Microorganism experiment bacterial colony image processing method

Abstract

The invention provides a microbial experiment bacterial colony image processing method, and belongs to the field of image enhancement. Through combination of a student enhancement network and a mean value teacher auxiliary learning network, limited labeled data and a large amount of unlabeled data are fully utilized; the student enhancement network extracts image features and generates enhanced potential variables, and the mean value teacher network generates pseudo tags through nonlinear transformation and adaptive regularization to guide learning; a Gaussian process regression loss module is introduced, a learning process is optimized by combining weighted errors and covariance matrix uncertainty measurement, and overfitting is avoided; the optimization potential variable selection module screens a high-quality potential variable set based on the peak signal-to-noise ratio and diversity score; the multi-loss function module combines the L1 loss, the SSIM loss and the perception loss to carry out weighted optimization so as to further improve the image enhancement effect; the method is superior to the prior art in the aspects of precision, robustness and applicability, and is suitable for high-quality image processing requirements in microbiological experiments.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a method for processing microbial experimental colony images. Background Art

[0002] In microbial experiments, the clarity and feature visibility of colony images are crucial to the accuracy and reliability of experimental results; traditional colony image processing methods mainly rely on simple image enhancement techniques, such as contrast adjustment and filtering, but these methods often fail to meet the needs of complex microbial images; microbial colony images usually have the characteristics of low contrast, noise interference, and blurred features, which makes accurate identification and analysis of colonies extremely challenging; in addition, with the deepening of microbial research, higher requirements are placed on the accuracy and efficiency of image processing; therefore, developing an advanced method that can effectively enhance the features of microbial experimental colony images and improve image clarity is of great significance to promoting the development of microbial experimental research.

[0003] At present, the processing of microbial experimental colony images mainly faces two major technical bottlenecks. On the one hand, traditional image enhancement algorithms do not work well when dealing with complex backgrounds and noise interference, and it is difficult to effectively extract colony features. On the other hand, although existing deep learning methods have made significant progress in the field of image processing, they still have shortcomings in microbial colony image processing. For example, deep learning models usually require a large amount of labeled data for training, while obtaining a large number of labeled images in microbial experiments is time-consuming and laborious. In addition, existing models are prone to overfitting problems when processing unlabeled data, resulting in insufficient generalization ability. Therefore, how to make full use of information in limited labeled data and a large amount of unlabeled data to improve the performance of the model is an urgent problem to be solved in the current microbial experimental colony image processing.

[0004] In recent years, deep learning technology has made breakthrough progress in the field of image processing, providing new ideas for the processing of microbial experimental colony images; however, existing deep learning methods still have some limitations when processing microbial images; for example, most methods only rely on a single network structure and cannot fully utilize the advantages of different network modules; in addition, when dealing with image enhancement tasks, existing methods often ignore the diversity and uncertainty of image features, resulting in the enhanced images still failing to meet experimental requirements in some cases; therefore, an innovative image processing method is needed that can make full use of labeled and unlabeled data through multi-module collaborative optimization, while considering the diversity and uncertainty of image features, so as to achieve more efficient and accurate microbial experimental colony image enhancement. Summary of the invention

[0005] The present invention provides a microbial experiment colony image processing method, which aims to significantly improve the clarity and feature visibility of microbial experiment colony images by combining a student enhancement network and a mean teacher assisted learning network, utilizing pseudo label generation, Gaussian process regression loss and multi-loss function optimization, and is suitable for high-quality image processing needs in microbial experiments.

