Wheat scab spore germination form detection method based on microscopic image

Abstract

The invention relates to a wheat gibberellic disease spore germination form detection method based on microscopic images. Compared with the prior art, the wheat gibberellic disease spore germination form detection method solves the defect that germination form spore microscopic images are difficult to detect. The method comprises the following steps: collecting and preprocessing a wheat scab spore microscopic image; construction of a germination form target key point detection model; training a germination form target key point detection model; acquiring and preprocessing a medicament treatment spore image; and outputting a spore detection result. Through the germination convolution cascade module and the germination recursion gating module, spore sub-pixel level details and a germination tube global contour are accurately captured in a microscopic image, the problem of missing detection and false detection of tiny targets is solved, and microscopic image accurate detection of germinated spores is realized.

Description

Technical Field

[0001] The invention relates to the technical field of microscopic image processing, in particular to a method for detecting the germination phenotype of wheat scab spores based on microscopic images. Background Art

[0002] Wheat fusarium head blight is a major crop disease caused by Fusarium graminearum. The germination status of its spores directly determines the efficiency of disease transmission and the effectiveness of pesticide control. Traditional detection methods rely on manual microscopic observation, requiring visual judgment of spore germination rate and germination tube length ratio. These methods suffer from significant drawbacks such as low efficiency, strong subjectivity, and difficulty in quantification, making them difficult to meet the demands of rapid and accurate analysis of large-scale samples in high-throughput drug screening scenarios. In recent years, computer vision technology has been gradually applied to the field of microscopic image analysis. Although existing target detection models can achieve coarse positioning of spore targets, they still face severe challenges in practical applications.

[0003] Microscopic images of spores are characterized by their tiny size, diverse germination phenotypes (e.g., unilateral or bilateral germination), and complex backgrounds susceptible to uneven illumination and noise. Consequently, conventional models are unable to perceive sub-pixel features such as spore germination fissures and the terminus of the germination tube, leading to significant problems with missed and false detections. Furthermore, existing pose estimation methods struggle to stably regress the coordinates of key points at the terminus of the germination tube in complex microscopic scenes, significantly reducing the reliability of subsequent calculations of germination length ratios. Most technologies require phased detection and measurement, resulting in redundant processes and cumulative errors, making efficient end-to-end analysis impossible, further limiting the robustness and generalizability of the systems.

[0004] Although some studies have attempted to introduce classic image processing algorithms (such as edge detection and morphological operations) to assist in analysis, they are insufficiently sensitive to changes in spore morphology, making it difficult to distinguish different germination phenotypes or process overlapping targets. In addition, existing germination length ratio calculations are still highly dependent on manual annotation or semi-automatic tools, requiring manual adjustment of key points frame by frame, which is time-consuming, labor-intensive, and has poor repeatability.

[0005] This technical bottleneck severely restricts the efficiency and accuracy of disease-resistant drug screening, especially in high-throughput scenarios, where existing methods cannot meet the demands of both real-time performance and precision. Therefore, an innovative approach combining high-precision detection, fine-grained key point localization, and automated computation is urgently needed to overcome the technical barriers to phenotypic analysis of spore germination under microscopic conditions. Summary of the Invention

[0006] The purpose of the present invention is to solve the defect in the prior art that it is difficult to detect germination phenotype spores using microscopic images, and to provide a method for detecting the germination phenotype of wheat fusarium spores based on microscopic images to solve the above problem.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for detecting the germination phenotype of wheat scab spores based on microscopic images comprises the following steps:

[0009] Acquisition and preprocessing of wheat scab spore microscopic images: Acquisition of wheat scab spore microscopic images, including images of germinated and ungerminated spores, and images of unilateral or bilateral germination phenotypes, and preprocessing of the microscopic images. Preprocessing includes annotation, data augmentation, and dataset partitioning. Annotation content includes spore categories and key points.

[0010] Construction of a model for detecting key points of germination phenotypic targets: Based on the original backbone network of the YOLOv11 pose estimation model, a germination convolution cascade module driven by recursive gated convolution and a germination recursive gating module are introduced to construct a model for detecting key points of germination phenotypic targets.

[0011] Training of the germination phenotype target key point detection model: The preprocessed microscopic images are input into the germination phenotype target key point detection model for end-to-end training. This model simultaneously optimizes target detection and key point localization tasks, and outputs spore germination status classification, detection frame coordinates, and key points at the end of the germination tube.

[0012] Acquisition and preprocessing of drug-treated spore images: Microscopic images of wheat scab spores treated with different concentrations of drugs were obtained. Image cleaning, annotation, and data enhancement were performed according to unified experimental standards to construct a test dataset covering multiple drug-concentration gradients.

[0013] Output of spore detection results: The test dataset is input into the trained germination phenotype target key point detection model, and the spore detection box and key point coordinates are output simultaneously.

[0014] The construction of the germination phenotype target key point detection model includes the following steps:

[0015] Set up the germination convolution cascade module:

[0016] The germination convolution cascade module is set to generate a set of soft gating vectors using 1×1 point-by-point convolution, and the input tensor is recalibrated channel by channel, thereby completing lightweight feature screening before entering the convolution operation. Then, λ-order 3×3 depth-wise separable convolutions are stacked in series, where λ is the order. Each order of convolution reads the output of the previous order and is dynamically modulated by the same gating coefficient, thereby recursively accumulating information within a single layer and explicitly arranging the channel and spatial coupling relationship. The outputs of each order are normalized according to the factorial weight and fused uniformly at the end using 1×1 convolution, so that the information of different orders is written back to the main branch in the form of residuals.

[0017] When λ=1, it degenerates into ordinary depth-wise separable convolution; when λ>1, high-order terms similar to the Volterra series are introduced to capture second-order and third-order cross-channel and cross-space interactions within a single layer of the network; high-order convolution adopts a depth-wise separable form and shares a lightweight gating vector;

[0018] Set up the germination recursive gating module;

[0019] A germination phenotype target key point detection model is set up, including an input module, a backbone network, a neck network, and a predicted head. The input module performs image resolution unification to 960×960, performs noise suppression and contrast enhancement, integrates random rotation, scale perturbation, and gamma jitter augmentation strategies, generates multi-task training samples, and encapsulates the data into a batch of 8 tensor flows after normalization. The backbone network is constructed by alternating stacking germination convolution cascade modules and germination recursive gating modules. The shallow layer focuses on high-frequency textures and deeply integrates long-range contexts to output the third-level feature map P3, the fourth-level feature map P4, and the fifth-level feature map P5. The neck network bidirectionally fuses multi-scale features through the path aggregation network-feature pyramid network, and embeds the germination bottleneck module to achieve cross-scale recursive aggregation and void diffusion alignment. The predicted head performs three-branch collaborative output: a bounding box branch, a classification branch, and a key point branch. The multi-task gradient is dynamically balanced through the germination loss function. The bounding box branch is coordinate regression, the classification branch is germination state discrimination, and the key point branch is sub-pixel positioning.

