High-throughput seed vigor grading evaluation method and system based on dual-flow visual framework

Abstract

The embodiment of the application provides a high-throughput seed vigor grading evaluation method and system based on a double-flow visual framework, comprising the following steps: S1: high-throughput seed image acquisition and preprocessing: acquiring high-resolution RGB panoramic images in the germination process of the same batch of seeds, and extracting single-seed images by adopting an instance-guided local cropping and adjacent mask strategy; S2: high-fidelity feature encoding extraction: inputting the single-seed images after preprocessing into a MorphoMamba network model, and extracting features by a full-resolution feature reservation encoder to reserve edge topological information of micro organs; S3: double-flow collaborative decoding and mask generation: adopting a double-flow collaborative decoder to decode and reconstruct deep semantic features. The application can overcome the loss problem of high-frequency features of micro organs in network extraction, and realize high-precision pixel-level segmentation and anti-distortion quantitative evaluation of seed micro organs.

Description

Technical Field

[0001] This application relates to the field of agricultural seed phenotypic analysis technology, and in particular to a high-throughput seed vigor grading assessment method and system based on a dual-stream visual framework. Background Technology

[0002] Seed vigor is a key physiological indicator of germination potential and stress resistance in crops such as wheat, corn, and soybeans, which possess morphological characteristics of tiny germination organs. During the early germination stage, the growth dynamics of these tiny organs (radicle and plumule) directly reflect seed vigor. Therefore, accurately quantifying the morphological characteristics of these tiny organs is crucial for efficient and non-destructive seed selection. However, existing visual assessment systems often overlook the influence of seed growth environment and batch-to-batch variations on morphological parameters. Without cross-sectional comparisons within the same batch, absolute morphological measurements (such as absolute length and absolute area) are easily affected by inconsistencies in initial physiological states and environmental conditions between different batches. Therefore, high-throughput, synchronous morphological quantification and grading within the same batch of seeds is a fundamental cornerstone for ensuring the scientific rigor and accuracy of vigor assessment.

[0003] Furthermore, traditional germination tests are typically time-consuming, inefficient, and destructive, making them unsuitable for large-scale screening. Current non-destructive testing technologies primarily rely on machine vision or hyperspectral imaging. However, while hyperspectral sensors can detect biochemical characteristics, their spatial resolution is extremely low, failing to finely characterize the morphological features of minute organs and hindering the establishment of accurate morphology-vitality mappings. In deep learning-based computer vision applications, existing image classification models can only output global probability labels, lacking the ability to spatially locate minute organs. Object detection-based methods (such as the YOLO series) inevitably introduce a high proportion of background noise pixels when dealing with curved, extremely small root and shoot organs using rigid bounding boxes. This coarseness in spatial representation prevents them from accurately identifying slow-developing or deformed "low-vitality" seeds.

[0004] Even more serious is the severe scale imbalance that exists during the conversion from high-throughput panoramic images to tiny targets. Most existing vision systems can only perform a rough binary classification of "germination" or rely on a single perspective of a two-dimensional image to estimate the absolute length of the germinal root. This method is highly susceptible to perspective distortion caused by the physical curvature of the organ. Regarding deep network architectures, traditional convolutional neural networks (CNNs) suffer irreversible loss of topological information at the edges of tiny organs during multi-level downsampling feature extraction. While the latest visual Transformers or state-space models (such as Mamba) possess strong macro-contextual awareness, they often induce feature smoothing effects when capturing global semantics, thus erasing extremely fragile local high-frequency boundary information.

[0005] In summary, overcoming the loss of high-frequency features of micro-organs in network extraction and achieving high-precision pixel-level segmentation and distortion-resistant quantitative evaluation of seed micro-organs in high-throughput complex scenarios with the same batch is a technical challenge that urgently needs to be solved in the field of agricultural seed phenotypic analysis. Summary of the Invention

[0006] The purpose of this application is to provide a high-throughput seed vigor grading assessment method and system based on a dual-stream vision framework to address the shortcomings of existing technologies. It can overcome the problem of loss of high-frequency features of micro-organs in network extraction and achieve high-precision pixel-level segmentation and distortion-resistant quantitative assessment of seed micro-organs.