[0006] The present invention provides a method for processing microbial experimental colony images, comprising the following steps: S1. Preparation of microbial experimental colony image dataset: collect microbial experimental colony images, 30% of which are labeled as labeled samples in the training dataset, and 70% are not labeled as unlabeled samples in the training dataset; S2, design a student enhancement network module to extract features from the input microbial experimental colony image and generate an enhanced image as the student network latent variable; S3. Design a mean teacher-assisted learning network module, map the input feature information to a high-dimensional feature space by improving the nonlinear transformation strategy, and introduce an adaptive regularization coefficient optimization model to generate pseudo labels and pseudo teacher network latent variables; S4. Design a mean teacher-assisted Gaussian process regression loss module, design the weighted error between the pseudo latent variable and the actual latent variable, combine the uncertainty measure of the covariance matrix, and use the Gaussian process regression loss function to guide the learning process; S5. Design an optimized latent variable selection OLVS module to select the best quality and most diverse latent variable set from the latent variables generated by the mean teacher-assisted learning network and the student-enhanced network by optimizing the peak signal-to-noise ratio and diversity score of the latent variables. S6. Design a loss function module, optimize and design different loss methods, and introduce weighted coefficients and scale factors to perform weighted optimization according to different scales and microbial experimental colony image characteristics; S7. Construct the final enhancement network to infer and enhance the image to be processed; integrate the above modules into a complete network, train the network by optimizing the loss function, and generate an image enhancement model; and use the trained model to process the input microbial experimental colony image and output the enhanced image to improve the image clarity and visibility of the microbial experimental colony features.

[0007] Preferably, the student enhancement network module in S2 is responsible for extracting key features from the input microbial experimental colony image and generating an enhanced image as a latent variable of the student network; the student enhancement network can effectively improve the clarity and feature visibility of the image by learning the feature expression of the image, and provide a high-quality image foundation for subsequent image processing and analysis; the design of this module enables the model to focus on the extraction and enhancement of image features, laying a solid foundation for further optimizing the image processing effect.

[0008] Preferably, the mean teacher-assisted learning network module in S3 is specifically implemented as follows: S31. The feature vector U extracted from the microbial experimental colony image is mapped to a high-dimensional feature space through nonlinear transformation. The calculation formula is as follows: ; in, is the feature vector of the input microbial experimental colony image, is the input feature vector The representation of the mapping in the high-dimensional feature space, To calculate the weight matrix of the linear transformation passed from the kth layer to the next layer, is the PReLu activation function, To calculate the weight matrix between the kth layer and each layer, To adjust the bias term of the k-th layer output information, It is the network layer information.

[0009] Preferably, S32, using weighted basis function expansion to model the output, and introducing a noise term to simulate the random disturbance of data during the actual microbial experiment colony image enhancement process, the specific formula is: ; in, is the output value, which represents the feature vector of the colony image according to the input microbial experiment and the transformed features The calculated forecast results; is the weight coefficient of the basis function, indicating the contribution intensity of each basis function to the final output; is the frequency parameter of each basis function, which is used to control the oscillation speed of each basis function, reflecting its sensitivity to different input features; The offset of each basis function controls the phase of the basis function; It is the noise term that simulates the random disturbance of data in the process of microbial colony image enhancement; is the number of design basis functions.

[0010] Preferably, S33, the prediction process is optimized by weighted regression and gradient regularization, and the complexity of the model is controlled to ensure its good generalization ability. The specific formula is: ; in, is the training loss function, which represents the goal of model optimization; is the weight of each sample, indicating the contribution of the data point in the loss function during the microbial experimental colony image enhancement process; is the data fitting error term, indicating the predicted value and the true label The error between is the gradient regularization term; is the adaptive regularization weight, the specific formula is: ; in, To initialize the regularization weights, To use the predicted value The variance of is used to reflect the uncertainty of the sample; Adjust the parameters to control the amplification effect of variance on the regularization coefficient; is the L2 norm of the current gradient, is the correlation between the target value and the predicted value, Factors that adjust the weights of correlation effects; To prevent extremely small numbers with a denominator of zero.

[0011] Preferably, the mean teacher assisted learning network module in S3 maps the input feature information to a high-dimensional feature space by improving the nonlinear transformation strategy, and introduces an adaptive regularization coefficient optimization model to generate pseudo labels and pseudo teacher network latent variables; the module uses nonlinear transformation and weighted basis function expansion to model the output, and introduces noise terms to simulate the random perturbation of data in the actual image enhancement process; the prediction process is optimized by weighted regression and gradient regularization to control the model complexity and ensure good generalization ability; the design of this module can effectively utilize unlabeled data and generate reliable pseudo labels, thereby improving the model's learning ability and overall performance for unlabeled data.

