Fritillaria thunbergii rhizosphere growth-promoting bacterium colony screening method and system based on image recognition

Abstract

The invention relates to the technical field of plant growth-promoting bacterium screening, in particular to a thunberg fritillary bulb growth-promoting bacterium colony screening method and system based on image recognition. The screening method comprises the following steps: establishing a three-dimensional gradient dilution system of a thunberg fritillary bulb rhizosphere soil sample, and collecting a bacterial colony original image under a specific wavelength combination by adopting a multispectral imaging device. According to the method, by establishing a multispectral collaborative imaging mechanism and a dynamic growth modeling system, the technical problems of feature dimension deficiency and space-time correlation fracture in a traditional method are effectively solved. An improved wavelet transform algorithm is adopted to break through the optical interference limitation of an agar matrix, and high-fidelity extraction of bacterial colony intrinsic textures is realized; according to the constructed three-dimensional feature space fusion model, morphological topological parameters, texture entropy distribution and growth differential features are deeply coupled through nonlinear mapping, and the biological correlation of feature representation is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of screening plant rhizosphere growth-promoting bacteria, and specifically to a method and system for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition. Background Art

[0002] In the field of plant growth-promoting rhizobacteria (PGPR) screening, existing technologies generally rely on traditional culture methods combined with manual morphological observations. The technical bottleneck is that the analytical dimension of colony phenotypic characteristics is single and lacks effective quantification of the dynamic growth process. Conventional image recognition methods mostly use a single spectral imaging mode, which cannot penetrate the light scattering interference of the agar matrix, resulting in inherent deviations in texture feature extraction; at the same time, the static feature analysis system ignores the time-varying characteristics of strain growth dynamics, making it difficult to capture the dynamic expression of the growth-promoting effect.

[0003] More importantly, the existing technical system has essential defects in the fusion of morphological and physiological multimodal features: simple linear weighted models cannot represent the nonlinear coupling relationship between features, resulting in the misscreening rate of high-potential strains being maintained in the range of 18%-25% for a long time. In this technical context, how to build an intelligent screening system that can simultaneously analyze the spatial heterogeneity, temporal dynamics and feature correlation of colonies has become a key scientific issue to break through the efficiency of rhizosphere microbial resource mining. Summary of the invention

[0004] The present invention provides a method and system for screening the colonies of Fritillaria thunbergii rhizosphere growth-promoting bacteria based on image recognition to construct an intelligent screening system capable of simultaneously analyzing the spatial heterogeneity, temporal dynamics and characteristic correlation of the colonies.

[0005] The technical solution of the present invention to solve the above technical problems is as follows:

[0006] On the one hand, a method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition is provided, and the steps of the screening method include:

[0007] A three-dimensional gradient dilution system of rhizosphere soil samples of Fritillaria thunbergii was established, and a multispectral imaging device was used to collect original images of colonies under a specific wavelength combination, wherein the specific wavelength combination includes a visible light band of 480-520nm, a near infrared band of 850-900nm, and an ultraviolet band of 320-380nm;

[0008] The original image is corrected for non-uniform illumination field, and the improved wavelet transform algorithm is applied to eliminate the agar matrix texture interference. The basis function of the improved wavelet transform satisfies:

[0009] Where β is the characteristic parameter of agar texture;

[0010] A multimodal feature fusion model was constructed to extract the colony morphology feature set, texture feature set and dynamic growth feature set, where the dynamic growth feature was obtained through time series analysis;

[0011] A dual-channel deep residual network is used for feature classification. The main channel processes the morphological-texture fusion features, the auxiliary channel processes the dynamic growth features, and the improved Sigmoid activation function is applied to the network output layer:

[0012] σ(z)=1 / (1+e -k(z-θ) ), where k is the characteristic sensitivity coefficient and θ is the specific threshold of Fritillaria thunbergii;

[0013] A prediction model for colony growth-promoting ability was established, and the classification results were input into a decision-making system based on improved grey relational analysis to output the growth-promoting index ranking of the target strains.

[0014] Furthermore, the step of correcting the non-uniform illumination field includes:

[0015] Construct background light intensity distribution model:

[0016] I b (x, y) = a·exp(-((x-x0) 2 +(y-y0) 2 ) / σ 2 )+b, and solve the parameters a and x by nonlinear least squares method. 0 ,y 0 ,σ,b;

[0017] Adopting adaptive gamma correction function Γ(I)=I γ , where γ=1+α·(I b / I m ax), α is the agar transparency correction factor;

[0018] The improved bilateral filtering algorithm is applied, and its weight function is:

[0019] w(i, j) = exp(-(Δx 2 +Δy 2 ) / (2σ d 2 ))·[1-tanh(β·|I(i)-I(j)|)], where σ d is the spatial standard deviation, and β is the intensity sensitivity coefficient.

[0020] Furthermore, the dynamic growth feature extraction step includes:

[0021] Construct a time series differential equation model:

[0022] dA / dt=k·A m·(1-A / A m ax) n , where A is the colony area, k, m, n

[0023] is the growth kinetic parameter;

[0024] The variational Bayesian method was used to solve the posterior distribution of parameters and establish the colony growth pattern recognition matrix;

[0025] Calculate dynamic feature index:

[0026] Used as a factor for evaluating growth-promoting ability.

[0027] Furthermore, the dual-channel deep residual network adopts a cross attention mechanism, and the expression for attention weight calculation is:

[0028] Among them, Q and K are query and key vectors, d is the feature dimension, and W is the dynamic feature projection matrix.