[0020] Input module settings: All microscopic images were resampled to 960 × 960 using minimum edge padding, and contrast was enhanced using 3 × 3 bilateral filtering for denoising and contrast-constrained adaptive histogram equalization.

[0021] Setting the backbone network: The backbone network adopts a staged hierarchical architecture. The first two stages are connected in series twice through a 3×3 germination convolution cascade module to extract the high-frequency details of the spore wall texture and germination cracks; the subsequent three stages switch to the germination recursive gating module, with a void rate of 2 j-1 An exponentially increasing dilation strategy expands the receptive field, and a channel attention mechanism is embedded to focus on key areas. Three-scale feature maps, P3, P4, and P5, are output. The channel width is strictly aligned with the YOLOv11 pose estimation model baseline, and the residual connection retains the original skip-layer structure.

[0022] Setting up the neck network: The three-scale feature maps P3, P4, and P5 output by the backbone network are bidirectionally cascaded from top to bottom through a feature pyramid network and a path aggregation network from bottom to top; the upstream path uses nearest neighbor upsampling and 1×1 channel compression with a learnable scaling factor, and the downstream path uses depthwise separable 3×3 convolution to inject positioning signals; each splicing node embeds a germination bottleneck module, combining channel attention, void convolution, and deformable offset to achieve cross-scale feature balance;

[0023] In the key convergence layers of the backbone and neck networks, the original third-layer module is replaced with a budding convolutional cascade module, allowing features to simultaneously perform gating screening, recursive interaction, and multi-stage fusion when flowing through this layer. A budding recursive gating module is embedded in the deep semantic segments of the backbone network and the cross-scale fusion nodes of the neck network. This aggregates long-range contextual semantic information through recursive gating, optimizing the input feature quality of the downstream detection head.

[0024] Set the predicted head and loss function: input the multi-scale enhanced features output by the neck network into the predicted head, and the key point branch outputs the key point coordinates The germination loss function is used to calculate the key point error, and the effective gradient is screened by visibility marking. Error backpropagation is performed only on visible key points, and gradients are accumulated only for key points with v=1. The multi-task loss is adaptively normalized according to the gradient amplitude through a dynamic equalizer. The bounding box branch adopts the generalized intersection-over-union loss, and the classification branch adopts the focal loss. To reduce the memory usage, the germination bottleneck module in the prediction head truncates the gated vector gradient after forward calculation to achieve lightweight training.

[0025] The training of the germination phenotype target key point detection model includes the following steps:

[0026] Input module processing: The preprocessed microscopic image is input to the model, and bilateral filtering denoising, contrast-limited adaptive histogram equalization and contrast enhancement are performed. Random rotation, scale perturbation, and gamma dithering are applied for data augmentation. The data is normalized and packaged into a tensor flow with a batch size of 8.

[0027] Backbone network feature extraction: Tensors flow into the backbone network for feature interaction and aggregation processing;

[0028] Neck network feature fusion:

[0029] Multi-scale feature maps enter the neck network and are bidirectionally cascaded from bottom to top through a path aggregation network and a feature pyramid network. The upstream path uses nearest neighbor upsampling and 1×1 channel compression, while the downstream path uses depthwise separable 3×3 convolution to inject positioning signals. A germination bottleneck module is embedded in the feature splicing node, combining channel attention, void convolution, and deformable offset to optimize cross-scale alignment through recursive gating to aggregate long-range context.

[0030] Predicting head multi-task output: Fusion feature input predicts the head;

[0031] End-to-end backpropagation: The backbone network, neck network, and predicted head jointly update parameters: the germination loss function is used to simultaneously optimize the target detection and key point localization tasks, and output the spore category, detection box coordinates, and key points at the end of the germination tube.

[0032] The step of setting the germination recursive gating module comprises the following steps:

[0033] The germination recursive gating module is set to perform global average pooling on the input tensor to capture long-range contextual statistical summaries. The gating vector is generated by a branch consisting of a lightweight 1×1 convolutional layer, a batch normalization layer, and a sigmoid activation in series, and the original features are recalibrated channel by channel.

[0034] The recalibrated features are written into a convolutional "bottle" consisting of a stack of K-level recurrent gating units. Each unit level contains a 3×3 depthwise separable convolution and a residual bypass. The gating vector is cyclically reused as a shared coefficient in the recurrent units of the germinating recurrent gating module, driving each level of convolution to dynamically enhance important channels and suppress redundant information.

[0035] In the germinal recursive gating module, the depth convolution of the j-th level recurrent unit adopts a dilation rate of 2 j-1 The dilated convolution, where j is the level number,

[0036] Right now:

[0037] Level 1: Expansion rate 2 0 =1, covering local details;

[0038] Level 2: Expansion rate 2 1 =2, perceiving mid-range structure;

[0039] Level 3: Expansion rate 2 2 =4, capture global contour;

[0040] The outputs of each convolution are gated and superimposed to the bottle accumulation state M. Finally, 1×1 fused convolution is used to compress M back to the original channel dimension and merged with the pixel-by-pixel residual of the input feature.

[0041] The germination bottleneck module compresses deep features into dense discriminative representations for use by subsequent positioning branches through global gating constraints, recursive stack expansion of field of view, void convolution diffusion of information and residual steady-state maintenance.

[0042] The backbone network feature extraction comprises the following steps:

[0043] The shallow layer is connected twice in series through the germinal convolution cascade module to input feature F∈R C×H×W Perform interactive operations: split into the main branch F through 1×1 point-by-point convolution m Interaction branch F g , each occupying C / 2 channels, the interactive branch enters the n-th order recursive gated convolution process, where the t-th order is executed in sequence: channel compression, that is, 1×1 convolution is reduced to C / n, with expansion rate d t Depthwise separable convolution extracts the receptive field R t , linear mapping output γ t ,

[0044] All order outputs are scaled by a learnable factor λ t Linear combination and activation by σ(·) to generate the gated coefficient map

[0045] Among them, Y t Represents the output features of the t-th order recursive gated convolution, generated by channel compression, depth-wise separable convolution and linear mapping, and its dimension is the same as the interaction branch F g Consistent, λ t Represents a learnable scaling factor, dynamically adjusting the t-th order output Y t The contribution weight of , σ represents the activation function, which compresses the weighted sum to the interval [0, 1] to generate the gating weight, and n represents the total order of recursive gated convolution, which controls the diversity of the receptive field;

[0046] Main branch F m G k Residual back-injection after element-wise modulation: in is element-wise multiplication;

[0047] Deep switch to the germination recursive gating module, perform recursive gating aggregation on feature F1: split into F m With F g , the gated branch generates the gated tensor G through k=2 layers of recursive stack m , modulate the main branch to get Final output F2 = ΔF + F1;

[0048] Synchronize with the dilation rate of the dilated convolution to 2 j-1 The receptive field is expanded exponentially and multi-scale feature maps P3, P4, and P5 are output.

