Seed germination parameter measurement method based on double YOLO and topological skeleton extraction

Abstract

The present application relates to the technical field of agricultural production, and more particularly to a seed germination parameter measurement method based on double YOLO and topological skeleton extraction, specifically comprising the following steps: S1, seed double-layer instance segmentation detection; S2, mask correlation matching and germination state determination; S3, seedling region skeleton refinement processing; S4, main skeleton purification and seedling length calculation; S5, germination index statistics and visual output. The present application filters out the seed coat background interference through the mask matching algorithm, and matches the topological skeleton purification algorithm to accurately restore the actual length of the radicle, overcoming various defects existing in the prior art from multiple dimensions, and completing high-precision and standardized automatic integrated detection of seed germination rate and seedling length.

Description

Technical Field

[0001] This invention relates to the field of agricultural production technology, and in particular to a method for measuring seed germination parameters based on dual YOLO and topological skeleton extraction. Background Technology

[0002] In recent years, deep learning technology has made groundbreaking progress in the field of computer vision. Convolutional neural networks have far surpassed traditional image processing methods in terms of accuracy and robustness in tasks such as object detection, instance segmentation, and feature extraction. As machine vision technology penetrates deeper into the agricultural field, seed germination detection schemes based on deep learning have gradually become a research hotspot. Existing related patent solutions have made some progress in the acquisition of seed germination information, but there are still obvious technical shortcomings.

[0003] Publication number CN116612139B implements germination rate calculation based on deep learning. It uses a deep learning object detection model to identify seeds and germination points, and counts the seeds to calculate the germination rate, solving the problem of low efficiency of manual counting. However, its technical solution only stays at the object detection level and cannot perform pixel-level fine-grained segmentation of the seed coat and bud, nor does it have the ability to measure morphological parameters such as bud length and root length. The detection dimension is single and cannot comprehensively evaluate the seed vigor level.

[0004] Publication number CN119516247A possesses sprout length measurement capabilities. It adds a sprout length measurement step to target detection, sequentially binarizing, dilating, blurring, and extracting the skeleton from the cropped image of germinating seeds, then converting the sprout length data using the longest straight-line distance of the skeleton. However, the algorithm has several inherent flaws, covering the three core steps of target differentiation, skeleton extraction, and length measurement.

[0005] (1) This patent directly extracts the skeleton from the binary image of the whole germinated seed without peeling off the seed coat area. The seed coat will interfere with the skeleton generation, causing the skeleton starting point to be misaligned and the path to be redundant.

[0006] (2) This patent relies on the traditional morphological dilation and erosion algorithm to extract the skeleton. The skeleton cannot guarantee the width of a single pixel and is prone to problems such as adhesion, breakage and center line offset.

[0007] (3) The patent uses straight-line distance to represent the length of curved seed buds, which deviates greatly from the actual growth path of the seed buds, resulting in a problem of seed bud length measurement distortion.

[0008] In summary, existing technologies cannot achieve precise separation of seed coat and bud, or accurate path-based length measurement of curved young roots. Furthermore, they lack complete automated calculation logic for germination information indicators, and their detection accuracy and completeness fail to meet national seed quality inspection requirements. Therefore, this invention designs a dual YOLO hierarchical segmentation architecture to achieve coarse and fine dual-layer detection. Combined with mask matching to eliminate seed coat interference, and relying on a topological skeleton purification algorithm to accurately restore the true length of young roots, this invention comprehensively addresses the shortcomings of existing technologies, achieving high-precision, standardized, and fully automated measurement of seed germination rate and bud length. Summary of the Invention

[0009] This invention discloses a method for measuring seed germination parameters based on dual YOLO and topological skeleton extraction. The invention designs a dual YOLO hierarchical segmentation architecture to construct a coarse and fine two-layer detection system. It filters out seed coat background interference through a mask matching algorithm and accurately restores the actual length of young roots by combining it with a topological skeleton purification algorithm. This method overcomes various defects of existing technologies from multiple dimensions and achieves high-precision, standardized, fully automatic integrated detection of seed germination rate and seedling length.