[0012] Preferably, the mean teacher assisted Gaussian process regression loss module in S4 is specifically implemented as follows: S41. Calculate the weighted error between the pseudo latent variable and the actual latent variable in the process of microbial experiment colony image enhancement. The specific formula is: ; in, is a pseudo latent variable, which is an estimated value of the latent variable generated by the mean teacher-assisted learning network module in step S2; is the mean of the latent variable in the Gaussian process model; is the covariance matrix, which represents the dependency between latent variables, and its inverse matrix represents the weighted error; To measure the pseudo latent variable With the Gaussian process model mean The weighted difference between .

[0013] Preferably, S42, by taking the logarithm of the determinant of the covariance matrix, the model can control the prediction uncertainty of the latent variables and avoid overfitting. The specific formula is: ; in, is the covariance matrix The logarithm of the determinant of , which represents a measure of model uncertainty; is a regularization hyperparameter used to control the weight of the regularization term; is the trace of the inverse of the covariance matrix and the covariance matrix of the pseudo latent variable; Calculate the logarithm of the covariance term to control the uncertainty of the model.

[0014] Preferably, S43, combining the above two results, optimizing the potential variable prediction of the microbial experiment colony image enhancement result by maximizing the posterior distribution, the specific formula is: ; in, and are the hyperparameters for adjusting the error and prior knowledge terms.

[0015] Preferably, the mean teacher-assisted Gaussian process regression loss module in S4 calculates the weighted error between the pseudo latent variable and the actual latent variable, and combines the uncertainty measure of the covariance matrix to guide the learning process using the Gaussian process regression loss function; this module controls the prediction uncertainty of the latent variable by taking the logarithm of the determinant of the covariance matrix to avoid the occurrence of overfitting; and optimizes the prediction of the latent variable by maximizing the posterior distribution, thereby further improving the model's ability to handle complex image data and the accuracy of the image enhancement effect.

[0016] Preferably, the optimized latent variable selection OLVS module in S5 selects a latent variable set with the best quality and greater diversity from the latent variables generated by the teacher network and the student network, and the specific method is: S51. The PSNR score is obtained by calculating the peak signal-to-noise ratio to measure the difference between each potential variable and the real microbial experimental colony image. The specific formula is: ; in, is the PSNR score of the ith latent variable, is a real image, is the ith latent variable, is the maximum pixel value of the true image used to normalize the PSNR score, and is the mean square error between the true image and the i-th latent variable.

[0017] Preferably, S52, using a selection function to select the optimal latent variable according to the PSNR score, the specific formula is: ; ; in, is the optimal latent variable set, including the selected N latent variables; i represents the i-th latent variable; is the real image label; Latent variables generated for the teacher network; Latent variables generated for student networks; is the weight of the PSNR score; the weighting given to scoring diversity; is the penalty coefficient of the error term; Differences between latent variables generated to measure teacher and student networks; To score diversity, we measure the degree of difference between the current latent variable and other latent variables; is the maximum independent variable point set function.

[0018] Preferably, the optimized latent variable selection (OLVS) module in S5 selects a set of latent variables with the best quality and greater diversity from the latent variables generated by the mean teacher-assisted learning network and the student enhancement network by calculating the peak signal-to-noise ratio and diversity score; the module uses the PSNR score to measure the difference between the latent variables and the real image, and combines the diversity score to ensure that the selected latent variables have high quality and diversity; in this way, the model can screen out more representative and effective latent variables, further improving the accuracy and robustness of the image enhancement effect.

[0019] Preferably, the loss function module in S6 uses multiple loss functions, including L1 loss, SSIM loss and perceptual loss, and the specific method is: S61. Calculate the unsupervised loss to process the unlabeled microbial experimental colony image data, and calculate the L1 loss by combining the pseudo labels generated by the student network and the teacher network to ensure that the model learns effective information from the pseudo labels. The specific formula is: ; in, Generate latent variables for the student network at the jth scale; Latent variables generated for the teacher network; is the weighting coefficient used to adjust the contribution of unsupervised loss at different scales.