[0049] The method of predicting the head multi-task output includes the following steps:

[0050] The bounding box branch regresses the detection frame coordinates, the classification branch discriminates the spore germination state, and the key point branch outputs the key point coordinates at the end of the germination tube. The germination loss function dynamically optimizes key point positioning:

[0051] Coordinate normalization and error calculation: The coordinates of the key points predicted by the network are normalized. The manually annotated key point coordinates (x, y, v) are divided by the input image size W and H respectively for normalization. The calculation formula is as follows:

[0052]

[0053] in, is the network predicted coordinate, (x, y) is the manually marked coordinate, W and H are the width and height of the input image, Δx is the horizontal coordinate offset, and Δy is the vertical coordinate offset;

[0054] Piecewise hybrid loss function design:

[0055] In order to take into account both high-resolution sensitivity to small errors and robustness to large errors, a logarithmic-linear hybrid loss L(e) is adopted, which is defined as follows:

[0056] L(e)=log(1+μe), when e≤δ;

[0057] L(e)=log(1+μδ)+μ(e-δ) / (1+μδ), when e>δ,

[0058] Among them, e is the normalized error, δ is the turning radius, and μ is the hyperparameter for adjusting the curvature;

[0059] Set visibility mask v∈{0,1} and filter gradient: L=Σv·L(e);

[0060] Dynamic recalibration of scaling factors:

[0061] In order to adapt to the error size at different training stages, a scaling factor α_t is designed:

[0062]

[0063] in, is the exponential sliding mean of the key point loss of the current batch, ε is a small constant to prevent division by zero, and C is the initial scale constant;

[0064] Loss Normalization and Multi-Task Weighting:

[0065] The total keypoint loss L_kp is normalized by the number of visible keypoints N_v:

[0066] L_kp^norm=α_t×L_kp / (N_v+ε),

[0067] Among them, N v is the number of visible points; at the same time, the gated vector gradient is truncated to reduce the memory usage, w kp is the multi-task weight, L_kp^norm is the normalized total loss of key points,

[0068] The normalized loss is combined with other task losses according to preset weights to ensure stable training.

[0069] A method for detecting the germination phenotype of wheat fusarium spores based on microscopic images. A computer program is stored on the storage medium. When the computer program is executed by a processor, the method for detecting the germination phenotype of wheat fusarium spores based on microscopic images is implemented.

[0070] A computer device comprises a memory, a processor and a computer program stored in the memory and runnable on the processor. When the computer program is executed by the processor, a method for detecting the germination phenotype of wheat fusarium spores based on microscopic images is realized.

[0071] Beneficial effects

[0072] Compared with the existing technology, the microscopic image-based wheat fusarium spore germination phenotype detection method of the present invention accurately captures the sub-pixel details of spores and the global outline of the germination tube in the microscopic image through the germination convolution cascade module (high-order cross-channel / spatial modeling) and the germination recursive gating module (exponential void diffusion + channel attention), solves the problem of missed detection and false detection of small targets, and realizes the precise detection of germination phenotypic spores in microscopic images.

[0073] The present invention adopts an end-to-end unified architecture to synchronously output the detection frame, germination status and key point coordinates, and combines the germination loss function (segmented mixed loss + dynamic normalization) to achieve multi-task gradient balance, significantly improving the sub-pixel positioning robustness; relying on bilateral filtering and adaptive enhancement preprocessing, cross-scale deformable alignment neck network and long-range context recursive gating, it can strongly adapt to complex scenes such as illumination noise, spore overlap and morphological variation; through gated gradient truncation and depthwise separable convolution, it reduces video memory consumption, and uses void diffusion strategy and multi-agent data enhancement to ensure efficient model training and strong generalization, greatly improving the efficiency of disease resistance agent screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a method sequence diagram of the present invention;

[0075] Figure 2 This is a schematic diagram of the structure of the germination convolution cascade module involved in the present invention;

[0076] Figure 3 This is a schematic diagram of the structure of the germination recursive gating module involved in the present invention;

[0077] Figure 4 This is a microscopic image of wheat fusarium spore germination;

[0078] Figure 5 This is a schematic diagram of marking a microscopic image model of wheat scab spore germination;

[0079] Figure 6 Schematic diagram of the model for marking the key points of germination of a single wheat scab spore;

[0080] Figure 7 For the traditional YOLOv11 pose estimation model Figure 4 Detection effect diagram;

[0081] Figure 8 The method of the present invention is Figure 4 Detection effect diagram;

[0082] Figure 9 This is another microscopic image of wheat fusarium spore germination;

[0083] Figure 10 for Figure 9 The corresponding attention heat map visualization result diagram. DETAILED DESCRIPTION

[0084] In order to provide a further understanding and appreciation of the structural features and effects achieved by the present invention, a detailed description is provided with reference to preferred embodiments and accompanying drawings as follows:

[0085] like Figure 1 As shown, the method for detecting the germination phenotype of wheat scab spores based on microscopic images of the present invention comprises the following steps:

[0086] The first step is to collect and preprocess microscopic images of wheat fusarium spores: collect microscopic images of wheat fusarium spores, including images of germinated and ungerminated spores, and images of unilateral or bilateral germination phenotypes, and preprocess the microscopic images. The preprocessing includes annotation, data enhancement, and dataset partitioning. The annotation content includes spore categories and key points.

[0087] The second step is to construct a germination phenotype target key point detection model: based on the original backbone network of the YOLOv11 posture estimation model, the germination convolution cascade module driven by recursive gated convolution and the germination recursive gating module are introduced to construct the germination phenotype target key point detection model.