[0007] This application provides a high-throughput seed vigor classification assessment method based on a two-stream visual framework, including the following steps: S1: High-throughput seed image acquisition and preprocessing: Acquire high-resolution RGB panoramic images of the same batch of seeds during germination, and extract individual seed images using instance-guided local cropping and adjacent masking strategies. S2: High-fidelity feature encoding extraction: The preprocessed single seed image is input into the MorphoMamba network model, and features are extracted through the full-resolution feature preservation encoder to preserve the edge topological information of the micro-organs; S3: Dual-stream collaborative decoding and mask generation: A dual-stream collaborative decoder is used to decode and reconstruct deep semantic features, and lossless addition is performed to fuse global macroscopic state perception and local high-frequency detail features to output a high-precision pixel-level semantic segmentation mask. S4: Quantitative phenotypic analysis and automated vigor rating: Based on the semantic segmentation mask, the projected area ratio of each organ is calculated, and an image vigor index is constructed to achieve fine-grained automated rating of rice seed vigor.

[0008] Optionally, in step S1, the instance-guided local cropping and adjacent mask strategy specifically includes: The original image is mirrored to prevent edge target truncation; a local area of ​​a set size is cropped using the geometric center of a single seed as the anchor point to preserve the original physical resolution of the target at 1:1; an instance mask is used as a spatial filter to retain only the core semantics of the current anchor point seed, and other adjacent seeds are forced to cover the background pixels to generate a high-resolution single target image that eliminates adjacent interference.

[0009] Optionally, in step S2, the full-resolution feature-preserving encoder has the following specific structure: A convolutional neural network is used to extract preliminary features, and two high-fidelity bypass branches are constructed in the shallow network: H / 1 branch: uses a lightweight convolutional sequence to directly extract high-resolution spatial features at 1:1 resolution from the original image; H / 2 branch: intercepts intermediate transition features at 1:2 resolution before max pooling operation; the features of the H / 1 and H / 2 branches bypass the deep downsampling operation and directly participate in the subsequent decoding upsampling process through long skip connections; between the encoding and decoding layers, the complex bottleneck layer structure is abandoned, and a 1×1 convolutional structure is used for channel dimension alignment to achieve high-fidelity feature transition.

[0010] Optionally, in step S3, the dual-stream collaborative decoder specifically includes a Mamba global branch and a CNN local branch, and adopts the following additive feature fusion strategy: Mamba's global branch uses a state-space model to perform global dependency analysis and extracts a low-frequency semantic basis graph, denoted as... CNN local branches extract high-frequency edge responses and fine-grained texture features, denoted as The fusion is performed through a pure linear superposition mechanism, and the fused output is... The calculation formula is: .

[0011] Optionally, step S4 includes the following specific steps: For each individual seed instance in the image, count the actual number of pixels in the seed body, the number of pixels in the radicle, and the number of pixels in the plumule; calculate the radicle projection area ratio: , in, Represents the ratio of the projected area of ​​the radicle. Represents the number of pixels in the germinal root. Represents the actual number of pixels in the seed body; Calculate the ratio of the projected area of ​​the germ: ,in Represents the ratio of the projected area of ​​the germ. Represents the number of pixels in the germ; By incorporating physiological weights, an image vitality index is constructed: Seeds with IVI > 0.15 are considered high-vitality seeds; seeds that have germinated but have IVI < 0.15 are considered low-vitality seeds; seeds with both radicle and plumule pixel counts less than 50 are considered dead or ungerminated seeds.

[0012] Optionally, when training the MorphoMamba network model, an equal-weighted combined loss function that integrates Focal Loss and Tversky Loss is used. : Focal Loss is used to suppress the weights of a large number of easily classified background samples; Tversky Loss is used to evaluate region overlap and prevents topological breaks and missed detections of tiny embryonic roots by introducing asymmetric penalty weights.