[0020] Preferably, S62, calculate L1 loss , SSIM loss and perceived loss , the specific formula is: ; in, is the pixel value of the real image at the jth scale, is the predicted value of the student network at the jth scale; is the total number of scales; is the weight coefficient for each scale; ; in, is the structural similarity between the real image and the image generated by the student network at the jth scale; is the weight coefficient for each scale; ; at last, Features extracted by the jth layer of the pre-trained VGG network; is the weight coefficient of the j-th scale feature; It is used to measure the L2 norm of feature differences.

[0021] Preferably, S63, L1 loss, SSIM loss and perceptual loss are combined to form a weighted combined loss function to optimize the learning objective of the model. The specific formula is: ; in, and It is a hyperparameter that controls the impact of different loss terms on the final total loss.

[0022] Preferably, the loss function module in S6 combines L1 loss, SSIM loss and perceptual loss, and introduces weighting coefficients and scale factors to perform weighted optimization according to different scales and microbial experimental colony image features; the module processes unlabeled image data through unsupervised loss, and calculates L1 loss in combination with pseudo-labels generated by student network and teacher network to ensure that the model learns effective information from pseudo-labels; by combining multiple loss methods to form a weighted combination loss function, the learning objectives of the model are further optimized, and the overall performance of the image enhancement effect is significantly improved.

[0023] Preferably, the final enhanced network is constructed in S7 to infer and enhance the image to be processed; the input data of this step is the original microbial experimental colony image; all the modules of the above steps are integrated into a complete network, and the network is trained by optimizing the loss function; the trained model is used to process the input microbial experimental colony image to generate an output enhanced microbial experimental colony image with higher clarity and feature visibility.

[0024] Compared with the existing technology, this technology combines the student enhancement network and the mean teacher assisted learning network, makes full use of labeled and unlabeled data, realizes the combination of supervised and unsupervised learning, and improves the generalization ability of the model; at the same time, introduces the Gaussian process regression loss module and the optimized latent variable selection module to accurately control the prediction uncertainty, screen out a high-quality and diverse set of latent variables, and significantly improves the image clarity and feature visibility; in addition, through the weighted optimization of multiple loss function modules (including L1 loss, SSIM loss and perceptual loss), the image enhancement effect is further improved; this technology is superior to the existing technology in terms of accuracy, robustness and applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a method for processing images of microbial experimental colonies.

[0026] Figure 2 Schematic diagram of the process to obtain the optimal latent variables.

[0027] Figure 3 The following is a schematic diagram of the final enhanced network flow.

[0028] Figure 4 The enhanced output contrast image for the final network. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] Please see attached Figure 1 -Attached Figure 4 , the present invention provides a method for processing microbial experimental colony images.

[0031] The present invention proposes a method for processing microbial colony images, as shown in the attached Figure 1 As shown in the overall flow chart, it specifically includes the following steps: Step S1, microbial experiment colony image data set preparation, use Python's Pillow library to load microbial experiment colony images, and use the os module to traverse the image folder to obtain the image path list; use random.shuffle() to randomly shuffle the image path list, and then select labeled samples at a ratio of 30%, and the remaining 70% as unlabeled samples; the labeled samples are labeled by manual labeling tools (such as LabelImg), and the labeling results are saved in JSON format; the unlabeled samples directly store the path information for the subsequent semi-supervised learning process; the labeled samples and the unlabeled samples are stored in different folders for easy calling during the training process.