[0088] Automated detection of germination phenotypes of wheat scab spores in microscopic images faces several key challenges that make direct use of existing models ineffective:

[0089] The targets are tiny and complex in shape: Spores themselves are extremely small (micrometer-scale) and have diverse germination phenotypes (ungerminated, unilaterally germinated, bilaterally germinated). The germination fissures and the ends of the germination tubes are sub-pixel fine structures, and conventional convolutional layers have difficulty effectively capturing their high-frequency details and spatial relationships. Background noise interference is severe: Microscopic images often have problems such as uneven lighting, impurities, complex background textures, and noise, which interfere with the recognition of tiny targets and the positioning of key points. Key point positioning accuracy requires extremely high precision: The positioning accuracy of the key point at the end of the germination tube directly affects the accuracy of the subsequent calculation of the germination length ratio. Sub-pixel positioning errors will be significantly amplified on tiny targets. There is a contradiction between receptive field and positioning accuracy: Understanding the overall morphology of the spores (especially the long-distance germination tubes) requires a large receptive field, but accurately locating key points requires retaining high-resolution details. Standard models are difficult to take into account simultaneously; multi-task collaborative optimization is difficult: it is necessary to complete target detection (spore position, category) and key point positioning (germination tube end) with high precision in one model. The two tasks have different characteristics (classification / regression, global / local), and gradients may conflict. Simple combinations can easily lose sight of one thing while focusing on another; computational efficiency constraints: high-throughput screening scenarios require efficient models, but the introduction of complex modules will increase the computational burden, and a balance must be struck between accuracy and efficiency.

[0090] To address the above difficulties, the core innovation here lies in deeply customizing the backbone network of the original YOLOv11 pose estimation model and introducing two key modules.

[0091] It includes the following steps:

[0092] (1) Setting up the germination convolution cascade module:

[0093] like Figure 2 As shown in the figure, the germinal convolution cascade module is set to generate a set of soft gating vectors with 1×1 point-by-point convolution, and the input tensor is recalibrated channel by channel to suppress the noise / background channels and focus on the information-rich channels. Thus, lightweight feature screening is completed before entering the convolution operation; followed by serial stacking of λ-order 3×3 depth-wise separable convolutions, where λ is the order. Each order of convolution reads the output of the previous order and is dynamically modulated by the same gating coefficient, thereby recursively accumulating information within a single layer and explicitly arranging the channel and spatial coupling relationship; the output of each order is normalized according to the factorial weight and fused uniformly at the end with 1×1 convolution, so that the information of different orders is written back to the main branch in the form of residuals;

[0094] When λ=1, it degenerates into ordinary depth-wise separable convolution; when λ>1, high-order terms approximating the Volterra series are introduced to capture second-order and third-order cross-channel and cross-space interactions within a single layer of the network; high-order convolution adopts a depth-wise separable form and shares a lightweight gating vector. After the outputs of each order are fused according to the weights, they are injected back into the main branch in the form of a residual, retaining the original information and injecting enhanced high-order features.

[0095] This module is the core of shallow feature extraction, specifically addressing the challenges of "small targets, critical details, and complex backgrounds." While extracting basic textures, it significantly enhances the model's perception of subtle but highly discriminative local features, such as spore boundaries and germination points, through recursion and high-order interactions. This reduces shallow information loss and lays the foundation for subsequent precise positioning.

[0096] (2) Setting up the germination recursive gating module:

[0097] like Figure 3 As shown in the figure, the germination recursive gating module is set to perform global average pooling on the input tensor to capture the long-range context statistical summary; the gating vector is generated by a branch consisting of a lightweight 1×1 convolution layer, a batch normalization layer, and a sigmoid activation in series, and the original features are recalibrated channel by channel;

[0098] The recalibrated features are written into a convolutional "bottle" consisting of a stack of K-level recurrent gating units. Each unit level contains a 3×3 depthwise separable convolution and a residual bypass. The gate vector is cyclically reused as a shared coefficient in the recurrent units of the germinating recurrent gating module. Each unit level contains a dilated convolution and a residual connection, driving each level of convolution to dynamically enhance important channels, achieve channel attention fusion, and suppress redundant information.

[0099] In the germination recursive gating module, the depth convolution of the j-th level recurrent unit adopts a dilation rate of 2 j-1 The dilated convolution, where j is the level number,

[0100] Right now:

[0101] Level 1: Expansion rate 2 0 =1, covering local details;

[0102] Level 2: Expansion rate 2 1 =2, perceiving mid-range structure;

[0103] Level 3: Expansion rate 2 2 =4, capture global contour;

[0104] The outputs of each convolution are gated and superimposed to the bottle accumulation state M. Finally, 1×1 fused convolution is used to compress M back to the original channel dimension and merged with the pixel-by-pixel residual of the input feature.

[0105] The germination bottleneck module compresses deep features into a dense discriminative representation through global gating constraints, recursive stacking to expand the field of view, dilated convolution to diffuse information, and residual steady-state maintenance. This is used by the subsequent localization branch. The final fusion result is concatenated with the input feature residual to ensure smooth information flow.

[0106] This module is the core of deep feature abstraction and contextual understanding, specifically addressing the challenges of large receptive fields, understanding overall structure, and distinguishing complex phenotypes. Through recursive gating to aggregate long-range dependencies and multi-scale perception using exponentially dilated convolutions, it effectively integrates information ranging from local spore details to global germ tube morphology and even the surrounding environment. This is crucial for distinguishing unilateral from bilateral germination, understanding germ tube orientation, and reliably locating terminal keypoints in complex backgrounds or with slight overlap.

[0107] (3) Setting up the key point detection model for the germination phenotype target,

[0108] The model consists of an input module, a backbone network, a neck network, and a predicted head. The input module unifies the image resolution to 960×960, striking a balance between retaining sufficient spore detail information (avoiding excessive loss of sub-pixel features due to downsampling) and computational efficiency. This provides a spatial basis for subsequent high-precision detection and key point positioning, performs noise suppression and contrast enhancement, and integrates random rotation, scale perturbation, and gamma jitter augmentation strategies to simulate the natural variation in spore posture, size, and imaging brightness in actual microscopic observations. This greatly enhances the model's generalization ability to posture changes, scale differences, and fluctuations in imaging conditions, avoiding overfitting. Multi-task training samples are generated, and the data is normalized and packaged into a batch size of 8 tensor flows.

[0109] The backbone network is constructed by alternating stacks of germination convolutional cascade modules and germination recursive gating modules. It shallowly focuses on high-frequency textures and deeply integrates long-range context, outputting third-level feature maps P3, fourth-level feature maps P4, and fifth-level feature maps P5. This network significantly enhances the perception of high-frequency details and local complex patterns, such as the spore outline, cell wall texture, and, in particular, the germination initiation point (fissure), within the high-resolution feature map (P3). This is key to addressing missed and false detections of microspores and accurately identifying the germination initiation state. It efficiently aggregates long-range context within the medium- and low-resolution feature maps (P4 and P5). This enables the model to understand the overall orientation of the germ tube (especially its long-distance extension), the relative positional relationships between spores, and global phenotypic features such as distinguishing unilateral from bilateral germination. This provides semantic support for the precise location of key points (the terminus of the germ tube). The network provides features at different levels of abstraction, ranging from fine local details to global semantics, paving the way for effective cross-scale fusion within the neck network, ensuring that spores of varying sizes can be effectively detected and characterized.