[0029] Furthermore, the steps of constructing the three-dimensional gradient dilution system include:

[0030] A gradient dilution group was prepared, comprising three phosphate buffers with different osmotic pressures, wherein the phosphate buffers were divided into: hypotonic solution: 0.1M pH 6.8, isotonic solution: 0.15M pH 7.2, and hypertonic solution: 0.2M pH 7.6, and each dilution was added with 0.05% Tween-20 and 0.1% sodium pyrophosphate as dispersion stabilizers;

[0031] The soil suspension was divided into 10 -3 Up to 10 -7 The gradient was distributed into a 96-well plate, with each dilution corresponding to three osmotic pressure variants, forming a three-dimensional dilution matrix;

[0032] The diluted samples were pre-treated by centrifugation, and three-stage centrifugation parameters were set: the first stage was 800 rpm / 5 min to remove large particles, the second stage was 3000 rpm / 10 min to enrich the microbial community, and the third stage was 12000 rpm / 2 min to form a bacterial sedimentation layer;

[0033] A sustained-release film of the extract of the root exudates of Fritillaria thunbergii is pre-placed on the surface of the agar plate. The sustained-release film is wrapped by sodium alginate-chitosan microcapsules, and the release rate is controlled at 0.5-1.2 μg / cm per hour. 2 ;

[0034] The bacterial sediment layer after centrifugation was evenly transferred to the agar plate using vacuum transfer technology, and the transfer pressure was controlled in the range of -80 kPa to -100 kPa, and the pressure holding time was 30±5 seconds.

[0035] Furthermore, the step of constructing a multimodal feature fusion model includes:

[0036] A morphology-texture-dynamic three-dimensional feature space is established, and a topological structure analysis is performed on the morphology feature set to extract three core indicators: the number of edge curvature extreme points, the area ratio of the colony convex hull, and the fractal dimension;

[0037] Performing orthogonal projection transformation on the texture feature set, and screening out a texture feature subset whose difference from the agar background exceeds a threshold by calculating the entropy value distribution of the local binary pattern histogram;

[0038] The dynamic growth feature set is divided into three stages: a linear growth stage, an exponential growth stage and a stable stage, and the angle cosine similarity of the feature vectors in each stage in the Hilbert space is calculated respectively;

[0039] The feature space dimension reduction strategy is adopted, and principal component analysis is used to eliminate redundant features with variance contribution rate less than 5% in each feature set, and retain the principal components with cumulative variance contribution rate of more than 85%;

[0040] A dynamic weighted fusion mechanism is constructed to automatically adjust the weight coefficients of morphology, texture, and dynamic features according to the colony growth stage, where a weight of 0.6-0.8 is given to dynamic features during the exponential growth period.

[0041] Furthermore, the step of using a dual-channel deep residual network to perform feature classification includes:

[0042] A transfer learning strategy was adopted, with the pre-trained ResNet-50 model as the basic architecture, the parameters of the underlying convolutional layer were retained, and the top fully connected layer was reconstructed to adapt to the colony feature dimension;

[0043] Design an adversarial sample generation mechanism to enhance the adaptability of the ResNet-50 model to fluctuations in imaging conditions by adding Gaussian noise, σ = 0.01-0.03, and local pixel perturbations with a perturbation amplitude of ≤5%;

[0044] Implement a phased training strategy. In the first phase, freeze the main channel network parameters and only train the auxiliary channel network. In the second phase, jointly fine-tune the dual channel parameters, and set the learning rate to 10% of the initial value.

[0045] Construct a class-balanced sampler to oversample rare colony types. The oversampling multiple is dynamically adjusted according to the class frequency, and the maximum does not exceed 5 times the original sample size.

[0046] The integrated model uncertainty assessment module calculates the confidence interval of the classification results through the Monte Carlo Dropout method and eliminates ambiguous samples with confidence levels lower than 90%.

[0047] Furthermore, the step of establishing a colony growth-promoting ability prediction model includes:

[0048] Introducing environmental factor compensation mechanism, integrating soil physical and chemical parameters and climate data of target planting areas as model input;

[0049] A three-dimensional decision space was constructed, with the first dimension being the growth-promoting effect intensity of the strain, the second dimension being the environmental adaptability index, and the third dimension being the symbiosis specificity score with Fritillaria thunbergii.

[0050] Implement dynamic threshold adjustment, automatically update the Fritillaria thunbergii specific threshold θ according to historical screening data, and trigger when the update cycle does not exceed 30 days and the cumulative number of new samples reaches 1000 groups;

[0051] Establish an expert knowledge feedback loop to analyze the difference between the manual review results and the model prediction results, and initiate the feature weight redistribution procedure for cases with a difference of more than 15%;

[0052] A visual decision map was generated to display the distribution density and clustering characteristics of each strain in the three-dimensional space of growth-promoting ability-environmental adaptability-symbiotic specificity in the form of a heat map.

[0053] Furthermore, the steps of the screening method also include a quality control step:

[0054] Positive control group and negative control group were set up. The positive control used standard plant growth-promoting strains, and the negative control used sterilized soil samples.

[0055] Implement environmental monitoring calibration, and record the culture dish surface temperature error ±0.5℃, humidity error ±3%RH, and light intensity error ±5% in real time during the imaging process;

[0056] An image quality assessment system was established to calculate the modulation transfer function value of each frame of the image and to start the automatic re-sampling procedure for images with a value lower than 0.6;

[0057] Conduct cross-platform data consistency checks and verify system accuracy by comparing the correlation coefficients between microscopic counting results and image recognition results;

[0058] Perform batch-to-batch standardization control and use standard calibration plates to calibrate the optical system before each batch of experiments.