[0032] Step S2, use Python's torchvision library to build a student enhancement network. The network structure adopts a convolutional neural network, and specifically selects the UNet architecture; the network input is a microbial experimental colony image, and the output is an enhanced image; torch.nn.Conv2d is used to define the convolution layer in the network, and the convolution kernel size is 3×3, the step size is 1, and the padding is 1 to maintain the image size; the activation function uses torch.nn.PReLU to introduce nonlinear characteristics and accelerate the training process; the encoder part of the network contains 4 downsampling modules, each module consists of two convolution layers and a maximum pooling layer; the decoder part contains 4 upsampling modules, each module consists of a transposed convolution layer and two convolution layers, which are used to gradually restore the image resolution; the final output enhanced image has the same resolution as the input image.

[0033] Step S3: Design attachment Figure 2The mean teacher-assisted learning network module in the paper uses Python's torch.nn module to implement the mean teacher-assisted learning network. The network maps features to a high-dimensional space through a custom nonlinear transformation. The MLP contains two fully connected layers. The first layer maps the input feature dimension 512 to 1024, and the second layer maps 1024 back to 512. The activation function uses torch.nn.PReLU to introduce nonlinear characteristics. The network introduces an adaptive regularization coefficient to reflect the uncertainty of the sample by calculating the variance of the predicted value, and dynamically adjusts the regularization weight by combining the L2 norm of the gradient and the correlation between the target value and the predicted value.

[0034] Further, S31, the feature vector U extracted from the microbial experimental colony image is mapped to a high-dimensional feature space through nonlinear transformation, and the calculation formula is as follows: ; in, is the feature vector of the input microbial experimental colony image, is the input feature vector The representation of the mapping in the high-dimensional feature space, To calculate the weight matrix of the linear transformation passed from the kth layer to the next layer, is the PReLu activation function, To calculate the weight matrix between the kth layer and each layer, To adjust the bias term of the k-th layer output information, It is the network layer information.

[0035] Further, S32, the output is modeled using weighted basis function expansion, and a noise term is introduced to simulate the random disturbance of data in the process of colony image enhancement in actual microbial experiments. The specific formula is: ; in, is the output value, which represents the feature vector of the colony image according to the input microbial experiment and the transformed features The calculated forecast results; is the weight coefficient of the basis function, indicating the contribution intensity of each basis function to the final output; is the frequency parameter of each basis function, which is used to control the oscillation speed of each basis function, reflecting its sensitivity to different input features; The offset of each basis function controls the phase of the basis function; It is the noise term that simulates the random disturbance of data in the process of microbial colony image enhancement; is the number of design basis functions.

[0036] Furthermore, S33 optimizes the prediction process through weighted regression and gradient regularization, and controls the complexity of the model to ensure its good generalization ability. The specific formula is: ; in, is the training loss function, which represents the goal of model optimization; is the weight of each sample, indicating the contribution of the data point in the loss function during the microbial experimental colony image enhancement process; is the data fitting error term, indicating the predicted value and the true label The error between is the gradient regularization term; is the adaptive regularization weight, the specific formula is: ; in, To initialize the regularization weight, here it is equal to 0.5, To use the predicted value The variance of is used to reflect the uncertainty of the sample; To control the amplification effect of variance on the regularization coefficient, the adjustment parameter is equal to 0.75 here; is the L2 norm of the current gradient, is the correlation between the target value and the predicted value, is the factor that adjusts the weight of the correlation, which is equal to 0.42 here; To prevent the denominator from being zero, here it is 10 -6 That is 0.000001.

[0037] Step S4: Design attachment Figure 2 The mean teacher-assisted Gaussian process regression loss module in the paper uses Python's torch library to implement the Gaussian process regression loss module; the module guides the learning process by calculating the weighted error between the pseudo latent variables and the actual latent variables and combining the uncertainty measure of the covariance matrix.

[0038] Further, S41, the weighted error between the pseudo latent variable and the actual latent variable in the process of microbial experiment colony image enhancement is calculated, and the specific formula is: ; in, is a pseudo latent variable, which is an estimated value of the latent variable generated by the mean teacher-assisted learning network module in step S2; is the mean of the latent variable in the Gaussian process model; is the covariance matrix, which represents the dependency between latent variables, and its inverse matrix represents the weighted error; To measure the pseudo latent variable With the Gaussian process model mean The weighted difference between .