[0110] The neck network bidirectionally fuses multi-scale features through the path aggregation network-feature pyramid network, effectively integrating the multi-scale features output by the backbone network, ensuring that shallow high-resolution features contain rich semantic information, and deep low-resolution features are integrated into precise positioning signals. The embedded germination bottleneck module realizes cross-scale recursive aggregation and void diffusion alignment, which significantly improves the quality of the fused features, making the features received by the predicted head highly aligned in space and semantics, providing more reliable and consistent input for subsequent detection box regression, classification and key point positioning.

[0111] The prediction head performs three-branch collaborative output: bounding box branch, classification branch, and key point branch. The multi-task gradient is dynamically balanced through the germination loss function. The bounding box branch is coordinate regression, the classification branch is germination state discrimination, and the key point branch is sub-pixel positioning.

[0112] (4) Input module settings: All microscopic images were resampled to 960 × 960 using minimum edge padding, and contrast was enhanced using 3 × 3 bilateral filtering for denoising and contrast-constrained adaptive histogram equalization.

[0113] (5) Setting the backbone network: The backbone network adopts a staged hierarchical architecture. The first two stages are connected in series twice through a 3×3 germination convolution cascade module to extract the high-frequency details of the spore wall texture and germination cracks; the subsequent three stages switch to the germination recursive gating module with a void rate of 2. j-1 An exponentially increasing hole diffusion strategy expands the receptive field, and a channel attention mechanism is embedded to focus on key areas. Three-scale feature maps, P3, P4, and P5, are output. The channel width is strictly aligned with the baseline of the YOLOv11 pose estimation model, and the residual connection retains the original skip-layer structure.

[0114] (6) Setting the neck network: The three-scale feature maps P3, P4, and P5 output by the backbone network are bidirectionally cascaded from top to bottom through the feature pyramid network and the path aggregation network from bottom to top; the upstream path uses the nearest neighbor upsampling and 1×1 channel compression with a learnable scaling factor, and the downstream path uses a depth-separable 3×3 convolution to inject the positioning signal; each splicing node is embedded with a germination bottleneck module, and the cross-scale feature balance is achieved by combining channel attention, void convolution and deformable offset.

[0115] (7) In the key convergence layer of the backbone network and the neck network, the original third-layer module is replaced with a germination convolution cascade module, so that the triple operations of gating screening, recursive interaction and multi-stage fusion are performed simultaneously when the features flow through this layer; in the deep semantic segment of the backbone network and the cross-scale fusion node of the neck network, a germination recursive gating module is embedded, which aggregates long-range contextual semantic information through recursive gating to optimize the input feature quality of the downstream detection head.

[0116] (8) Set the prediction head and loss function: Input the multi-scale enhanced features output by the neck network into the prediction head, and the key point branch outputs the key point coordinates The germination loss function is used to calculate the key point error, and the effective gradient is screened by visibility marking. Error backpropagation is performed only on visible key points, and gradients are accumulated only for key points with v=1. The multi-task loss is adaptively normalized according to the gradient amplitude through a dynamic equalizer. The bounding box branch adopts the generalized intersection-over-union loss, and the classification branch adopts the focal loss. To reduce the memory usage, the germination bottleneck module in the prediction head truncates the gated vector gradient after forward calculation to achieve lightweight training.

[0117] The third step is to train the germination phenotype target key point detection model: the preprocessed microscopic images are input into the germination phenotype target key point detection model for end-to-end training, and the target detection and key point positioning tasks are simultaneously optimized to output the spore germination status classification, detection frame coordinates, and key points at the end of the germination tube.

[0118] Training an end-to-end, high-accuracy spore germination phenotype keypoint detection model on microscopic images faces unique challenges that go beyond general object detection or pose estimation models:

[0119] First, task coupling and gradient conflicts arise: Three highly related but distinct tasks must be optimized simultaneously: object detection (position regression), classification (germination state), and keypoint localization (sub-pixel regression). Object detection and classification tasks are relatively macroscopic, potentially leading to larger gradients; whereas keypoint localization involves extremely detailed sub-pixel regression, resulting in smaller and more sensitive gradients. Simply summing the loss functions leads to gradient conflicts: the macroscopic task may dominate the optimization direction, suppressing the convergence of the keypoint localization task, resulting in inaccurate keypoint localization at the end, directly impacting the accuracy of the germinal tube length ratio calculation. Dynamically balancing the gradients of these multiple tasks is a key challenge.

[0120] Secondly, keypoint positioning is extremely sensitive: the end of the germinal tube is a sub-pixel target and may only occupy a few pixels in a 960x960 image. Conventional regression losses (such as L1 / L2) are insensitive to small errors (a positioning offset of 1-2 pixels is negligible in macroscopic detection, but has a significant impact on the calculation of the germinal tube length ratio) and are not robust enough to large errors (susceptible to outliers). It is extremely difficult to design a keypoint loss function that is highly sensitive to small errors, robust to large errors, and adaptable to different stages of training (large errors in the early stage and small errors in the later stage).

[0121] Finally, keypoint visibility and annotation noise: The end of the germinal tube may be occluded, blurred, or located outside the field of view, and may be visible (v=1) or invisible (v=0) during annotation. Invisible keypoints should not contribute to the gradient, otherwise they will introduce noise and mislead the model. An accurate visibility mask mechanism is required.

[0122] (1) Input module processing: The preprocessed microscopic image is input to the model, and bilateral filtering denoising and contrast-limited adaptive histogram equalization are performed to enhance the contrast. Random rotation, scale perturbation, and gamma dithering are applied for data enhancement. After data normalization, it is packaged into a tensor flow with a batch size of 8.

[0123] (2) Backbone network feature extraction: Tensors flow into the backbone network for feature interaction and aggregation processing;

[0124] A1) The shallow layer is connected twice in series through the germinal convolution cascade module to input feature F∈R C×H×W Perform interactive operations: split into the main branch F through 1×1 point-by-point convolution m Interaction branch F g , each occupying C / 2 channels, the interactive branch enters the n-th order recursive gated convolution process, where the t-th order is executed in sequence: channel compression, that is, 1×1 convolution is reduced to C / n, with expansion rate d t Depthwise separable convolution extracts the receptive field R t , linear mapping output γ t ,

[0125] All order outputs are scaled by a learnable factor λ t Linear combination and activation by σ(·) to generate the gated coefficient map

[0126] Among them, Y t Represents the output features of the t-th order recursive gated convolution, generated by channel compression, depth-wise separable convolution and linear mapping, and its dimension is the same as the interaction branch F g Consistent, λ t Represents a learnable scaling factor, dynamically adjusting the t-th order output Y t The contribution weight of , σ represents the activation function, which compresses the weighted sum to the interval [0, 1] to generate the gating weight, and n represents the total order of recursive gated convolution, which controls the diversity of the receptive field;

[0127] Main branch F m G k Residual back-injection after element-wise modulation: in is element-wise multiplication;

[0128] B2) Deeply switch to the germination recursive gating module and perform recursive gating aggregation on feature F1: split into F m With F g , the gated branch generates the gated tensor G through k=2 layers of recursive stack m , modulate the main branch to get Final output F2 = ΔF + F1;

[0129] B3) Synchronously use the dilation rate of the dilated convolution to 2 j-1 The receptive field is expanded exponentially and multi-scale feature maps P3, P4, and P5 are output.