[0013] This application provides a sub-vitality grading assessment system, including: The image acquisition and preprocessing module is used to acquire high-resolution RGB panoramic images of rice seed germination process, and to extract single seed images using instance-guided local cropping and adjacent masking strategies. The high-fidelity feature encoding module is used to input the preprocessed single seed image into the MorphoMamba network model and extract features through the full-resolution feature preservation encoder to preserve the edge topological information of the micro-organs. The dual-stream collaborative decoding module is used to decode and reconstruct deep semantic features using a dual-stream collaborative decoder. It performs lossless addition and fusion of global macro-state perception and local high-frequency detail features to output a high-precision pixel-level semantic segmentation mask. The vitality quantification and grading module is used to calculate the projected area ratio of each organ based on the semantic segmentation mask and construct an image vitality index, thereby realizing fine-grained automated grading of rice seed vitality.

[0014] Optionally, the dual-stream collaborative decoder specifically includes a Mamba global branch and a CNN local branch. The Mamba global branch uses a state-space model to perform global dependency analysis and extract a low-frequency semantic basis map. The CNN local branch extracts high-frequency edge responses and fine-grained texture features, which are then fused through a pure linear superposition mechanism.

[0015] Optionally, for an independent seed instance in the image, the projection area ratio includes the radicle projection area ratio and the plumule projection area ratio; the radicle projection area ratio is equal to the ratio of the number of radicle pixels to the number of actual pixels of the seed body. The bud projection area ratio is equal to the ratio of the number of pixels in the embryo to the number of actual pixels in the seed body.

[0016] Optionally, the image viability index is equal to 0.6 times the ratio of the radicle projection area plus 0.4 times the ratio of the germ projection area; When the image vitality index is greater than 0.15, it is judged as a high-vitality seed; when the seed has germinated but the image vitality index is 0.15, it is judged as a low-vitality seed; when the number of radicle pixels is less than 50 and the number of plumule pixels is less than 50, it is judged as a dead seed or a seed that has not germinated.

[0017] Compared with the prior art, this application has the following beneficial effects. This paper proposes an instance-level local cropping and masking strategy to eliminate adjacent interference in high-density scenes and preserve the original 1:1 physical resolution of tiny targets without interpolation scaling. It innovatively proposes the MorphoMamba architecture, employing full-resolution skip connections to directly transmit H / 1 and H / 2 high-definition features, abandoning deep bottleneck layers that easily lead to feature smoothing, and avoiding the loss of high-frequency features of tiny organs during network extraction. Simultaneously, a dual-stream collaborative decoder is designed, perfectly combining the global perception of Mamba and the sharp local edges of CNNs through a simple additive fusion mechanism, achieving lossless detail reconstruction. The generated pixel-level segmentation results are used to extract the projected area ratio to construct an image vitality index. Compared to traditional bounding boxes or one-dimensional length estimation, the projected area ratio exhibits strong robustness to perspective distortion, thus achieving truly fine-grained, automated, high-throughput seed grading.

[0018] The model designed in this application significantly reduces computational redundancy. While achieving an average intersection-union ratio (mIoU) of 84.41% and an average F1 score of 91.40%, the computational cost is only 18.98 GFLOPs, which greatly improves the feasibility of its implementation on agricultural edge computing devices. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the steps of the high-throughput seed vigor classification assessment method based on a dual-stream visual framework proposed in the embodiments of this application. Figure 2 This is a flowchart of the image cropping and masking strategy guided by an example in the embodiments of this application; Figure 3 This is a diagram showing the overall architecture of the MorphoMamba network in the embodiments of this application; Figure 4 This is a network structure diagram of bottleneck layer 1 of the MorphoMamba network in the embodiments of this application; Figure 5 This is a diagram of the bottleneck layer 2 network structure of the MorphoMamba network in this embodiment of the application; Figure 6 This is a diagram of the encoder network structure in an embodiment of this application; Figure 7 This is a closed-loop flowchart of quantitative reasoning and automated vitality rating in the embodiments of this application; Figure 8This is a visualization of the segmentation results compared with other models in the embodiments of this application. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] In the description of this invention, it should be understood that the terms "longitudinal," "lateral," "thickness," "upper," "lower," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means two or more. In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] Please see Figures 1-8 As shown, this application discloses a high-throughput seed vigor classification evaluation method based on a two-stream visual framework, comprising the following steps: S1: High-throughput seed image acquisition and preprocessing: Acquire high-resolution RGB panoramic images of the same batch of seeds during germination, and extract individual seed images using instance-guided local cropping and adjacent masking strategies; the instance-guided local cropping and adjacent masking strategies specifically include: The original image is mirrored to prevent edge target truncation; a local area of ​​a set size is cropped using the geometric center of a single seed as the anchor point to preserve the original physical resolution of the target at 1:1; an instance mask is used as a spatial filter to retain only the core semantics of the current anchor point seed, and other adjacent seeds are forced to cover the background pixels to generate a high-resolution single target image that eliminates adjacent interference.