[0039] Furthermore, S42, by taking the logarithm of the determinant of the covariance matrix, the model can control the prediction uncertainty of the latent variables and avoid overfitting. The specific formula is: ; in, is the covariance matrix The logarithm of the determinant of , which represents a measure of model uncertainty; is the regularization hyperparameter, which is equal to 0.65 here and is used to control the weight of the regularization term; is the trace of the inverse of the covariance matrix and the covariance matrix of the pseudo latent variable; Calculate the logarithm of the covariance term to control the uncertainty of the model.

[0040] Further, S43, combining the above two results, the latent variable prediction of the microbial experiment colony image enhancement result is optimized by maximizing the posterior distribution. The specific formula is: ; in, and To adjust the hyperparameters of the error and prior knowledge terms, here is equal to 0.4, Equal to 0.6.

[0041] Step S5, design an optimized latent variable selection OLVS module, and use Python's torch library to implement the optimized latent variable selection module; the module selects the optimal set of latent variables from the generated latent variables by calculating the peak signal-to-noise ratio and diversity score; the PSNR score is obtained by calculating the mean square error between the real image and the latent variable; the diversity score is obtained by calculating the Euclidean distance between the latent variables, which is used to measure the difference between the variables; finally, the latent variable set with the best quality and strong diversity is screened out from the latent variables generated by the teacher network and the student network through the selection function.

[0042] Furthermore, S51, the PSNR score is obtained by calculating the peak signal-to-noise ratio to measure the difference between each latent variable and the real microbial experimental colony image. The specific formula is: ; in, is the PSNR score of the ith latent variable, is a real image, is the ith latent variable, is the maximum pixel value of the true image used to normalize the PSNR score, and is the mean square error between the true image and the i-th latent variable.

[0043] Further, S52, a selection function is used to select the optimal latent variable according to the PSNR score. The specific formula is: ; ; in, is the optimal latent variable set, including the selected N latent variables; i represents the i-th latent variable; is the real image label; Latent variables generated for the teacher network; Latent variables generated for student networks; is the weight of the PSNR score; the weighting given to scoring diversity; is the penalty coefficient of the error term; Differences between latent variables generated to measure teacher and student networks; To score diversity, we measure the degree of difference between the current latent variable and other latent variables; is the maximum independent variable point set function.

[0044] Step S6, design a loss function module, use Python's torch library to implement a multi-loss function module, and combine L1 loss, SSIM loss and perceptual loss; L1 loss is obtained by calculating the pixel-level difference between the real image and the enhanced image, and is implemented using torch.nn.L1Loss; SSIM loss is obtained by calculating the structural similarity of the image.

[0045] Furthermore, S61, unsupervised loss is calculated to process unlabeled microbial experimental colony image data, and the L1 loss is calculated by combining the pseudo labels generated by the student network and the teacher network to ensure that the model learns effective information from the pseudo labels. The specific formula is: ; in, Generate latent variables for the student network at the jth scale; Latent variables generated for the teacher network; is the weighting coefficient used to adjust the contribution of unsupervised loss at different scales.

[0046] Further, S62, calculate L1 loss , SSIM loss and perceived loss , the specific formula is: ; in, is the pixel value of the real image at the jth scale, is the predicted value of the student network at the jth scale; is the total number of scales; is the weight coefficient for each scale; ; in, is the structural similarity between the real image and the image generated by the student network at the jth scale; is the weight coefficient for each scale; ; at last, Features extracted by the jth layer of the pre-trained VGG network; is the weight coefficient of the j-th scale feature; It is used to measure the L2 norm of feature differences.

[0047] Further, S63, L1 loss, SSIM loss and perceptual loss are combined to form a weighted combination loss function to optimize the learning objective of the model. The specific formula is: ; in, and To control the hyperparameters of the impact of different loss items on the final total loss, here is equal to 0.8, Equal to 0.8.