[0130] (3) Neck network feature fusion:

[0131] Multi-scale feature maps enter the neck network and are bidirectionally cascaded from bottom to top through the path aggregation network and the feature pyramid network from top to bottom. The upstream path uses nearest neighbor upsampling and 1×1 channel compression, and the downstream path uses depth-wise separable 3×3 convolution to inject positioning signals. A germination bottleneck module is embedded in the feature splicing node, combining channel attention, void convolution and deformable offset, and optimizing cross-scale alignment through recursive gating to aggregate long-range context.

[0132] (4) Predicting head multi-task output: Fusion feature input predicts the head;

[0133] C1) The bounding box branch regresses the detection box coordinates, the classification branch discriminates the spore germination state, and the key point branch outputs the coordinates of the key points at the end of the germination tube. The germination loss function dynamically optimizes key point positioning:

[0134] C2) Coordinate normalization and error calculation: The coordinates of the key points predicted by the network are normalized. The manually annotated key point coordinates (x, y, v) are divided by the input image size W and H respectively for normalization. The calculation formula is as follows:

[0135] in, is the network predicted coordinate, (x, y) is the manually marked coordinate, W and H are the width and height of the input image, Δx is the horizontal coordinate offset, and Δy is the vertical coordinate offset;

[0136] C3) Piecewise hybrid loss function design:

[0137] In order to take into account both high-resolution sensitivity to small errors and robustness to large errors, a logarithmic-linear hybrid loss L(e) is adopted, which is defined as follows:

[0138] L(e)=log(1+μe), when e≤δ;

[0139] L(e)=log(1+μδ)+μ(e-δ) / (1+μδ), when e>δ,

[0140] Where e is the normalized error, δ is the turning radius, and μ is a hyperparameter for adjusting the curvature. δ and μ are carefully set according to the spore size and positioning accuracy requirements to ensure that the benefits of the logarithmic term are maximized within the typical error range.

[0141] C4) Set visibility mask v∈{0,1} and filter gradient: L=Σv·L(e);

[0142] C5) Dynamic recalibration of scaling factor: To adapt to the error size at different training stages, a scaling factor α_t is designed:

[0143]

[0144] in, is the exponential sliding mean of the key point loss of the current batch, ε is a small constant to prevent division by zero, and C is the initial scale constant in the early training ( Large): α_t automatically becomes smaller, reducing the weight of key point loss to prevent large errors from dominating the training and causing instability. In the later stage of training ( Small): α_t automatically increases, increasing the weight of key point loss. When the error is already small, the optimization efforts are increased to pursue sub-pixel accuracy. This achieves adaptive "speed change" in the training process, significantly accelerating convergence and improving final accuracy.

[0145] C6) Loss Normalization and Multi-Task Weighting:

[0146] The total keypoint loss L_kp is normalized by the number of visible keypoints N_v:

[0147] L_kp^norm=α_t×L_kp / (N_v+ε),

[0148] Among them, N v is the number of visible points; at the same time, the gated vector gradient is truncated to reduce the memory usage, w kp is the multi-task weight, and L_kp^norm is the normalized total keypoint loss, which is divided by the total number of visible keypoints in the current batch, N_v. This ensures that the loss is independent of the number of spores in the image, is comparable across batches, and has more stable gradients.

[0149] (5) End-to-end backpropagation: The backbone network, neck network, and predicted head network jointly update parameters: the germination loss function is used to simultaneously optimize the target detection and key point localization tasks, outputting the spore category, detection frame coordinates, and key points at the end of the germination tube. The three tasks share the rich features extracted from the backbone and neck, avoiding the error accumulation and context information loss problems of the staged model (first detecting spores and then locating key points on the cropped image).

[0150] The fourth step is the collection and pre-processing of images of spores treated with chemicals: microscopic images of wheat scab spores treated with chemicals of different concentrations are obtained. Figure 4 、 Figure 9 As shown in the figure, the image is cleaned and labeled according to the unified experimental standards, such as Figure 5 、 Figure 6As shown, the boxes, key points, and group IDs are annotated, and data enhancement is performed to construct a test dataset covering multiple drug concentration gradients;

[0151] Step 5: Output of spore detection results: Input the test dataset into the trained germination phenotype target key point detection model, and simultaneously output the spore detection box and key point coordinates.

[0152] like Figure 8 As shown, it is the method of the present invention Figure 4 The detection effect diagram, from Figure 8 It can be seen that the detection results of the model in complex real microscopic scenes (including dense spores, background noise, uneven lighting, and morphological variation) are as follows: colored detection frames, key point marks at the end of the germination tube (red, yellow, green, and white dots), confidence levels, and category labels (all in the germination state). All visible spores in the figure are accurately detected and framed without omissions. Background noise, impurities, etc. are not mistakenly identified as spores, solving the problem of "missed detection of small targets and false detection of background interference". The model correctly distinguishes between ungerminated spores and germinated spores, achieving strong discrimination of complex germination phenotypes. The key point at the end of the germination tube is precisely located at the tip of the germination tube, even for curved and slender germination tubes. The point position is highly consistent with naked eye observation or expert annotation expectations, and sub-pixel accuracy is reflected, demonstrating that a single forward propagation synchronously outputs all detection frames, categories, and key points without any traces of staged processing, achieving efficient analysis. Figure 7 Traditional methods may cause false detection and key points may be offset. Figure 10 As shown, it is Figure 9 The corresponding attention heat map visualization result diagram shows that the model is processing Figure 9 When focusing on spore images, the warmer the heatmap color, the greater the contribution of that region to the model's decision (classification or keypoint localization). Strongly warm regions are highly concentrated in the spore body (especially the cell wall and germination point) and the germ tube (especially its growth direction and terminal end). Background areas are mostly cool, indicating that the model successfully suppresses irrelevant noise. The terminal region of the germ tube often exhibits a significant hotspot, which intuitively explains why the model is able to accurately locate this location: it is identified as the most discriminative location.