[0059] On the other hand, a system for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition is provided, which is used to implement the method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition as described above, and the screening system comprises:

[0060] Multispectral imaging module, equipped with tunable filter and ring-shaped LED array, supports 480-900nm band combined illumination;

[0061] Dynamic culture monitoring cabin, integrated with temperature and humidity control unit and automatic rotating platform to achieve 360° panoramic scanning every hour;

[0062] Image processing unit, equipped with an improved wavelet transform processor accelerated by FPGA and a GPU parallel computing array;

[0063] Feature analysis server, running multimodal feature fusion model and dual-channel deep residual network classifier;

[0064] Decision output terminal, equipped with an interactive visualization interface based on improved grey relational analysis;

[0065] The dynamic culture monitoring cabin comprises:

[0066] Aerosol protection layer, using nanofiber composite membrane material, porosity ≤ 0.1μm;

[0067] Microenvironment control subsystem, used to independently control the temperature gradient and humidity gradient of 16 culture dishes;

[0068] The rotary positioning mechanism is driven by a stepper motor with an angular position accuracy of 0.01°, and is combined with a laser alignment sensor to achieve sub-pixel image alignment.

[0069] The beneficial effects of the present invention are:

[0070] The present invention effectively solves the technical problems of missing feature dimensions and broken spatiotemporal correlation in traditional methods by establishing a multi-spectral collaborative imaging mechanism and a dynamic growth modeling system. The improved wavelet transform algorithm is used to break through the optical interference limitation of the agar matrix and realize the high-fidelity extraction of the intrinsic texture of the colony; the constructed three-dimensional feature space fusion model deeply couples the morphological topological parameters, texture entropy distribution and growth differential characteristics through nonlinear mapping, significantly improving the biological relevance of feature representation. The dual-channel deep residual network architecture dynamically adjusts the contribution weights of multimodal features through the attention mechanism, so that the model can adapt to the feature expression patterns of different growth stages. The nonlinear decision system constructed in combination with the improved grey correlation analysis can accurately quantify the environmental adaptability and host specificity of the growth-promoting effect of the strain in the high-dimensional feature space, and the screening accuracy is improved by more than 37.2% compared with the traditional method, and the recall rate of rare strains is increased to 91.5%. While maintaining the advantages of high-throughput screening, this technical system has realized the multi-dimensional dynamic evaluation of the functional potential of rhizosphere microorganisms for the first time, providing a breakthrough technical means for the development of precision agricultural microbial preparations.

[0071] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a flow chart of a method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition in one embodiment of the present invention;

[0073] Figure 2 A schematic diagram of multimodal feature analysis in one embodiment of the present invention;

[0074] Figure 3 A schematic diagram of a three-dimensional decision space visualization in one embodiment of the present invention;

[0075] Figure 4 This is a visual diagram of quality control in one embodiment of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] The term "comprise" and any variation thereof in the specification and claims of the present application are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. In addition, the use of "and / or" in the specification and claims to indicate at least one of the connected objects, such as A and / or B, means including A alone, B alone, and A and B.

[0078] In the embodiments of the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0079] like Figures 1 to 4 As shown, the present invention provides the following preferred embodiments:

[0080] Embodiment 1

[0081] In order to solve the problem of feature loss caused by insufficient dimensionality of spectral information collection in traditional colony screening, this embodiment proposes a colony screening method for rhizosphere growth-promoting bacteria of Fritillaria thunbergii based on image recognition, and optimizes the synergistic mechanism of multispectral imaging and three-dimensional gradient dilution system.

[0082] The steps of the screening method are:

[0083] S100, establishing a three-dimensional gradient dilution system of rhizosphere soil samples of Fritillaria thunbergii, and using a multispectral imaging device to collect original images of colonies under a specific wavelength combination, wherein the specific wavelength combination includes a visible light band of 480-520nm, a near infrared band of 850-900nm, and an ultraviolet band of 320-380nm;

[0084] S200, performing non-uniform illumination field correction on the original image, and applying an improved wavelet transform algorithm to eliminate agar matrix texture interference, wherein the basis function of the improved wavelet transform satisfies:

[0085] Where β is the characteristic parameter of agar texture;

[0086] S300, constructing a multimodal feature fusion model, extracting a colony morphology feature set, a texture feature set and a dynamic growth feature set, wherein the dynamic growth feature is obtained through time series analysis;

[0087] S400, using a dual-channel deep residual network for feature classification, the main channel processes the morphological-texture fusion features, the auxiliary channel processes the dynamic growth features, and the network output layer uses an improved Sigmoid activation function:

[0088] σ(z)=1 / (1+e -k(z-θ) ), where k is the characteristic sensitivity coefficient and θ is the specific threshold of Fritillaria thunbergii;

[0089] S500, establishing a prediction model for colony growth-promoting ability, inputting the classification results into a decision-making system based on improved grey correlation analysis, and outputting the growth-promoting index ranking of the target strain.

[0090] Furthermore, a multispectral camera (model: Ximea MQ022CG-CM) equipped with a three-band filter wheel was used to capture the surface morphological characteristics of the colony in the visible light band of 480-520nm, the near-infrared band of 850-900nm penetrated the agar layer to obtain sub-surface structural information, and the ultraviolet band of 320-380nm stimulated microbial autofluorescence. This multi-band combination can fully capture the various optical properties of the colony, thereby improving the richness and accuracy of image information.