[0048] Step S7: Specifically, all the above modules are integrated into a complete network to build the final enhanced network; the optimal latent variables are used for training under Python 3.10.12 version, and the PyTorch 2.5.1 framework and Adam optimizer are used for 1000 iterations of training. Since the video memory size is 12GB, the batch size is set to 32, the initial learning rate is 0.004, and the learning rate is dynamically adjusted during the training process. After the training data is completed, the optimal enhanced model file is obtained; as shown in the attached figure. Figure 3 The model can infer and enhance the colony images of microbial experiments to be processed, and generate enhanced images with higher clarity and stronger feature visibility; this integration method gives full play to the advantages of each module, realizes the combination of supervised learning and unsupervised learning, effectively utilizes labeled and unlabeled data, significantly improves the generalization ability of the model and the image enhancement effect, and meets the needs of high-quality image processing in microbial experiments. Figure 4 Schematic diagram of the comparison of microbial experimental colony images before and after processing.

Claims

1. A method for processing microbial experimental colony images, characterized in that: The following steps are involved: S1. Microbial experimental colony image dataset preparation: collect microbial experimental colony images, 30% of which are annotated and 70% are not annotated; S2, design the student enhancement network module, extract the features of the microbial experimental colony image, and generate the enhanced image as the student network latent variable; S3. Design a mean teacher-assisted learning network module, improve the nonlinear transformation strategy to map the input feature information to a high-dimensional feature space, and introduce an adaptive regularization coefficient optimization model to generate pseudo labels and pseudo teacher network latent variables; S4. Design a mean teacher-assisted Gaussian process regression loss module, design the weighted error between the pseudo latent variable and the actual latent variable, combine the uncertainty measure of the covariance matrix, and use the Gaussian process regression loss function to guide the learning process; S5. Design an optimized latent variable selection OLVS module to select the best quality and most diverse latent variable set from the latent variables generated by the mean teacher-assisted learning network and the student-enhanced network by optimizing the peak signal-to-noise ratio and diversity score of the latent variables. S6. Design a loss function module, optimize and design different loss methods, and introduce weighted coefficients and scale factors to perform weighted optimization according to different scales and microbial experimental colony image characteristics; S7. Construct the final network to generate the image enhancement model, integrate the above modules into a complete network, generate the image enhancement model; and use the trained model to process the input microbial experimental colony image and output the enhanced image.

2. A microbial experiment colony image processing method according to claim 1, characterized in that: The mean teacher-assisted learning network module in S3 has the following specific method: S31. The feature vector U extracted from the microbial experimental colony image is mapped to a high-dimensional feature space through nonlinear transformation. The calculation formula is as follows: ； in, is the feature vector of the input microbial experimental colony image, The input feature vector The representation of the mapping in the high-dimensional feature space, To calculate the weight matrix of the linear transformation passed from the kth layer to the next layer, is the PReLu activation function, To calculate the weight matrix between the kth layer and each layer, To adjust the bias term of the k-th layer output information, Is the number of layers of the network; S32. Use weighted basis function expansion to model the output, and introduce noise terms to simulate the random disturbance of data in the process of colony image enhancement in actual microbial experiments. The specific formula is: ； in, is the output value, which represents the feature vector of the colony image according to the input microbial experiment and the transformed features The calculated forecast results; is the weight coefficient of the basis function, indicating the contribution intensity of each basis function to the final output; is the frequency parameter of each basis function, which is used to control the oscillation speed of each basis function, reflecting its sensitivity to different input features; The offset of each basis function controls the phase of the basis function; It is the noise term that simulates the random disturbance of data in the process of microbial colony image enhancement; is the number of design basis functions; S33, optimize the prediction process through weighted regression and gradient regularization, and control the complexity of the model to ensure its good generalization ability. The specific formula is: ； in, is the training loss function, which represents the goal of model optimization; is the weight of each sample, indicating the contribution of the data point in the loss function during the microbial experimental colony image enhancement process; is the data fitting error term, indicating the predicted value and the true label The error between is the gradient regularization term; is the adaptive regularization weight, the specific formula is: ； in, To initialize the regularization weights, To use the predicted value The variance of is used to reflect the uncertainty of the sample; Adjust the parameters to control the amplification effect of variance on the regularization coefficient; is the L2 norm of the current gradient, is the correlation between the target value and the predicted value, Factors that adjust the weights of correlation effects; To prevent extremely small numbers with a denominator of zero.