[0010] The technical implementation scheme of the present invention is as follows:

[0011] A method for measuring seed germination parameters based on dual YOLO and topological skeleton extraction includes the following steps:

[0012] S1. Seed Dual-Level Instance Segmentation and Detection: Construct a two-level instance segmentation and detection model architecture, perform parallel inference on the collected panoramic images of seeds in petri dishes, and output the global mask of the overall seed target and the fine-grained local mask of the seed coat and bud structure inside the seed, forming a dual-level detection mode of global coarse localization segmentation and local fine structure segmentation.

[0013] S2. Mask association matching and germination status determination: A mask matching strategy is adopted to establish a one-to-one mapping between the global mask of the seed and the local masks of the seed coat and the germination of the seed bud, thereby determining the germination status of a single seed.

[0014] S3. Refinement of the budding region skeleton: In step S2, the skeleton of the budding region mask that is determined to be a germinating seed is refined, and the budding region with a certain width is shrunk to a single-pixel width and a topologically continuous central skeleton line.

[0015] S4. Main skeleton purification and seed bud length calculation: Based on the neighborhood topology features of the skeleton pixels, the initial central skeleton is purified, the lateral branch burrs are removed and the end pixels of the main trunk are restored, and the actual physical length of the seed bud is calculated based on the purified main skeleton.

[0016] S5. Germination index statistics and visualization output: Statistically analyze the germination status and seed bud length of all seeds, calculate the core indicators of germination rate and average seed bud length, and generate a visualization image with superimposed detection and labeling information for output.

[0017] As a further description of the above technical solution, step S1 includes the following steps:

[0018] S1-1. Use image acquisition equipment with a fixed position, and agree on the shooting height, focal length and image resolution to acquire panoramic images of the culture dish;

[0019] S1-2. Based on homologous seed images, two types of labeled datasets were created: the first type is a coarse-grained labeled dataset, which only labels the complete outline of the planted individual without distinguishing the internal structure and germination status; the second type is a fine-grained labeled dataset, which separates and labels the seed coat and seed bud into two independent structural regions.

[0020] S1-3. Train a first-level baseline segmentation model based on a coarse-grained labeled dataset, and infer the localization results, counting results, and global mask of all seeds in the petri dish.

[0021] S1-4. A two-level subdivision segmentation model is trained based on a fine-grained labeled dataset. The model infers and outputs two types of fine-grained local masks: seed coat mask and seed bud mask. The seed coat region and seed bud region are isolated to eliminate the interference of the seed coat on subsequent length measurement.

[0022] As a further description of the above technical solution, step S2 specifically includes the following:

[0023] S2-1. The mask coverage is used as the matching evaluation index for the two-level mask. The formula for calculating the mask coverage is as follows:

[0024]

[0025] In the formula, For mask coverage; This represents the total number of pixels in the intersection region of the primary seed mask and the secondary local mask (seed coat / bud). This represents the total number of pixels in the second-level local mask itself.

[0026] Differential matching thresholds are set for seed coat and seed bud respectively, with the seed coat matching threshold set to 0.7 and the seed bud matching threshold set to 0.6; when the mask coverage is greater than the corresponding threshold, the two sets of masks are considered to be successfully matched.

[0027] S2-2. A two-round greedy pairing strategy is used to perform the matching. The first round completes the pairing of the global planting mask and the seed coat mask; the second round completes the pairing of the global seed mask and the seed bud mask.

[0028] In each round, the mask coverage is paired in descending order. Masks that have been occupied are selected and marked in sequence to ensure that there are no duplicates in the corresponding matches.

[0029] S2-3. In step S2-2, if the global mask of a single seed matches both the seed coat mask and the bud mask, the seed is determined to be a germinated seed; if only the seed coat mask is matched, the seed is determined to be an ungerminated seed.