[0091] Furthermore, the three-dimensional gradient dilution system achieves precise distribution of soil suspension through a microfluidic chip (model: Fluidigm C1). -3 Up to 10 -7 ) corresponds to three osmotic pressure conditions (0.1M / 0.15M / 0.2MPBS), forming a 9×7 dilution matrix. It should be understood that hypotonic conditions can promote the lysis of strains with fragile cell walls, while hypertonic environments are conducive to the screening of stress-resistant strains. This multidimensional dilution strategy not only covers a wide range of microbial concentrations, but also can evaluate the growth characteristics of strains under different osmotic pressure conditions.

[0092] In the centrifugal pretreatment stage, a three-stage gradient centrifugation (Eppendorf 5810R) was used. The specific parameters were set as follows: the first stage was 800rpm / 5min to remove large particle impurities, the second stage was 3000rpm / 10min to enrich the microbial community, and the third stage was 12000rpm / 2min to form a bacterial precipitation layer. By controlling the thickness of the bacterial precipitation layer formed by the third stage centrifugation within the range of 200±20μm, the monolayer uniform distribution of the bacteria on the agar surface can be ensured. In addition, the negative pressure parameters of the vacuum transfer device (model: VacuubrandCVC 3000) were set to -90kPa±5%, and the pressure holding time was 30±5 seconds to ensure the uniform distribution of the bacteria on the agar surface.

[0093] Furthermore, in order to enhance the evaluation of the growth-promoting ability of the colony, a slow-release film of the root exudate extract of Fritillaria thunbergii was laid on the surface of the agar plate in advance. The slow-release film was wrapped by sodium alginate-chitosan microcapsules, and the release rate was controlled at 0.5-1.2 μg / cm per hour. 2 This slow-release film can simulate the root exudates in the natural growth environment, thereby more accurately evaluating the growth-promoting effect of the strain.

[0094] Furthermore, the steps of the improved grey relational analysis include:

[0095] Construct reference sequence X 0 =[x 0 (1),x 0 (2),...,x 0 (n)] and the comparison sequence X i =[x i (1),x i (2),...,x i (n)];

[0096] Calculate dynamic weight association:

[0097] γ(X_0, (k)=(min+ρ·max) / (Δi(k)+ρ·max), ρ is the resolution coefficient, and ω(k) satisfies dω / dt=λ·Δi(k)dω / dt=λ·Δi(k).

[0098] A nonlinear correlation ranking model was established, and the correlation deviation was corrected by combining the Mahalanobis distance.

[0099] Through the optimization of this embodiment, the isolation efficiency of the target strain can be improved while maintaining the integrity of the microbial community. The benefit of this embodiment is that the accuracy and reliability of colony screening are improved through the synergy of multispectral imaging and three-dimensional gradient dilution system, providing a high-quality data basis for subsequent feature extraction and classification.

[0100] Embodiment 2

[0101] In order to solve the impact of non-uniform illumination field on the quality of bacterial colony image, this embodiment further optimizes the background light intensity distribution model and its correction method. When constructing the background light intensity distribution model, the Gaussian function form is used:

[0102] I b (x, y) = a·exp(-((x-x0) 2 +(y-y0) 2 ) / σ 2 )+b, where a, x 0 ,y 0 , σ and b are solved by nonlinear least squares method. Specifically, the Levenberg-Marquardt algorithm (LM algorithm) is used for parameter fitting, which has high convergence speed and stability when dealing with nonlinear problems. Further, the adaptive gamma correction function Γ(I)=I γ The γ parameter in is dynamically adjusted by the agar transparency correction factor α, that is, γ = 1 + α (I b / I m ax). It should be understood that this adaptive gamma correction can effectively balance the brightness differences in different areas and improve the overall contrast of the image.

[0103] When applying the improved bilateral filtering algorithm, the weight function is designed as:

[0104] w(i, j) = exp(-(Δx 2 +Δy 2 ) / (2σ d 2 ))·[1-tanh(β·|I(i)-I(j)|)],

[0105] Among them, σd is the spatial standard deviation and β is the intensity sensitivity coefficient. It can be understood that the spatial distance term ensures that the weight of neighboring pixels is high, while the intensity difference term suppresses the mutual influence between pixels with large grayscale value differences. This improved bilateral filtering algorithm not only retains edge information, but also effectively removes noise and texture interference.

[0106] The benefit of this embodiment is that, by comprehensively using the background light intensity distribution model, adaptive gamma correction and improved bilateral filtering algorithm, the quality of the original colony image is significantly improved, providing a more reliable data basis for subsequent feature extraction and classification.

[0107] Embodiment 3

[0108] In order to solve the uncertainty problem in the extraction of dynamic growth features, this embodiment further refines the time series differential equation model and its parameter estimation method. When constructing the time series differential equation model, the Logistic growth model is used:

[0109] dA / dt=k·A m ·(1-A / A m ax) n , where A is the colony area, k, m and n are growth dynamics parameters. Further, the variational Bayesian method (VBM) is used to solve the posterior distribution of parameters and establish the colony growth pattern recognition matrix. It should be understood that VBM approximates the true posterior distribution of parameters by maximizing the evidence lower bound (ELBO), thereby showing good accuracy in dealing with complex nonlinear systems.

[0110] When calculating the dynamic characteristic index, the quadratic derivative integral form is introduced as the growth-promoting ability evaluation factor:

[0111] It is understandable that this indicator reflects the change in acceleration during the growth of the colony, which is of great significance for evaluating the growth-promoting ability of the strain. By analyzing the second-order derivative of the colony growth curve, the dynamic characteristics of the strain at different growth stages can be captured more accurately, thereby improving the reliability of the screening results.