3. A microbial experiment colony image processing method according to claim 2, characterized in that: The mean teacher assisted Gaussian process regression loss module in S4 is specifically: S41. Calculate the weighted error between the pseudo latent variable and the actual latent variable in the process of microbial experiment colony image enhancement. The specific formula is: ； in, is a pseudo latent variable, which is an estimated value of the latent variable generated by the mean teacher-assisted learning network module in step S2; is the mean of the latent variable in the Gaussian process model; is the covariance matrix, which represents the dependency between latent variables, and its inverse matrix represents the weighted error; To measure the pseudo latent variable With the Gaussian process model mean The weighted gap between S42. By taking the logarithm of the determinant of the covariance matrix, the model can control the prediction uncertainty of the latent variables and avoid overfitting. The specific formula is: ； in, is the covariance matrix The logarithm of the determinant of , which represents a measure of model uncertainty; is a regularization hyperparameter used to control the weight of the regularization term; is the trace of the inverse of the covariance matrix and the covariance matrix of the pseudo latent variable; is the logarithm of the covariance term, controlling for model uncertainty; S43. Combining the above two results, the latent variable prediction of the microbial experiment colony image enhancement result is optimized by maximizing the posterior distribution. The specific formula is: ； in, and are the hyperparameters for adjusting the error and prior knowledge terms.

4. A microbial experiment colony image processing method according to claim 3, characterized in that: The optimized latent variable selection OLVS module in S5 selects a latent variable set with the best quality and greater diversity from the latent variables generated by the teacher network and the student network. The specific method is as follows: S51. The PSNR score is obtained by calculating the peak signal-to-noise ratio to measure the difference between each potential variable and the real microbial experimental colony image. The specific formula is: ； in, is the PSNR score of the ith latent variable, For real images, is the ith latent variable, is the maximum pixel value of the real image used to normalize the PSNR score, is the mean square error between the real image and the i-th latent variable; S52. Use the selection function to select the optimal latent variable according to the PSNR score. The specific formula is: ； ； in, is the optimal latent variable set, including the selected N latent variables; i represents the i-th latent variable; is the real image label; Latent variables generated for the teacher network; Latent variables generated for student networks; is the weight of the PSNR score; the weighting given to scoring diversity; is the penalty coefficient of the error term; Differences between latent variables generated to measure teacher and student networks; To score diversity, we measure the degree of difference between the current latent variable and other latent variables; is the maximum independent variable point set function.

5. A microbial experiment colony image processing method according to claim 3, characterized in that: The loss function module in S6 uses multiple loss functions, including L1 loss, SSIM loss and perceptual loss. The specific method is as follows: S61. Calculate the unsupervised loss to process the unlabeled microbial experimental colony image data, and calculate the L1 loss by combining the pseudo labels generated by the student network and the teacher network to ensure that the model learns effective information from the pseudo labels. The specific formula is: ； in, Generate latent variables for the student network at the jth scale; Latent variables generated for the teacher network; is the weighting coefficient used to adjust the contribution of unsupervised loss at different scales; S62. Calculate L1 loss , SSIM loss and perceived loss , the specific formula is: ； in, is the pixel value of the real image at the jth scale, is the predicted value of the student network at the jth scale; is the total number of scales; is the weight coefficient for each scale; ； in, is the structural similarity between the real image and the image generated by the student network at the jth scale; is the weight coefficient for each scale; ； at last, Features extracted by the jth layer of the pre-trained VGG network; is the weight coefficient of the j-th scale feature; L2 norm is used to measure feature differences; S63. Combine L1 loss, SSIM loss and perceptual loss to form a weighted combination loss function to optimize the learning objective of the model. The specific formula is: ； in, and It is a hyperparameter that controls the impact of different loss terms on the final total loss.

Images