[0030] As a further description of the above technical solution, step S3 employs a parallel thinning algorithm to perform skeleton thinning processing, specifically including:

[0031] S3-1, Using a 3×3 neighborhood as the basic decision unit, with the current pixel to be detected... Centered on the target, the binary states of 8 neighboring pixels are extracted in a clockwise order, and two types of core decision parameters are defined: the total number of foreground pixels in the neighborhood. Number of background-to-foreground transitions The calculation formula is as follows:

[0032]

[0033]

[0034] S3-2. Each iteration includes two symmetrical sub-scanning stages, setting deletion constraints for boundary pixels in the southeast and northwest directions respectively. Pixels that simultaneously satisfy the constraints of the number of neighboring foreground pixels, the number of jumps, and the directional boundary are marked as deletable boundary pixels. After all scans in a single round are completed, all marked pixels are deleted in batches.

[0035] S3-3. Iterate and refine the algorithm repeatedly until there are no more pixels to delete in a single iteration. When the algorithm converges, it outputs a single-pixel continuous central skeleton that is consistent with the topology of the original seed bud region.

[0036] As a further description of the above technical solution, step S4 specifically includes:

[0037] S4-1: Use 8 neighborhood transition times As the core indicator for topology determination, the determination rules are as follows:

[0038] Pixels with a jump count of 1 are determined to be endpoint pixels, pixels with a jump count of 2 are determined to be middle pixels of the main branch, and pixels with a jump count of not less than 3 are determined to be branch node pixels.

[0039] S4-2: Initialize the deletion round mask matrix to record the pixel deletion level of each round, and remove all skeleton endpoint pixels in batches until the number of remaining skeleton endpoints is no more than 2, thus obtaining the preliminary trunk skeleton with side branches and burrs removed.

[0040] S4-3: Starting from the two endpoints of the initial main skeleton, traverse the neighboring pixels in reverse order from high to low according to the deletion level, restore the mistakenly deleted end pixels of the main skeleton layer by layer, and restore the true main skeleton of the seed bud.

[0041] S4-4: The total pixel length of the main skeleton is calculated using the directional weighting method, and then the actual physical length of the seed bud is obtained by converting the pixel to physical length ratio.

[0042] As a further description of the above technical solution, step S5 specifically includes:

[0043] In step S5, the formula for calculating the seed germination rate is as follows:

[0044]

[0045] In the formula, GR is the seed germination rate; This represents the total number of germinated seeds. This represents the total number of seeds in the petri dish;

[0046] The formula for calculating the average seedling length is:

[0047]

[0048] In the formula, The average physical length of the seed bud; Let be the physical length of the germinated seed (the length of the ungerminated seed is 0).

[0049] As a further description of the above technical solution, a mask annotation map is generated by overlaying a seed detection box, a seed coat segmentation mask, and a seed bud segmentation mask onto the original seed image; a skeleton annotation map is generated by overlaying the purified seed bud main skeleton line onto the original seed image, and two types of visual annotation images are output for result verification.

[0050] An automatic seed germination information measurement device is characterized by comprising an image acquisition module, a two-level segmentation and reasoning module, a mask matching and state determination module, a skeleton refinement module, a main skeleton purification and length calculation module, and an index calculation and visualization output module; the device is used to execute the automatic seed germination information measurement method according to any one of claims 1 to 6.

[0051] The present invention has the following advantages:

[0052] 1. This invention adopts a dual YOLO coarse-fine two-level segmentation architecture. The first-level model completes global seed detection, and the second-level model realizes pixel-level fine segmentation of seed coat and seed bud, which can avoid interference caused by the seed coat region to the seed bud length measurement.

[0053] 2. This invention uses a coverage-based dual-round greedy matching algorithm to match the two-level mask results, and relies on the two-level model inference results to achieve dual verification, thereby improving the recognition accuracy of the overall seed and germination information.

[0054] 3. This invention extracts the main framework of the bud based on the Guo-Hall refinement algorithm, and combines neighborhood topology analysis and a deletion round backtracking mechanism to purify the main framework, thoroughly removing interference from root lateral branches and burrs. It can stably extract the main framework for roots with different morphologies, curvatures, and numbers of branches, with measurement accuracy far superior to morphological methods based on dilatation and erosion calculations.

[0055] 4. The algorithm proposed in this invention is fully automated, with high processing efficiency. The detection time for a single petri dish image is only a few seconds, which is dozens of times more efficient than manual detection. It is suitable for large-scale seed quality inspection and experimental scenarios. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0057] Figure 2 The image shows a petri dish taken from a fixed camera position according to the present invention.