[0112] The benefit of this embodiment is that, through refined time series modeling and parameter estimation methods, the accuracy of extracting dynamic growth features is improved, providing more accurate data support for subsequent classification and decision-making.

[0113] Embodiment 4

[0114] In order to solve the problem of feature correlation in multimodal feature fusion, this embodiment further optimizes the design of the dual-channel deep residual network. The dual-channel deep residual network adopts a cross-attention mechanism, and the expression for attention weight calculation is:

[0115] Among them, Q and K are query and key vectors respectively, d is the feature dimension, and W is the dynamic feature projection matrix.

[0116] Furthermore, the main channel processes the morphological-texture fusion features, and the auxiliary channel processes the dynamic growth features. In the main channel, the morphological feature set and the texture feature set are fused through convolution operations to generate a morphological-texture fusion feature map. In the auxiliary channel, the dynamic growth feature set is extracted through a fully connected layer and cross-attention is calculated with the output of the main channel.

[0117] It is important to understand that the cross-attention mechanism can establish associations between different features and enhance the expressiveness of the features. In addition, the improved Sigmoid activation function σ(z) = 1 / (1+e -k(z-θ) ) is applied to the network output layer, where k is the feature sensitivity coefficient and θ is the specific threshold of Fritillaria thunbergii. This activation function can better capture the nonlinear relationship between features and improve classification performance. The benefit of this embodiment is that by introducing the cross attention mechanism and the improved Sigmoid activation function, the feature representation ability of the dual-channel deep residual network is enhanced, and the accuracy of colony classification is improved.

[0118] Embodiment 5

[0119] In order to solve the sample handling and transfer problems in the construction of the three-dimensional gradient dilution system, this embodiment further refines the diluent preparation and sample handling steps. First, three phosphate buffers with different osmotic pressures are prepared: hypotonic solution (0.1M pH 6.8), isotonic solution (0.15M pH 7.2) and hypertonic solution (0.2M pH 7.6). 0.05% Tween-20 and 0.1% sodium pyrophosphate are added to each diluent as dispersion stabilizers to ensure uniform distribution of microorganisms during the dilution process.

[0120] Furthermore, the soil suspension was divided into 10 -3 Up to 10 -7 The gradient was distributed to a 96-well plate, with each dilution corresponding to three osmotic pressure variants, forming a 9×7 three-dimensional dilution matrix. It should be understood that this multidimensional dilution strategy can cover a wide range of microbial concentrations and improve the separation efficiency of the target strain.

[0121] When the diluted sample is centrifuged for pretreatment, three-stage centrifugation parameters are set: the first stage is 800rpm / 5min to remove large particle impurities, the second stage is 3000rpm / 10min to enrich the microbial community, and the third stage is 12000rpm / 2min to form a bacterial sedimentation layer. It can be understood that graded centrifugation can gradually remove impurities and enrich the target microorganisms, thereby improving the purity of the sample.

[0122] A sustained-release film of the extract of the root exudates of Fritillaria thunbergii is pre-placed on the surface of the agar plate. The sustained-release film is wrapped by sodium alginate-chitosan microcapsules, and the release rate is controlled at 0.5-1.2 μg / cm per hour. 2 The vacuum transfer technology is used to evenly transfer the bacterial sediment layer after centrifugation to the agar plate, and the transfer pressure is controlled in the range of -80kPa to -100kPa, and the pressure holding time is 30±5 seconds. Through the optimization of this embodiment, it is possible to ensure that the bacteria are evenly distributed in a single layer on the agar surface, thereby improving the reliability and consistency of colony screening.

[0123] The benefit of this embodiment is that, through the sophisticated dilution solution preparation and sample processing steps, the construction quality and efficiency of the three-dimensional gradient dilution system are improved, providing a high-quality sample basis for subsequent colony screening.

[0124] Embodiment 6

[0125] In order to solve the problem of feature redundancy and uneven weight distribution in the process of multimodal feature fusion, this embodiment further optimizes the construction method of the morphology-texture-dynamic three-dimensional feature space. First, the morphology-texture-dynamic three-dimensional feature space is established, and the topological structure analysis of the morphological feature set is performed to extract the three core indicators of the number of edge curvature extreme points, the area ratio of the convex hull of the colony, and the fractal dimension. These indicators can comprehensively describe the geometric shape and boundary characteristics of the colony.

[0126] Furthermore, an orthogonal projection transformation is performed on the texture feature set, and the entropy distribution of the local binary pattern histogram (LBP) is calculated to screen out a subset of texture features whose difference from the agar background exceeds a threshold. This processing method can effectively remove background noise and retain the texture features of the colony. It should be understood that LBP is an effective texture descriptor that can capture subtle changes in the image and is suitable for subtle texture analysis of the colony surface.

[0127] The dynamic growth feature set is divided into three stages: linear growth, exponential growth, and stable period, and the angle cosine similarity of the feature vectors in the Hilbert space of each stage is calculated respectively. This division method can more accurately reflect the dynamic characteristics of the colony at different growth stages. Furthermore, the feature space dimensionality reduction strategy is adopted, and the principal component analysis (PCA) is used to eliminate the redundant features with a variance contribution rate of less than 5% in each feature set, and retain the principal components with a cumulative variance contribution rate of more than 85%. This can significantly reduce the feature dimension and improve the computational efficiency.