[0024] S2: High-fidelity feature encoding extraction: The preprocessed single seed image is input into the MorphoMamba network model, and features are extracted through a full-resolution feature preservation encoder to preserve the edge topological information of the micro-organs; the specific structure of the full-resolution feature preservation encoder is as follows: A convolutional neural network is used to extract initial features, and two high-fidelity bypass branches are constructed in the shallow network: the H / 1 branch, which directly extracts high-resolution spatial features at 1:1 resolution from the original image using a lightweight convolutional sequence; and the H / 2 branch, which intercepts intermediate transition features at 1:2 resolution before max pooling. Features from the H / 1 and H / 2 branches bypass deep downsampling operations and directly participate in the subsequent decoding upsampling process through long skip connections. Between the encoding and decoding layers, a complex bottleneck layer structure is abandoned, and a 1×1 convolutional structure is used for channel dimension alignment to prevent feature smoothing and achieve high-fidelity feature transition. This solves the problem of a sharp drop in resolution (down to 1 / 4) in extremely shallow layers caused by traditional Stem modules.

[0025] As one implementation, ResNet-50 is used as the basic backbone network to extract initial features. The H / 1 branch directly extracts high-resolution spatial features at 1:1 resolution from the original image using lightweight convolutional sequences; the H / 2 branch truncates intermediate transition features at 1:2 resolution without max pooling. These shallow features rich in edge texture directly participate in the last two upsampling operations of the decoding stage through long skip connections.

[0026] While complex bottleneck layers (such as ASPP and Transformer) can expand the receptive field, they can also cause severe feature smoothing effects, leading to blurred high-frequency features of small organs during deep transmission. Therefore, this model explicitly abandons the cumbersome Neck module. The deepest semantic features are aligned through simple 1×1 convolutions, i.e., input to the decoding stage, ensuring that deep information is in its most original, high-fidelity state.

[0027] S3: Dual-stream collaborative decoding and mask generation: A dual-stream collaborative decoder is used to decode and reconstruct deep semantic features, and lossless addition is performed to fuse global macroscopic state perception and local high-frequency detail features to output a high-precision pixel-level semantic segmentation mask. The dual-stream collaborative decoder specifically includes a Mamba global branch and a CNN local branch, and employs the following additive feature fusion strategy: Mamba's global branch uses a state-space model to perform global dependency analysis and extracts a low-frequency semantic basis graph, denoted as... CNN local branches extract high-frequency edge responses and fine-grained texture features, denoted as The fusion is performed through a pure linear superposition mechanism, and the fused output is... The calculation formula is: A purely linear additive feature fusion strategy is adopted to avoid the signal attenuation defect caused by weak high-frequency features in the traditional Sigmoid gating mechanism. At the physical level of this formula, traditional multimodal fusion typically employs gating mechanisms such as the Sigmoid or Softmax functions, whose expressions usually involve nonlinear multiplication operations (e.g., multiplying feature maps using attention weights). However, considering that the ends of the germinal roots or newly formed germinal buds appear as extremely weak high-frequency edges in the image, their pixel response values... Typically, the values ​​are extremely low. If the signal passes through a sigmoid activation function (whose value range is between 0 and 1), this multiplication operation will inevitably result in the fused response value being strictly smaller than the original observation value, thus triggering "signal attenuation" and secondary suppression of high-frequency features. This application innovatively employs pure linear superposition (i.e., abandoning all weighted penalty networks). Because... Focusing on extracting high-frequency local topological boundaries, Focusing on extracting low-frequency global macroscopic locations, both macroscopic and microscopic features naturally possess highly complementary positive interaction characteristics within the feature semantic space. By directly adding them together, this system ensures perfect overlap between macroscopic positioning and high-definition microscopic contours, achieving 100% lossless feature transmission and thus avoiding the phenomenon of edge topology breaks.