[0153] The final output (visualization see Figure 8 、 Figure 10 ) is the concentrated embodiment and ultimate verification of all the technological innovations previously described in this invention. It not only enables automated, high-precision, and efficient analysis of wheat scab spore germination phenotypes, but also directly supports high-throughput screening of disease-resistant agents through the reliable data it outputs (particularly the coordinates of key points), possessing significant practical value and promising prospects for widespread adoption.

[0154] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the germination phenotype of wheat scab spores based on microscopic images, characterized in that: The following steps are involved: 11) Collection and preprocessing of wheat scab spore microscopic images: Collect microscopic images of wheat scab spores, including images of germinated and ungerminated spores, and images of unilateral or bilateral germination phenotypes. Preprocess the microscopic images, including annotation, data augmentation, and dataset partitioning. Annotation includes spore types and key points. 12) Construction of a model for detecting key points of germination phenotype targets: Based on the original backbone network of the YOLOv11 pose estimation model, a germination convolution cascade module driven by recursive gated convolution and a germination recursive gating module are introduced to construct a model for detecting key points of germination phenotype targets; 13) Training of the germination phenotype target key point detection model: The preprocessed microscopic images are input into the germination phenotype target key point detection model for end-to-end training, which simultaneously optimizes the target detection and key point positioning tasks, and outputs the spore germination state classification, detection box coordinates, and key points at the end of the germination tube; 14) Acquisition and preprocessing of chemical-treated spore images: Microscopic images of wheat scab spores treated with chemical agents at different concentrations were obtained. Image cleaning, annotation, and data enhancement were performed according to unified experimental standards to construct a test dataset covering multiple chemical-concentration gradients. 15) Output of spore detection results: The test dataset is input into the trained germination phenotype target key point detection model, and the spore detection box and key point coordinates are output simultaneously.

2. A method for detecting wheat scab spore germination phenotype based on microscopic images according to claim 1, characterized in that: The construction of the germination phenotype target key point detection model includes the following steps: 21) Set up the germination convolution cascade module: The germination convolution cascade module is set to generate a set of soft gating vectors using 1×1 point-by-point convolution, and the input tensor is recalibrated channel by channel, thereby completing lightweight feature screening before entering the convolution operation. Then, λ-order 3×3 depth-wise separable convolutions are stacked in series, where λ is the order. Each order of convolution reads the output of the previous order and is dynamically modulated by the same gating coefficient, thereby recursively accumulating information within a single layer and explicitly arranging the channel and spatial coupling relationship. The outputs of each order are normalized according to the factorial weight and fused uniformly at the end using 1×1 convolution, so that the information of different orders is written back to the main branch in the form of residuals. When λ=1, it degenerates into ordinary depth-wise separable convolution; when λ>1, high-order terms similar to the Volterra series are introduced to capture second-order and third-order cross-channel and cross-space interactions within a single layer of the network; high-order convolution adopts a depth-wise separable form and shares a lightweight gating vector; 22) Setting up the germination recursive gating module; 23) Set up a germination phenotype target key point detection model, including an input module, a backbone network, a neck network, and a predicted head. The input module performs image resolution unification to 960×960, performs noise suppression and contrast enhancement, integrates random rotation, scale perturbation, and gamma jitter augmentation strategies, generates multi-task training samples, and encapsulates the data into a batch of 8 tensor flows after normalization; the backbone network is constructed by alternating stacking germination convolution cascade modules and germination recursive gating modules, focusing on high-frequency textures in the shallow layer and integrating long-range context in the deep layer, outputting the third-level feature map P3, the fourth-level feature map P4, and the fifth-level feature map P5; the neck network bidirectionally fuses multi-scale features through the path aggregation network-feature pyramid network, and embeds the germination bottleneck module to achieve cross-scale recursive aggregation and void diffusion alignment; the predicted head performs three-branch collaborative output: bounding box branch, classification branch, and key point branch. The multi-task gradient is dynamically balanced through the germination loss function, where the bounding box branch is coordinate regression, the classification branch is germination state discrimination, and the key point branch is sub-pixel positioning; 24) Input module settings: All microscopic images were resampled to 960 × 960 using minimum edge padding, and contrast was enhanced using 3 × 3 bilateral filtering for denoising and contrast-constrained adaptive histogram equalization. 25) Setting the backbone network: The backbone network adopts a staged hierarchical architecture. The first two stages are connected in series twice through a 3×3 germination convolution cascade module to extract the high-frequency details of the spore wall texture and germination cracks; the subsequent three stages switch to the germination recursive gating module, with a void rate of 2 j-1 An exponentially increasing dilation strategy expands the receptive field, and a channel attention mechanism is embedded to focus on key areas. Three-scale feature maps, P3, P4, and P5, are output. The channel width is strictly aligned with the YOLOv11 pose estimation model baseline, and the residual connection retains the original skip-layer structure. 26) Set up the neck network: The three-scale feature maps P3, P4, and P5 output by the backbone network are bidirectionally cascaded from top to bottom through the feature pyramid network and the path aggregation network from bottom to top; the upstream path uses nearest neighbor upsampling and 1×1 channel compression with a learnable scaling factor, and the downstream path uses depthwise separable 3×3 convolution to inject positioning signals; each splicing node embeds a germination bottleneck module, combining channel attention, void convolution, and deformable offset to achieve cross-scale feature balance; 27) In the key convergence layer of the backbone network and the neck network, the original third layer module is replaced with a germination convolution cascade module, so that the features flow through this layer and perform the triple operations of gated screening, recursive interaction, and multi-stage fusion simultaneously; in the deep semantic segment of the backbone network and the cross-scale fusion node of the neck network, a germination recursive gating module is embedded to aggregate long-range contextual semantic information through recursive gating, thereby optimizing the input feature quality of the downstream detection head; 28) Set the predicted head and loss function: input the multi-scale enhanced features output by the neck network into the predicted head, and the key point branch outputs the key point coordinates The germination loss function is used to calculate the key point error, the effective gradient is filtered by visibility marking, the error is only back-propagated for visible key points, and the gradient is only accumulated for the key point with v=1; the multi-task loss is adaptively normalized according to the gradient amplitude through a dynamic equalizer, the bounding box branch adopts the generalized intersection-over-union loss, and the classification branch adopts the focal loss; to reduce the video memory usage, the germination bottleneck module in the prediction head truncates the gated vector gradient after forward calculation to achieve lightweight training.