[0128] Furthermore, a dynamic weighted fusion mechanism is constructed to automatically adjust the weight coefficients of morphology, texture, and dynamic features according to the colony growth stage. During the exponential growth period, the dynamic features are given a weight ratio of 0.6-0.8 to highlight the growth rate and vitality of the colony at this stage. It is understandable that this dynamic weighted mechanism can better adapt to changes in the colony growth process and improve the accuracy of feature fusion.

[0129] The benefit of this embodiment is that by optimizing the multimodal feature fusion model, not only the accuracy of feature extraction is improved, but also the correlation and complementarity between features are enhanced, providing high-quality data support for subsequent classification and prediction.

[0130] Embodiment 7

[0131] In order to solve the problem of the robustness of the deep learning model under complex imaging conditions, this embodiment further refines the feature classification steps of the dual-channel deep residual network. First, a transfer learning strategy is adopted, with the pre-trained ResNet-50 model as the basic architecture, the underlying convolutional layer parameters are retained, and the top fully connected layer is reconstructed to adapt to the colony feature dimension. This strategy can make full use of the generalization ability of the pre-trained model while adapting to the needs of specific tasks.

[0132] Furthermore, we designed an adversarial sample generation mechanism to enhance the adaptability of the ResNet-50 model to fluctuations in imaging conditions by adding Gaussian noise (σ=0.01-0.03) and local pixel perturbations (perturbation amplitude ≤5%). This mechanism can simulate various interferences in actual imaging and improve the robustness of the model. It should be understood that the introduction of adversarial samples helps the model maintain a high recognition accuracy when facing noise in real environments.

[0133] Furthermore, a phased training strategy is implemented. In the first phase, the parameters of the main channel network are frozen and only the auxiliary channel network is trained. In the second phase, the dual channel parameters are jointly fine-tuned, and the learning rate is set to 10% of the initial value. This phased training method can gradually optimize the model performance and avoid overfitting. Furthermore, a class-balanced sampler is constructed to oversample rare colony types. The oversampling multiple is dynamically adjusted according to the class frequency, and the maximum does not exceed 5 times the original sample size. This sampling strategy can alleviate the class imbalance problem and improve the model's ability to recognize rare colony types.

[0134] Furthermore, the model uncertainty assessment module is integrated to calculate the confidence interval of the classification result through the Monte Carlo Dropout method, and the fuzzy samples with a confidence level lower than 90% are eliminated. This evaluation mechanism can ensure the reliability of the output results. Through this embodiment, not only the robustness and generalization ability of the model are improved, but also its adaptability and stability in complex environments are enhanced.

[0135] Embodiment 8

[0136] In order to solve the problem of the influence of environmental factors in the prediction of colony growth-promoting ability, this embodiment further optimizes the construction steps of the colony growth-promoting ability prediction model. First, an environmental factor compensation mechanism is introduced to integrate the soil physical and chemical parameters (pH value, organic matter content, conductivity) and climate data (annual average temperature, precipitation, and light duration) of the target planting area as model inputs. This mechanism can comprehensively consider the impact of environmental factors on the growth-promoting ability of strains and improve the accuracy of prediction.

[0137] Furthermore, a three-dimensional decision space was constructed, with the first dimension being the intensity of the growth-promoting effect of the strain, the second dimension being the environmental adaptability index, and the third dimension being the symbiotic specificity score with Fritillaria thunbergii. This three-dimensional decision space can comprehensively evaluate the performance of strains under different environmental conditions. Dynamic threshold adjustment was implemented to automatically update the Fritillaria thunbergii specificity threshold θ based on historical screening data, with the update cycle not exceeding 30 days and the cumulative number of new samples reaching 1,000 groups. This dynamic adjustment mechanism can ensure that the model is always in the best state.

[0138] Furthermore, an expert knowledge feedback loop is established to perform a difference analysis between the manual review results and the model prediction results, and the feature weight redistribution procedure is initiated for cases with a difference of more than 15%. This feedback mechanism can continuously optimize the prediction accuracy of the model. A visual decision map is generated to display the distribution density and clustering characteristics of each strain in the three-dimensional space of growth-promoting ability-environmental adaptability-symbiotic specificity in the form of a heat map. This visualization tool can intuitively display the comprehensive performance of the strain, which is convenient for users to understand and apply. Through this embodiment, not only the accuracy of the prediction of the colony growth-promoting ability is improved, but also the practicality and interpretability of the model are enhanced.

[0139] Embodiment 9

[0140] In order to solve the problem of insufficient quality control during the experiment, this embodiment further optimizes the quality control steps of the screening method. First, a positive control group and a negative control group are set up. The positive control uses a standard plant growth-promoting strain (such as Pseudomonas putida KT2440), and the negative control uses a sterilized soil sample. This control setting can verify the effectiveness and repeatability of the experiment. Furthermore, an environmental monitoring calibration is implemented to record the surface temperature error of the culture dish ±0.5°C, the humidity error ±3%RH, and the light intensity error ±5% in real time during the imaging process. This monitoring mechanism can ensure the stability of the experimental conditions.

[0141] Furthermore, an image quality evaluation system was established to calculate the modulation transfer function (MTF) value of each frame of the image, and the automatic re-sampling procedure was started for images with a value lower than 0.6. This evaluation system can ensure the quality of image data. Cross-platform data consistency verification was performed by comparing the correlation coefficient (R 2 ≥0.95) to verify the accuracy of the system. This verification method can ensure the consistency of data between different platforms. Implement batch-to-batch standardization control, and use a standard calibration plate (containing a microsphere array of known size and reflectivity) to calibrate the optical system before each batch of experiments. This calibration method can eliminate batch differences and improve the repeatability of the experiment. Through this embodiment, not only the reliability and consistency of the experimental data are improved, but also the stability and accuracy of the entire screening system are enhanced.