[0028] The MorphoMamba model described above was used for training multi-class micro-organ segmentation. The following hyperparameters and optimization settings were configured: the experiment was based on the PyTorch framework, the optimizer was AdamW, the base learning rate was set to 5e-5, and the weight decay coefficient was 1e-2. A cosine annealing strategy was used to dynamically adjust the learning rate throughout the training cycle. During training, data augmentation mechanisms were introduced, including random horizontal flipping, random vertical flipping, and color jittering, to enhance the model's generalization ability.

[0029] To address the severe imbalance between positive and negative samples caused by tiny organs, a composite loss function combining Focal Loss and Tversky Loss with equal weights is employed. Focal Loss dynamically scales the cross-entropy loss, suppressing the weights of a large number of easily classified background samples and forcing the model to focus on hard-to-classify edges; Tversky Loss introduces asymmetric penalty weights. This increases the penalty for false negatives (FN), effectively preventing topological breaks and missed detections of delicate embryonic roots.

[0030] Model performance evaluation metrics calculation: After model training, overall accuracy (OA), intersection-over-union ratio (IoU), F1 score (F1), and panorama quality (PQ) are used as the core evaluation metrics. The specific formulas are as follows: Wherein, TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. To rigorously evaluate the instance-level separation capability of small organs in high-density scenes, this embodiment also calculates panoramic quality (PQ): ,in, and These represent the predicted and actual instance regions, respectively.

[0031] To further verify the efficiency of the MorphoMamba model described in this embodiment, under the same hardware conditions (e.g., NVIDIA RTX 4090), the performance of the method in this application was compared with that of existing mainstream semantic segmentation networks (such as U-Net, TransUNet, SegFormer, etc.). While the traditional U-Net model can achieve a certain segmentation accuracy when processing the aforementioned high-throughput data, its massive number of parameters leads to extremely high computational complexity (e.g., FLOPs as high as 83.84G), and the large number of redundant convolutional blocks greatly limits its deployment on agricultural portable edge devices. On the other hand, large models based on pure attention mechanisms (such as TransUNet) tend to have lower F1 scores than the method in this application because their global self-attention computation easily acts as a "global low-pass filter," causing feature smoothing effects and blurring of the tiny germinal roots.

[0032] In comparison, the MorphoMamba model used in this application achieves a computational complexity of only 18.98 GFLOPs through high-fidelity bypass branches H / 1 and H / 2 and by discarding complex bottleneck layers, which is more than 77% lower than traditional methods. At the same time, it achieves the best performance in the field in various seed vitality indicators (such as OA, mIoU, F1, PQ, etc.).

[0033] S4: Quantitative phenotypic analysis and automated vigor rating: Based on semantic segmentation masking, the projected area ratio of each organ is calculated and an image vigor index is constructed to achieve fine-grained automated rating of rice seed vigor.

[0034] Step S4 includes the following specific steps: The system reverse-maps the local mask output by the network back to the panoramic image based on the recorded absolute center coordinates. Instance separation is performed using connected component analysis (CCA), and an area threshold of 100 pixels is set to filter out unstructured noise.

[0035] For an individual seed instance in the image, such as the i-th seed instance, count the number of real pixels in the seed body, the number of pixels in the radicle, and the number of pixels in the plumule; calculate the radicle projection area ratio: , in, Represents the ratio of the projected area of ​​the radicle. Represents the number of pixels in the germinal root. Represents the actual number of pixels in the seed body; Calculate the ratio of the projected area of ​​the germ: ,in Represents the ratio of the projected area of ​​the germ. Represents the number of pixels in the germ; By incorporating physiological weights, an image vitality index is constructed: Seeds with an IVI > 0.15 are classified as high-viability seeds and marked with a green border; seeds that have germinated but have an IVI < 0.15 are classified as low-viability seeds and marked with a yellow border; seeds with both radicle and plumule pixel counts less than 50 are classified as dead or ungerminated seeds and marked with a red border. This relative scoring mechanism based on the same batch greatly enhances the reference value of the evaluation results in actual agricultural screening.