[0058] Figure 3 This is a schematic diagram of the first-level YOLO segmentation result of the present invention.

[0059] Figure 4 This is a schematic diagram of the two-level YOLO segmentation results of the present invention.

[0060] Figure 5 This is a visual annotation diagram of the mask and skeleton of the present invention.

[0061] Figure 6 This is a diagram of the germination information visualization interface of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It is hereby declared that the directional terms such as up, down, left, right, front, back, inside, and outside used in this text are based solely on the accompanying drawings and are not intended to specifically limit the invention.

[0063] Example 1:

[0064] As attached Figure 1 The diagram shown is a detailed construction flowchart of the present invention. The seed germination information measurement method specifically includes:

[0065] Step 1: Construct a two-level YOLO instance segmentation model detection architecture, training a first-level baseline model and a second-level subdivision model respectively. Perform synchronous inference on the input petri dish seed image to obtain the overall seed target mask, seed coat and fine-grained pixel masks respectively, realizing a two-layer detection mode of "global coarse detection + local fine segmentation", specifically including:

[0066] 1. For example Figure 2The image shown is a sample image. A panoramic image of the petri dish was acquired by using a fixed industrial camera position, with uniform shooting height, focal length, and image resolution.

[0067] 2. Two sets of independent pixel-level mask annotations are performed based on images from the same source: The first set of coarse-grained annotations only annotates complete seed individuals, without distinguishing germination status or dissecting internal structures, and is used for primary model training; the second set of fine-grained annotations accurately separates and annotates two independent regions, the seed coat and the bud, and is used for secondary model training.

[0068] 3. For example Figure 3 As shown, YOLO s11-seg was selected as the base network to build a first-level baseline model. The complete seed was used as the sole segmentation target, and end-to-end training was completed using a coarse-grained labeled dataset. During the model inference phase, the location, counting, and global mask output of all seeds in the petri dish were completed, providing a foundation for subsequent mask matching and total seed count.

[0069] 4. For example Figure 4 As shown, the secondary subdivision model is also built on YOLOs11-seg, trained on a fine-grained labeled dataset, with seed coat and seed bud as two independent segmentation targets, and outputs refined seed coat and seed bud masks.

[0070] It should be noted that this model accurately segments the internal functional structure of germinating seeds, completely isolating the seed coat region from the seed bud region to be tested, thus eliminating the interference of the seed coat on subsequent skeleton extraction and root length measurement from the source.

[0071] The second step, the seed germination information measurement method, involves a two-round greedy matching of masks and determination of seed germination status. This means establishing a one-to-one correspondence between the primary germinating seed mask and the secondary seed coat and young root masks, thus achieving a precise association between germination characteristics and individual seeds. Specifically, this includes the following steps:

[0072] 1. Using mask coverage as the core evaluation metric for matching, the degree of overlap and correlation between two-level masks is quantified. The coverage calculation formula is as follows:

[0073] Formula 1;

[0074] In the formula, For mask coverage; This represents the total number of pixels in the intersection region of the primary seed mask and the secondary local mask (seed coat / bud). The total number of pixels in the secondary local mask; a differential threshold is set based on the morphological differences between the two types of targets. For seed coats with regular morphology, the matching threshold is set to 0.7, while for seed buds with narrow and varied morphology, the matching threshold is set to 0.6, balancing matching accuracy and error tolerance. If the value is greater than the threshold, the two sets of masks are considered to be matched successfully; otherwise, the match is considered to be unsuccessful.

[0075] 2. A two-round independent greedy pairing strategy is adopted. In the first round, the seed mask and seed coat mask are paired, and in the second round, the seed mask and bud mask are paired. In each round, the matching pairs are sorted in descending order of coverage, and the occupied masks are selected and marked in sequence to ensure one-to-one correspondence and no duplicate matching.

[0076] 3. For seed masks that match both the bud and the seed coat, they are considered as germinated seeds; for seed masks that only match the seed coat, they are considered as ungerminated seeds.