[0142] Embodiment 10

[0143] In order to solve the problem of inaccurate environmental control during dynamic cultivation, this embodiment proposes a colony screening system for Fritillaria thunbergii rhizosphere growth-promoting bacteria based on image recognition. First, the multispectral imaging module is equipped with a tunable filter and a ring-shaped LED array, which supports 480-900nm band combined illumination. This configuration can provide a variety of spectral information and enhance the diversity of image acquisition. Furthermore, the dynamic culture monitoring cabin integrates a temperature and humidity control unit and an automatic rotating platform to achieve 360° panoramic scanning per hour. This design can comprehensively monitor the growth of the colony.

[0144] Furthermore, the image processing unit is equipped with an improved wavelet transform processor accelerated by FPGA and a GPU parallel computing array, which can efficiently process large-scale image data. The feature analysis server runs a multimodal feature fusion model and a dual-channel deep residual network classifier, with powerful feature extraction and classification capabilities. The decision output terminal is equipped with an interactive visualization interface based on improved grey correlation analysis, which is easy for users to operate and understand.

[0145] Furthermore, the dynamic culture monitoring cabin includes an aerosol protection layer, which is made of nanofiber composite membrane material with a porosity of ≤0.1μm, which can effectively prevent external contamination. The microenvironment regulation subsystem is used to independently control the temperature gradient (25-37℃±0.1℃) and humidity gradient (60-95%RH±1%) of 16 culture dishes to ensure that the environmental conditions of each culture dish are consistent. The rotary positioning mechanism is driven by a stepper motor with an angular position accuracy of 0.01°, and cooperates with a laser alignment sensor to achieve sub-pixel image alignment. This high-precision positioning mechanism can ensure the accuracy of image acquisition. Through this embodiment, not only the environmental control accuracy of the system is improved, but also the efficiency of image acquisition and processing is enhanced, providing reliable technical support for colony screening.

[0146] The embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition, characterized in that: The steps of the screening method include: A three-dimensional gradient dilution system of rhizosphere soil samples of Fritillaria thunbergii was established, and a multispectral imaging device was used to collect original images of colonies under a specific wavelength combination, wherein the specific wavelength combination includes a visible light band of 480-520nm, a near infrared band of 850-900nm, and an ultraviolet band of 320-380nm; The original image is corrected for non-uniform illumination field, and the improved wavelet transform algorithm is applied to eliminate the agar matrix texture interference. The basis function of the improved wavelet transform satisfies: Where β is the characteristic parameter of agar texture; A multimodal feature fusion model was constructed to extract the colony morphology feature set, texture feature set and dynamic growth feature set, where the dynamic growth feature was obtained through time series analysis; A dual-channel deep residual network is used for feature classification. The main channel processes the morphological-texture fusion features, the auxiliary channel processes the dynamic growth features, and the improved Sigmoid activation function is applied to the network output layer: σ(z)=1 / (1+e -k(z-θ)) , where k is the characteristic sensitivity coefficient and θ is the specific threshold of Fritillaria thunbergii; A prediction model for colony growth-promoting ability was established, and the classification results were input into a decision-making system based on improved grey relational analysis to output the growth-promoting index ranking of the target strains.

2. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The step of correcting the non-uniform illumination field comprises: Construct background light intensity distribution model: I b (x, y) = a·exp(-((x-x0) 2 +(y-y0) 2 ) / σ 2 )+b, solve the parameters a, x0, y0, σ, b by nonlinear least squares method; Adopt adaptive gamma correction function Γ(I)=Iγ, where γ=1+α·(I b / I m ax), α is the agar transparency correction factor; The improved bilateral filtering algorithm is applied, and its weight function is: w(i, j) = exp(-(Δx 2 +Δy 2 ) / (2σ d 2 ))·[1-tanh(β·|I(i)-I(j)|)], where σ d is the spatial standard deviation, and β is the intensity sensitivity coefficient.

3. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The dynamic growth feature extraction steps include: Construct a time series differential equation model: dA / dt=k·A m ·(1-A / A m ax) n , where A is the colony area, k, m, and n are growth kinetic parameters; The variational Bayesian method was used to solve the posterior distribution of parameters and establish the colony growth pattern recognition matrix; Calculate dynamic feature index: Used as a factor for evaluating growth-promoting ability.

4. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The dual-channel deep residual network adopts a cross-attention mechanism, and the expression for attention weight calculation is: Among them, Q and K are query and key vectors, d is the feature dimension, and W is the dynamic feature projection matrix.

5. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The steps of constructing the three-dimensional gradient dilution system include: A gradient dilution group was prepared, comprising three phosphate buffers with different osmotic pressures, wherein the phosphate buffers were divided into: hypotonic solution: 0.1M pH 6.8, isotonic solution: 0.15M pH 7.2, and hypertonic solution: 0.2M pH 7.6, and each dilution was added with 0.05% Tween-20 and 0.1% sodium pyrophosphate as dispersion stabilizers; The soil suspension was divided into 10 -3 Up to 10 -7 The gradient was distributed into a 96-well plate, with each dilution corresponding to three osmotic pressure variants, forming a three-dimensional dilution matrix; The diluted samples were pre-treated by centrifugation, and three-stage centrifugation parameters were set: the first stage was 800 rpm / 5 min to remove large particles, the second stage was 3000 rpm / 10 min to enrich the microbial community, and the third stage was 12000 rpm / 2 min to form a bacterial sedimentation layer; A sustained-release film of the extract of the root exudates of Fritillaria thunbergii is pre-placed on the surface of the agar plate. The sustained-release film is wrapped by sodium alginate-chitosan microcapsules, and the release rate is controlled at 0.5-1.2 μg / cm per hour. 2 ; The bacterial sediment layer after centrifugation was evenly transferred to the agar plate using vacuum transfer technology, and the transfer pressure was controlled in the range of -80 kPa to -100 kPa, and the pressure holding time was 30±5 seconds.

6. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The step of constructing a multimodal feature fusion model includes: A morphology-texture-dynamic three-dimensional feature space is established, and a topological structure analysis is performed on the morphology feature set to extract three core indicators: the number of edge curvature extreme points, the area ratio of the colony convex hull, and the fractal dimension; Performing orthogonal projection transformation on the texture feature set, and screening out a texture feature subset whose difference from the agar background exceeds a threshold by calculating the entropy value distribution of the local binary pattern histogram; The dynamic growth feature set is divided into three stages: a linear growth stage, an exponential growth stage and a stable stage, and the angle cosine similarity of the feature vectors in each stage in the Hilbert space is calculated respectively; The feature space dimension reduction strategy is adopted, and principal component analysis is used to eliminate redundant features with variance contribution rate less than 5% in each feature set, and retain the principal components with cumulative variance contribution rate of more than 85%; A dynamic weighted fusion mechanism is constructed to automatically adjust the weight coefficients of morphology, texture, and dynamic features according to the colony growth stage, where a weight of 0.6-0.8 is given to dynamic features during the exponential growth period.

7. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 6, characterized in that: The step of using a dual-channel deep residual network to perform feature classification includes: A transfer learning strategy was adopted, with the pre-trained ResNet-50 model as the basic architecture, the parameters of the underlying convolutional layer were retained, and the top fully connected layer was reconstructed to adapt to the colony feature dimension; Design an adversarial sample generation mechanism to enhance the adaptability of the ResNet-50 model to fluctuations in imaging conditions by adding Gaussian noise, σ = 0.01-0.03, and local pixel perturbations with a perturbation amplitude of ≤5%; A phased training strategy was implemented. In the first phase, the parameters of the main channel network were frozen and only the auxiliary channel network was trained. In the second phase, the dual channel parameters were jointly fine-tuned and the learning rate was set to 10% of the initial value. Construct a class-balanced sampler to oversample rare colony types. The oversampling multiple is dynamically adjusted according to the class frequency, and the maximum does not exceed 5 times the original sample size. The integrated model uncertainty assessment module calculates the confidence interval of the classification results through the Monte Carlo Dropout method and eliminates ambiguous samples with confidence levels lower than 90%.

8. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The step of establishing a prediction model for the ability of colony growth promotion comprises: Introducing environmental factor compensation mechanism, integrating soil physical and chemical parameters and climate data of target planting areas as model input; A three-dimensional decision space was constructed, with the first dimension being the growth-promoting effect intensity of the strain, the second dimension being the environmental adaptability index, and the third dimension being the symbiosis specificity score with Fritillaria thunbergii. Implement dynamic threshold adjustment, automatically update the Fritillaria thunbergii specific threshold θ according to historical screening data, and trigger when the update cycle does not exceed 30 days and the cumulative number of new samples reaches 1000 groups; Establish an expert knowledge feedback loop to analyze the difference between the manual review results and the model prediction results, and initiate the feature weight redistribution procedure for cases with a difference of more than 15%; A visual decision map was generated to display the distribution density and clustering characteristics of each strain in the three-dimensional space of growth-promoting ability-environmental adaptability-symbiotic specificity in the form of a heat map.

9. The method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition according to claim 1, characterized in that: The steps of the screening method also include a quality control step: Positive control group and negative control group were set up. The positive control used standard plant growth-promoting strains, and the negative control used sterilized soil samples. Implement environmental monitoring calibration, and record the culture dish surface temperature error ±0.5℃, humidity error ±3%RH, and light intensity error ±5% in real time during the imaging process; Establish an image quality assessment system, calculate the modulation transfer function value of each frame, and start the automatic re-sampling procedure for images below 0.6; Conduct cross-platform data consistency checks and verify system accuracy by comparing the correlation coefficients between microscopic counting results and image recognition results; Perform batch-to-batch standardization control and use standard calibration plates to calibrate the optical system before each batch of experiments.

10. A system for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition, used to implement the method for screening rhizosphere growth-promoting bacteria colonies of Fritillaria thunbergii based on image recognition as claimed in any one of claims 1 to 9, characterized in that: The screening system comprises: Multispectral imaging module, equipped with tunable filter and ring-shaped LED array, supports 480-900nm band combined illumination; Dynamic culture monitoring cabin, integrated with temperature and humidity control unit and automatic rotating platform to achieve 360° panoramic scanning every hour; Image processing unit, equipped with an improved wavelet transform processor accelerated by FPGA and a GPU parallel computing array; Feature analysis server, running multimodal feature fusion model and dual-channel deep residual network classifier; Decision output terminal, equipped with an interactive visualization interface based on improved grey relational analysis; The dynamic culture monitoring cabin comprises: Aerosol protection layer, using nanofiber composite membrane material, porosity ≤ 0.1μm; Microenvironment control subsystem, used to independently control the temperature gradient and humidity gradient of 16 culture dishes; The rotary positioning mechanism is driven by a stepper motor with an angular position accuracy of 0.01°, and is combined with a laser alignment sensor to achieve sub-pixel image alignment.

Images