[0036] The following is a specific example application: Sample Preparation. The rice seed samples used in this example were selected from the same variety, origin, and harvest period of the same batch to eliminate genetic and environmental differences. This batch of seeds underwent standardized germination treatment in an artificial climate incubator. To achieve cross-sectional comparison of seed vigor within the same batch, the seeds were evenly arranged in petri dishes, with 25 seeds precisely placed in each dish. This arrangement ensured that the seeds of the same batch were under completely consistent temperature, humidity, and light conditions during germination. This specific arrangement effectively prevented severe cross-adhesion in the later stages of germination. The incubator parameters were set to a constant temperature of 26°C and provided 24-hour continuous light to ensure a stable germination environment.

[0037] High-resolution RGB images were acquired. A high-throughput intelligent monitoring platform for rice germination was constructed, using an 8-megapixel high-definition RGB camera mounted from the top to acquire image data. This camera is equipped with a distortion-free wide-angle lens, achieving a single image resolution of up to 3840×2160 pixels. This example acquired 260 high-throughput rice images, containing 6500 individual seeds, with the sample evenly covering the three typical developmental stages: non-germination, initial white emergence (appearance of tiny radicles / plumules), and obvious germination.

[0038] Dataset Labeling and Partitioning. To meet the requirement of accurate calculation of projected area ratio (PAR), this study abandoned coarse bounding box annotation and used LabelMe software for rigorous pixel-level instance annotation on keyframes. Image pixels were finely divided into four categories: background, seed body, radicle, and germinal root. Finally, the complete dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for model parameter fitting, the validation set was used to monitor the training process and save the optimal model weights, and the test set was completely independent and used to evaluate the generalization ability and actual recognition performance of the final saved model.

[0039] Instance-guided coarse-to-fine preprocessing. To address the scale imbalance between the panoramic view and the tiny targets, a mirror fill operation is first applied to the original image to prevent truncation of edge targets. Then, a 224×224 pixel local region is cropped using the geometric center of each seed as an anchor point. This operation preserves the original physical resolution of the target at a 1:1 ratio, completely avoiding high-frequency detail distortion caused by interpolation scaling. Simultaneously, to mitigate neighbor interference in high-density scenes, an instance mask is used as a spatial filter, retaining only the core semantics of the current anchor point seed, while neighboring seeds are forcibly overlaid as background pixels, generating high-quality, high-resolution single-target image patches.

[0040] The high-throughput seed vigor grading assessment method based on the dual-stream visual framework described above was used to grade the test set (650 seeds in total). The grading results are shown in Table 1.

[0041] The rating results for the test set are shown in Table 1: The total germination rate in the table above is 42.46%, which is relatively low. More than half of the seeds (57.54%) are completely inactive, and only 27.7% of the seeds are effective and highly viable, which is less than one-third. The overall seed quality is poor.

[0042] Although the embodiments of this application are described in detail using rice seeds as an example, those skilled in the art should understand that the seed vigor grading and evaluation method of this application is also applicable to the screening of other crop seeds with micro-germination organ morphological characteristics, such as wheat, corn, and soybeans. Furthermore, the artificial climate incubator environment (26°C constant temperature, 24-hour light) mentioned in this invention is only one of the optimal implementation environments. In actual industrialized high-throughput agricultural screening lines, a conveyor belt assembly line combined with a multi-angle stroboscopic light source can also be used for alternative sampling. These equivalent substitution schemes based on the core concept of this invention all fall within the protection scope of this application.