[0077] Step 3: During the germination information measurement process, the skeleton of the germination region is refined. The Guo-Hall parallel refinement algorithm is used to perform morphological shrinkage on the binary mask of the germination region. That is, through multiple rounds of iteration, the boundary pixels of the root region are peeled off layer by layer, shrinking the germination region with a certain width to a central skeleton line with a single pixel width. The specific steps include:

[0078] 1. Using a 3×3 neighborhood as the basic decision unit, and the current pixel to be detected... Extract 8 neighboring pixels in clockwise order, centered on the target. The binary state (foreground pixel value is 1, background pixel value is 0);

[0079] It should be noted that two types of core decision parameters are defined: the total number of neighborhood foreground pixels and the number of background-to-foreground transitions. The calculation formulas are as follows:

[0080] Formula 2;

[0081] Formula 3;

[0082] In the formula, This represents the total number of foreground pixels in the neighborhood. The number of transitions from background to foreground.

[0083] 2. Each iteration consists of two symmetrical sub-scanning phases: the first sub-scan sets a specific deletion constraint for the boundary pixels in the southeast direction, while simultaneously satisfying... , When the following four conditions are met, the pixel is marked as deletable: the pixels of due north, due east, and due south are not simultaneously foreground, and the pixels of due east, due south, and due west are not simultaneously foreground; the second sub-scan sets symmetry constraints on the boundary pixels in the northwest direction, and marks the pixels as deletable when the corresponding conditions are met; after the two sub-scans are completed, all marked pixels are deleted in batches to ensure that the skeleton is centered and symmetrical in shape.

[0084] 3. The iterative process is executed repeatedly until no pixel is marked as deletable in a single iteration, at which point the algorithm reaches convergence and outputs a continuous skeleton with a single-pixel width that is consistent with the original root system morphology and topology. Furthermore, the algorithm employs a parallel thinning architecture, which can process more boundary pixels per iteration, resulting in a faster convergence speed than traditional serial thinning algorithms. It also maintains better connectivity for obliquely growing root skeletons, making it suitable for detection scenarios with varying seed root morphologies.

[0085] The fourth step involves refining the main skeleton based on the neighborhood topology number and backtracking mechanism. This involves iteratively removing topological endpoints from the initial skeleton to eliminate all lateral branches and burrs, then restoring the main stem ends along the deletion path to obtain a unique main skeleton without branching. The seed bud length is then precisely calculated using a direction-weighted method and the conversion ratio K between physical length and pixels. The specific method is as follows:

[0086] 1: Use 8 neighbor transition counts As the core indicator for topology determination, the determination rules are as follows:

[0087] For endpoint pixels, including the two endpoints of the main trunk and the burr endpoints of the lateral branches;

[0088] The middle pixel of the main stem belongs to the continuous path pixel of the main stem of the seed bud;

[0089] The pixels at the branching nodes represent the nodes where the main trunk and side branches intersect.

[0090] 2: Initialize the deletion round mask matrix and record the pixel deletion level for each round; batch-remove all skeleton endpoints round by round until the number of skeleton endpoints is less than or equal to 2, at which point all side branches and burrs are removed, resulting in a preliminary branchless backbone skeleton.

[0091] 3: For example Figure 5 As shown, due to the iterative removal, the pixels at the end of the main stem are mistakenly deleted, resulting in an underestimation of the root length. Starting from the two ends of the main stem, the neighboring pixels are traversed in reverse order from high to low according to the deletion rounds, and the mistakenly deleted pixels at the end of the main stem are restored layer by layer to restore the true skeleton of the seed bud.

[0092] 4. The precise root length is calculated using a direction-weighted method. The number of adjacent pixel pairs in four categories—horizontal, vertical, main diagonal, and secondary diagonal—is counted separately, with weights assigned to the horizontal and vertical directions. Diagonal weight The precise pixel length of the root system's main framework is obtained by summing the data. Then, the actual physical length of the seed bud is calculated based on the conversion ratio K (pixels per centimeter) between the physical length and the pixels. The calculation method is as follows:

[0093] Formula 4;

[0094] Formula 5;

[0095] In the formula, The total pixel length of the main skeleton of the seed bud. This refers to the actual physical length corresponding to the main framework of the seed bud. This represents the number of horizontally adjacent pixel pairs. This represents the number of vertically adjacent pixel pairs. The number of adjacent pixel pairs on the main diagonal. This represents the number of adjacent pixel pairs on the secondary diagonal.