3. A method for detecting wheat scab spore germination phenotype based on microscopic images according to claim 1, characterized in that: The training of the germination phenotype target key point detection model includes the following steps: 31) Input module processing: The pre-processed microscopic image is input to the model, and bilateral filtering denoising, contrast-limited adaptive histogram equalization and contrast enhancement are performed. Random rotation, scale perturbation, and gamma dithering are applied for data enhancement. The data is normalized and packaged into a tensor flow with a batch size of 8. 32) Backbone network feature extraction: Tensors flow into the backbone network for feature interaction and aggregation processing; 33) Neck network feature fusion: Multi-scale feature maps enter the neck network and are bidirectionally cascaded from bottom to top through a path aggregation network and a feature pyramid network. The upstream path uses nearest neighbor upsampling and 1×1 channel compression, while the downstream path uses depthwise separable 3×3 convolution to inject positioning signals. A germination bottleneck module is embedded in the feature splicing node, combining channel attention, void convolution, and deformable offset to optimize cross-scale alignment through recursive gating to aggregate long-range context. 34) Predicting head multi-task output: Fusing feature input to predict the head; 35) End-to-end backpropagation: The backbone network, neck network, and predicted head jointly update parameters: the germination loss function is used to simultaneously optimize the target detection and key point localization tasks, and output the spore category, detection box coordinates, and key points at the end of the germination tube.

4. a kind of wheat scab spore germination phenotype detection method based on microscopic image according to claim 2, is characterized in that, The step of setting the germination recursive gating module comprises the following steps: 41) Set the germination recursive gating module to perform global average pooling on the input tensor to capture the long-range contextual statistical summary; generate the gating vector through a branch consisting of a lightweight 1×1 convolutional layer, a batch normalization layer, and a sigmoid activation in series, and recalibrate the original features channel by channel; The recalibrated features are written into a convolutional "bottle" consisting of a stack of K recurrent gating units. Each unit level contains a 3×3 depthwise separable convolution and a residual bypass. The gate vector is cyclically reused as a shared coefficient in the recurrent units of the germinating recurrent gating module, driving each level of convolution to dynamically enhance important channels and suppress redundant information. 42) In the germinal recursive gating module, the depth convolution of the j-th level recurrent unit adopts a dilation rate of 2 j-1 The dilated convolution, where j is the level number, Right now: Level 1: Expansion rate 2 0 =1, covering local details; Level 2: Expansion rate 2 1 =2, perceiving mid-range structure; Level 3: Expansion rate 2 2 =4, capturing the global outline; The outputs of each convolution are gated and superimposed to the bottle accumulation state M. Finally, 1×1 fused convolution is used to compress M back to the original channel dimension and merged with the pixel-by-pixel residual of the input feature. 43) The germination bottleneck module compresses deep features into dense discriminative representations for use by subsequent positioning branches through global gating constraints, recursive stacking to expand the field of view, dilated convolution to diffuse information, and residual steady-state maintenance.

5. A method for detecting wheat scab spore germination phenotype based on microscopic images according to claim 3, characterized in that: The backbone network feature extraction comprises the following steps: 51) The shallow layer is connected twice in series through the germinal convolution cascade module to input feature F∈R C×H×W Perform interactive operations: split into the main branch F through 1×1 point-by-point convolution m Interaction branch F g , each occupying C / 2 channels, the interactive branch enters the b-order recursive gated convolution process, where the t-th order performs the following steps: channel compression, i.e., 1×1 convolution is reduced to C / n, with expansion rate d t Depthwise separable convolution extracts the receptive field R t , linear mapping output γ t , All order outputs are scaled by a learnable factor λ t Linear combination and activation by σ(·) to generate the gated coefficient map Among them, Y t Represents the output features of the t-th order recursive gated convolution, generated by channel compression, depth-wise separable convolution and linear mapping, and its dimension is the same as the interaction branch F g Consistent, λ t Represents a learnable scaling factor, dynamically adjusting the t-th order output Y t The contribution weight of , σ represents the activation function, which compresses the weighted sum to the interval [0, 1] to generate the gating weight, and n represents the total order of recursive gated convolution, which controls the diversity of the receptive field; Main branch F m G k Residual back-injection after element-wise modulation: in is element-wise multiplication; 52) Deeply switch to the germination recursive gating module and perform recursive gating aggregation on feature F1: split into F m With F g , the gated branch generates the gated tensor G through k=2 layers of recursive stack m , modulate the main branch to get Final output F2 = ΔF + F1; 53) Synchronize with the dilation rate of the dilated convolution at 2 j-1 The receptive field is expanded exponentially and multi-scale feature maps P3, P4, and P5 are output.

6. A method for detecting wheat scab spore germination phenotype based on microscopic images according to claim 3, characterized in that: The method of predicting the head multi-task output includes the following steps: 61) Bounding box branch regression detection frame coordinates, classification branch discrimination spore germination state, key point branch output germination tube end key point coordinates The germination loss function dynamically optimizes key point positioning: 62) Coordinate normalization and error calculation: The coordinates of the key points predicted by the network are normalized. The manually annotated key point coordinates (x, y, v) are divided by the input image size W and H respectively for normalization. The calculation formula is as follows: in, is the network predicted coordinate, (x, y) is the manually marked coordinate, W and H are the width and height of the input image, Δx is the horizontal coordinate offset, and Δy is the vertical coordinate offset; 63) Piecewise hybrid loss function design: In order to take into account both high-resolution sensitivity to small errors and robustness to large errors, a logarithmic-linear hybrid loss L(e) is adopted, which is defined as follows: L(e)=log(1+μe), when e≤δ; L(e)=log(1+μδ)+μ(e-δ) / (1+μδ), when e>δ, Among them, e is the normalized error, δ is the turning radius, and μ is the hyperparameter for adjusting the curvature; 64) Set visibility mask v∈{0,1} and filter gradient: L=Σv·L(e); 65) Dynamic recalibration of scaling factor: To adapt to the error size at different training stages, a scaling factor α_t is designed: in, is the exponential sliding mean of the key point loss of the current batch, ε is a small constant to prevent division by zero, and C is the initial scale constant; 66) Loss Normalization and Multi-Task Weighting: The total keypoint loss L_kp is normalized by the number of visible keypoints N_v: L_kp^norm=α_t×L_kp / (N_v+ε), Among them, N v is the number of visible points; at the same time, the gated vector gradient is truncated to reduce the memory usage, w kp is the multi-task weight, L_kp^norm is the normalized total loss of key points, The normalized loss is combined with other task losses according to preset weights to ensure stable training.

7. A method for detecting wheat scab spore germination phenotype based on microscopic images according to claim 1, characterized in that: The storage medium stores a computer program. When the computer program is executed by the processor, the method for detecting the germination phenotype of wheat fusarium spores based on microscopic images according to any one of claims 1 to 6 is implemented.

8. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for detecting the germination phenotype of wheat fusarium spores based on microscopic images according to any one of claims 1 to 6 is realized.

Images