[0043] This application also discloses a sub-vitality grading assessment system, including: The image acquisition and preprocessing module is used to acquire high-resolution RGB panoramic images of rice seed germination process, and to extract single seed images using instance-guided local cropping and adjacent masking strategies. The high-fidelity feature encoding module is used to input the preprocessed single seed image into the MorphoMamba network model and extract features through the full-resolution feature preservation encoder to preserve the edge topological information of the micro-organs. The dual-stream collaborative decoding module is used to decode and reconstruct deep semantic features using a dual-stream collaborative decoder. It performs lossless addition and fusion of global macro-state perception and local high-frequency detail features to output a high-precision pixel-level semantic segmentation mask. The vitality quantification and grading module is used to calculate the projected area ratio of each organ based on semantic segmentation mask and construct an image vitality index, thereby realizing fine-grained automated grading of rice seed vitality.

[0044] Preferably, the dual-stream collaborative decoder specifically includes a Mamba global branch and a CNN local branch. The Mamba global branch uses a state-space model to perform global dependency analysis and extract a low-frequency semantic basis map. The CNN local branch extracts high-frequency edge responses and fine-grained texture features, which are then fused through a pure linear superposition mechanism.

[0045] Preferably, for an independent seed instance in an image, the projection area ratio includes the radicle projection area ratio and the plumule projection area ratio; the radicle projection area ratio is equal to the ratio of the number of radicle pixels to the number of actual pixels of the seed body. The bud projection area ratio is equal to the ratio of the number of pixels in the embryo to the number of actual pixels in the seed body.

[0046] The image viability index is equal to 0.6 times the ratio of the radicle projection area plus 0.4 times the ratio of the plumule projection area; When the image vitality index is greater than 0.15, it is judged as a high-vitality seed; when the seed has germinated but the image vitality index is 0.15, it is judged as a low-vitality seed; when the number of radicle pixels is less than 50 and the number of plumule pixels is less than 50, it is judged as a dead seed or a seed that has not germinated.

[0047] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0048] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A high-throughput seed vigor classification assessment method based on a two-stream visual framework, comprising the following steps: S1: High-throughput seed image acquisition and preprocessing: Acquire high-resolution RGB panoramic images of the same batch of seeds during germination, and extract individual seed images using instance-guided local cropping and adjacent masking strategies. S2: High-fidelity feature encoding extraction: The preprocessed single seed image is input into the MorphoMamba network model, and features are extracted through the full-resolution feature preservation encoder to preserve the edge topological information of the micro-organs; S3: Dual-stream collaborative decoding and mask generation: A dual-stream collaborative decoder is used to decode and reconstruct deep semantic features, and lossless addition is performed to fuse global macroscopic state perception and local high-frequency detail features to output a high-precision pixel-level semantic segmentation mask. S4: Quantitative phenotypic analysis and automated vigor rating: Based on the semantic segmentation mask, the projected area ratio of each organ is calculated, and an image vigor index is constructed to achieve fine-grained automated rating of rice seed vigor.

2. The high-throughput seed vigor grading assessment method based on a dual-stream visual framework according to claim 1, characterized in that, In step S1, the instance-guided local cropping and adjacent mask strategy specifically includes: The original image is mirrored to prevent edge target truncation; a local area of ​​a set size is cropped using the geometric center of a single seed as the anchor point to preserve the original physical resolution of the target at 1:1; an instance mask is used as a spatial filter to retain only the core semantics of the current anchor point seed, and other adjacent seeds are forced to cover the background pixels to generate a high-resolution single target image that eliminates adjacent interference.

3. The high-throughput seed vigor classification assessment method based on a dual-stream visual framework according to claim 1, characterized in that, In step S2, the full-resolution feature-preserving encoder has the following specific structure: A convolutional neural network is used to extract preliminary features, and two high-fidelity bypass branches are constructed in the shallow network: H / 1 branch: a lightweight convolutional sequence is used to directly extract high-resolution spatial features with a 1:1 resolution from the original image; H / 2 branch: Intercepts intermediate transition features at 1:2 resolution before max pooling operation; The features of the H / 1 and H / 2 branches bypass the deep downsampling operation and directly participate in the subsequent decoding upsampling process through long skip connections. Between the encoding and decoding layers, the complex bottleneck layer structure is abandoned, and a 1×1 convolution structure is used for channel dimension alignment to achieve high-fidelity feature transition.