[0096] Step 5: During the seed germination information measurement process, it is necessary to calculate the germination information indicators and the output of the visualization image, specifically including:

[0097] 1. The formula for calculating the germination rate is as follows:

[0098] Formula 6;

[0099] The formula for average seedling length is as follows:

[0100] Formula 7:

[0101] In the formula, GR is the seed germination rate. Total number of germinated seeds This represents the total number of seeds in the petri dish. The average physical length of the seed bud. Let be the physical length of the germinated seed (the length of the ungerminated seed is 0).

[0102] 2. Overlay the seed detection box and seed coat / bud segmentation mask onto the original image to generate a mask annotation map; based on the purified main framework, further overlay the bud centerline to generate a framework annotation map (e.g., ...). Figure 6 As shown in the figure, it is convenient for laboratory personnel to visually verify the test results.

[0103] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for measuring seed germination parameters based on dual YOLO and topological skeleton extraction, characterized in that, Includes the following steps: S1. Seed Dual-Level Instance Segmentation and Detection: Construct a two-level instance segmentation and detection model architecture, perform parallel inference on the collected panoramic images of seeds in petri dishes, and output the global mask of the overall seed target and the fine-grained local mask of the seed coat and bud structure inside the seed, forming a dual-level detection mode of global coarse localization segmentation and local fine structure segmentation. S2. Mask association matching and germination status determination: A mask matching strategy is adopted to establish a one-to-one mapping between the global mask of the seed and the local masks of the seed coat and the germination of the seed bud, thereby determining the germination status of a single seed. S3. Refinement of the budding region skeleton: In step S2, the skeleton of the budding region mask that is determined to be a germinating seed is refined, and the budding region with a certain width is shrunk to a single-pixel width and a topologically continuous central skeleton line. S4. Main skeleton purification and seed bud length calculation: Based on the neighborhood topology features of the skeleton pixels, the initial central skeleton is purified, the lateral branch burrs are removed and the end pixels of the main trunk are restored, and the actual physical length of the seed bud is calculated based on the purified main skeleton. S5. Germination index statistics and visualization output: Statistically analyze the germination status and seed bud length of all seeds, calculate the core indicators of germination rate and average seed bud length, and generate a visualization image with superimposed detection and labeling information for output.

2. The seed germination parameter measurement method based on dual YOLO and topological skeleton extraction as described in claim 1, characterized in that, Step S1 includes the following steps: S1-1. Use image acquisition equipment with a fixed position, and agree on the shooting height, focal length and image resolution to acquire panoramic images of the culture dish; S1-2, Based on homologous seed images, and labeled with two types of datasets: The first type is a coarse-grained labeled dataset, which only labels the complete outline of the planted individual without distinguishing the internal structure and germination status; The second category is fine-grained labeled datasets, which separate the labeling of seed coat and seed bud into two independent structural regions; S1-3. Train a first-level baseline segmentation model based on a coarse-grained labeled dataset, and infer the localization results, counting results, and global mask of all seeds in the petri dish. S1-4. A two-level subdivision segmentation model is trained based on a fine-grained labeled dataset. The model infers and outputs two types of fine-grained local masks: seed coat mask and seed bud mask. The seed coat region and seed bud region are isolated to eliminate the interference of the seed coat on subsequent length measurement.

