi-stage denoising network model is as follows: Let ； in, This indicates the time when the network model loss value reaches its minimum. The values ​​that the variables in the text can take. This represents the values ​​of the weight parameters at each stage of the network model. The network model calculates its output using the loss function. Compared with the actual target value The distance between them is then used to determine the weight parameters of the network model, and the optimizer uses the loss value as a feedback signal.

6. The image denoising system based on multi-scale information extraction according to claim 1, characterized in that, The construction of the dataset required to complete model training includes: Cell microscopic image data is obtained by taking multiple photos of the same scene and averaging the images from these multiple photos.

7. The image denoising system based on multi-scale information extraction according to claim 1, characterized in that, Also includes: The data preprocessing module performs data augmentation, image scaling, image rotation, or random segmentation on the images in the model training pre-dataset.

8. An image denoising method based on multi-scale information extraction, characterized in that, Based on the system according to any one of claims 1 to 7, the method includes: Training dataset construction: Receive cell microscopic image data acquired by the microscopic image acquisition device and complete the construction of the dataset required for model training; Model building and training: A multi-stage denoising network model is constructed, and the model parameters are constructed using data from the model training dataset; the multi-stage denoising network model includes a channel association unit, a multi-scale denoising unit, and a fusion compression unit; The model is saved by setting a loss function that gradually decreases with the number of training epochs to observe the optimal number of training epochs for the network model; the parameter weights with the best performance are saved by testing on the validation set; finally, the saved model is used to process the input image and output a denoised image. Real-time processing: The saved model is used to process the cell images of the sample to be processed acquired by the microscopic image acquisition device and output the denoised image. The channel association unit uses a 10-layer structure to process the input image and concatenates the output with the input image in the channel domain as the input to the multi-scale denoising unit. The 10-layer network structure uses Conv+ReLU in the first layer, Conv only in the last layer, and the middle 8 layers use the same structure, namely Conv+BN+ReLU. In the convolution operations of the first and middle layers of the channel association unit, the number of feature channels is set to 64, and the kernel size is set to 3x3. The multi-scale denoising unit adopts a five-layer parallel feature pyramid design structure to simultaneously capture target information at different scales. First, the graph output by the channel association unit is subjected to parallel pooling operations of 1x1, 2x2, 4x4, 8x8 and 16x16 to obtain feature representations at different scales. Then, these are input into their respective U-Nets for denoising. Finally, the multi-level denoising feature information is fused and concatenated in parallel with the graph output by the channel association unit in the channel domain as the input to the fusion compression unit. The fusion compression unit employs four BottleNeck units. Each BottleNeck unit uses two convolutional layers with Batch Normalization (BN) operations, followed by ReLU layers and convolutional layer operations. In each convolutional layer of the BottleNeck unit, the growth rate of the number of feature channels is set to 32, and the kernel size is 3x3.

Images