4. The high-throughput seed vigor classification assessment method based on a dual-stream visual framework according to claim 1, characterized in that, In step S3, the dual-stream collaborative decoder specifically includes a Mamba global branch and a CNN local branch, and adopts the following additive feature fusion strategy: Mamba's global branch uses a state-space model to perform global dependency analysis and extracts a low-frequency semantic basis graph, denoted as... CNN local branches extract high-frequency edge responses and fine-grained texture features, denoted as The fusion is performed through a pure linear superposition mechanism, and the fused output is... The calculation formula is: .

5. The high-throughput seed vigor classification assessment method based on a dual-stream visual framework according to claim 1, characterized in that, Step S4 includes the following specific steps: For each individual seed instance in the image, count the actual number of pixels in the seed body, the number of pixels in the radicle, and the number of pixels in the plumule; calculate the radicle projection area ratio: ， in, Represents the ratio of the projected area of ​​the radicle. Represents the number of pixels in the germinal root. Represents the actual number of pixels in the seed body; Calculate the ratio of the projected area of ​​the germ: ,in, Represents the ratio of the projected area of ​​the germ. Represents the number of pixels in the germ; By incorporating physiological weights, an image vitality index is constructed: Seeds with IVI > 0.15 are considered high-vitality seeds; seeds that have germinated but have IVI < 0.15 are considered low-vitality seeds; seeds with both radicle and plumule pixel counts less than 50 are considered dead or ungerminated seeds.

6. The high-throughput seed vigor classification assessment method based on a dual-stream visual framework according to claim 1, characterized in that, When training the MorphoMamba network model, an equal-weighted combined loss function that integrates Focal Loss and Tversky Loss was used. : Focal Loss is used to suppress the weights of a large number of easily classified background samples; Tversky Loss is used to evaluate regional overlap, and it prevents topological breaks and missed detections of tiny embryonic roots by introducing asymmetric penalty weights.

7. A seed vigor grading and assessment system, characterized in that, include: The image acquisition and preprocessing module is used to acquire high-resolution RGB panoramic images of rice seed germination process, and to extract single seed images using instance-guided local cropping and adjacent masking strategies. The high-fidelity feature encoding module is used to input the preprocessed single seed image into the MorphoMamba network model and extract features through the full-resolution feature preservation encoder to preserve the edge topological information of the micro-organs. The dual-stream collaborative decoding module is used to decode and reconstruct deep semantic features using a dual-stream collaborative decoder. It performs lossless addition and fusion of global macro-state perception and local high-frequency detail features to output a high-precision pixel-level semantic segmentation mask. The vitality quantification and grading module is used to calculate the projected area ratio of each organ based on the semantic segmentation mask and construct an image vitality index, thereby realizing fine-grained automated grading of rice seed vitality.

8. The seed vigor grading and evaluation system according to claim 7, characterized in that, The dual-stream collaborative decoder specifically includes a Mamba global branch and a CNN local branch. The Mamba global branch uses a state-space model to perform global dependency analysis and extracts a low-frequency semantic basis map. The CNN local branch extracts high-frequency edge responses and fine-grained texture features, which are then fused through a pure linear superposition mechanism.

9. The seed vigor grading and evaluation system according to claim 7, characterized in that, For individual seed instances in the image, the projected area ratio includes the radicle projected area ratio and the plumule projected area ratio; The ratio of the projected area of ​​the radicle is equal to the ratio of the number of pixels in the radicle to the number of actual pixels in the seed body; The bud projection area ratio is equal to the ratio of the number of pixels in the embryo to the number of actual pixels in the seed body.

10. The seed vigor grading and evaluation system according to claim 9, characterized in that, The image viability index is equal to 0.6 times the ratio of the radicle projection area plus 0.4 times the ratio of the germ projection area. When the image vitality index is greater than 0.15, it is judged as a high-vitality seed; when the seed has germinated but the image vitality index is 0.15, it is judged as a low-vitality seed; when the number of radicle pixels is less than 50 and the number of plumule pixels is less than 50, it is judged as a dead seed or a seed that has not germinated.

Images