3. The seed germination parameter measurement method based on dual YOLO and topological skeleton extraction as described in claim 1, characterized in that, The S2 step specifically includes the following: S2-1. The mask coverage is used as the matching evaluation index for the two-level mask. The formula for calculating the mask coverage is as follows: In the formula, For mask coverage; This represents the total number of pixels in the intersection region of the primary seed mask and the secondary local mask (seed coat / bud). This represents the total number of pixels in the second-level local mask itself. Differential matching thresholds are set for seed coat and seed bud respectively, with the seed coat matching threshold set to 0.7 and the seed bud matching threshold set to 0.6; when the mask coverage is greater than the corresponding threshold, the two sets of masks are considered to be successfully matched. S2-2. A two-round greedy pairing strategy is used to perform the matching. The first round completes the pairing of the global planting mask and the seed coat mask; the second round completes the pairing of the global seed mask and the seed bud mask. In each round, the mask coverage is paired in descending order. Masks that have been occupied are selected and marked in sequence to ensure that there are no duplicates in the corresponding matches. S2-3. In step S2-2, if the global mask of a single seed matches both the seed coat mask and the bud mask, the seed is determined to be a germinated seed; if only the seed coat mask is matched, the seed is determined to be an ungerminated seed.

4. The seed germination parameter measurement method based on dual YOLO and topological skeleton extraction as described in claim 1, characterized in that, Step S3 employs a parallel thinning algorithm to perform skeleton thinning processing, specifically including: S3-1, Using a 3×3 neighborhood as the basic decision unit, with the current pixel to be detected... Centered on the target, the binary states of 8 neighboring pixels are extracted in a clockwise order, and two types of core decision parameters are defined: the total number of foreground pixels in the neighborhood. Number of background-to-foreground transitions The calculation formula is as follows: S3-2. Each iteration includes two symmetrical sub-scanning stages, setting deletion constraints for boundary pixels in the southeast and northwest directions respectively. Pixels that simultaneously satisfy the constraints of the number of neighboring foreground pixels, the number of jumps, and the directional boundary are marked as deletable boundary pixels. After all scans in a single round are completed, all marked pixels are deleted in batches. S3-3. Iterate and refine the algorithm repeatedly until there are no more pixels to delete in a single iteration. When the algorithm converges, it outputs a single-pixel continuous central skeleton that is consistent with the topology of the original seed bud region.

5. The seed germination parameter measurement method based on dual YOLO and topological skeleton extraction according to claim 1, characterized in that, Step S4 specifically includes: S4-1: Use 8 neighborhood transition times As the core indicator for topology determination, the determination rules are as follows: Pixels with a jump count of 1 are determined to be endpoint pixels, pixels with a jump count of 2 are determined to be middle pixels of the main branch, and pixels with a jump count of not less than 3 are determined to be branch node pixels. S4-2: Initialize the deletion round mask matrix to record the pixel deletion level of each round, and remove all skeleton endpoint pixels in batches until the number of remaining skeleton endpoints is no more than 2, thus obtaining the preliminary trunk skeleton with side branches and burrs removed. S4-3: Starting from the two endpoints of the initial main skeleton, traverse the neighboring pixels in reverse order from high to low according to the deletion level, restore the mistakenly deleted end pixels of the main skeleton layer by layer, and restore the true main skeleton of the seed bud. S4-4: The total pixel length of the main skeleton is calculated using the directional weighting method, and then the actual physical length of the seed bud is obtained by converting the pixel to physical length ratio.

6. The seed germination parameter measurement method based on dual YOLO and topological skeleton extraction according to claim 1, characterized in that, Step S5 specifically includes: In step S5, the formula for calculating the seed germination rate is as follows: In the formula, GR is the seed germination rate; This represents the total number of germinated seeds. This represents the total number of seeds in the petri dish; The formula for calculating the average seedling length is: In the formula, The average physical length of the seed bud; Let be the physical length of the germinated seed (the length of the ungerminated seed is 0).

7. The seed germination parameter measurement method based on dual YOLO and topological skeleton extraction according to claim 6, characterized in that, Seed detection boxes, seed coat segmentation masks, and bud segmentation masks are overlaid on the original seed image to generate a mask annotation map; purified bud main skeleton lines are overlaid on the original seed image to generate a skeleton annotation map. Two types of visual annotation images are output for result verification.

8. An automatic seed germination information measuring device, characterized in that, The device includes an image acquisition module, a two-level segmentation and inference module, a mask matching and state determination module, a skeleton refinement module, a main skeleton purification and length calculation module, and an index calculation and visualization output module; the device is used to execute the automatic seed germination information measurement method according to any one of claims 1 to 6.

Images