Seed root length identification method based on cloud network and seed germination index detection system

Abstract

The invention relates to the field of image processing and fertilizer decomposition degree detection, and relates to a seed root length identification method and a seed germination index detection system based on a cloud network. The method comprises the following steps: S1, acquiring an original to-be-detected image; s2, establishing a target detection model; s3, preprocessing the original to-be-detected image; s4, detecting the preprocessed image by using a target detection model; s5, determining a scale factor; s6, establishing a key point detection model; s7, obtaining a heat map of a single seed image; s8, calculating the physical root length through the heat map and the scale factor; and S9, displaying the root length. The seed germination index detection system comprises a seed germination kit, a visual data processing unit and a seed root length recognition device. According to the invention, convenient uploading of seed images and real-time feedback of identification results are supported; the problems of large result difference and poor repeatability in the existing determination are solved; the standardization of seed germination index detection is realized, the accuracy is improved, and the use is more efficient and convenient.

Description

Technical Field

[0001] This invention relates to the fields of image processing and fertilizer maturity detection, specifically to a seed root length identification method and a seed germination index detection system based on cloud networks. Background Technology

[0002] Maturity is an important parameter for evaluating fertilizer product quality, comprehensively characterizing the content of phytotoxic substances in various fertilizers. Compost maturity assesses the degree to which the organic matter in compost has reached stability after microbial mineralization and humification during the composting process. Applying uncomposted compost to farmland produces phytotoxic substances such as ammonia and organic acids, and also generates heat during degradation, causing seedling burn and other adverse effects on seed germination, plant growth, and the soil environment.

[0003] The evaluation of compost maturity is crucial for its safe application in farmland. Currently, many methods can assess compost maturity, including traditional physical indicators (such as temperature, odor, and viscosity), chemical indicators (such as pH, conductivity, carbon-to-nitrogen ratio, and cation exchange capacity), and biological indicators (such as respiration rate and seed germination index), as well as emerging spectroscopic indicators and computer vision. Among these, the seed germination test is a sensitive, authoritative, and core biological indicator for evaluating compost maturity. Therefore, the standardization and repeatability of the seed germination test measurement method directly affect the reliability of the evaluation results, and the degree of integration of the detection device directly affects the detection efficiency and data reliability. However, the following problems exist when using this method: the lack of unified seed screening standards leads to poor comparability of experimental data due to differences in germination rate and stress resistance between different batches or types of seeds; in traditional manual sowing methods, seed positioning on filter paper relies on manual operation, which is easily affected by liquid flow or plant movement, causing seeds to clump together. This not only causes local differences in moisture and seed-extract contact efficiency but also leads to poor repeatability of GI values ​​due to root length measurement deviations. In existing technologies, Chinese utility model patent CN201967320U improved the seed germination container by using absorbent cotton strips to achieve continuous water supply; Chinese utility model patent CN201420205Y focused on optimizing the filter paper used for seed germination, using cotton pulp to enhance the water retention of the filter paper (water absorption rate > 400%). However, these improvements mainly focus on optimizing local materials and fail to achieve system integration. In existing integrated inventions, the combination method of seeds and filter paper has significant defects. The multi-layer seed paper proposed in Chinese invention application publication number CN105493693A, although achieving visual sowing through printing plant atlases on the upper paper layer, contains synthetic materials such as polyacrylamide in its adhesive system (biodegradation period > 180 days), which conflicts with the ecological compatibility requirements of compost testing. Most integrated inventions press the seeds inside the seed paper, and the filter paper needs to be broken during seed germination, resulting in the grown seeds getting tangled with the filter paper, increasing the difficulty of the experiment. Furthermore, obtaining the seed germination index requires measuring the seed root length. However, various plant root analyzers exist on the market, but they all require specific scanners and software. Therefore, using existing plant root analyzers to measure the seed root length as the seed germination index is costly and complex.

[0004] Therefore, as a frequently used indicator, the seed germination index currently requires a seed root length identification method and seed germination index detection system that can improve the seed germination rate and is easy to use. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the present invention aims to provide a cloud-based seed root length identification method and a seed germination index detection system. Seeds are cultured using an improved seed germination kit, and root images are acquired using a mobile device and sent to the cloud. The cloud uses image recognition and neural network technology to measure the root length and send the length back to the mobile device.

[0006] Specifically, on the one hand, the present invention provides a seed root length identification method based on a cloud network, which includes the following steps: S1. Obtain the original image to be tested: The original image to be tested is an image without trapezoidal distortion that includes the seed and the scale markings; S2. Establish an object detection model: The object detection model takes a square image as input and outputs detection boxes for various detection categories, including detection boxes with hairs and seeds, detection boxes without hairs and seeds, detection boxes with a baseline scale, and detection boxes with a preset scale. Data augmentation strategies are used to train the object detection model, resulting in a trained object detection model. S3. Preprocess the original image to be tested: Maintain the aspect ratio of the original image to be tested and process it into a square image required by the target detection model. S4. Use the target detection model to detect the preprocessed image: Input the preprocessed image into the trained target detection model to identify the areas of seeds with hairs, seeds without hairs, the reference scale, and the preset scale in the preprocessed image; crop the detection boxes of seeds with hairs to obtain a single seed image; for the detection boxes of seeds without hairs, directly return the root length as 0mm. S5. Determine the scale ratio factor: Based on the detected scale detection box, obtain the scale ratio factor λ between the pixel and the physical size; S6. Establish a key point detection model: The keypoint detection model takes a seed image as input and outputs a heatmap. K is the number of keypoints set, each keypoint occupies one channel, and each channel outputs a Gaussian probability heatmap of the keypoint's location. After training, the trained keypoint detection model is obtained. S7. Obtain the heatmap of a single seed image: Input the single seed image obtained in step S4 through the seed detection box with root hair into the key point detection model trained in step S6; output K heatmaps. S8. Calculate the root length using heatmaps, specifically: extract key point coordinates for each heatmap using subpixel interpolation; connect the key point coordinates of the K heatmaps sequentially according to the output K channels to form a polyline P; obtain the pixel value of the total length of polyline P, and obtain the physical root length based on the pixel value and the scale factor λ. S9, Display root length.

[0007] Preferably, S1 specifically comprises: S11. Place the seeds to be tested on the background board without overlap. Place a ruler on the background board where there are no seeds. The ruler and the seeds to be tested should be on the same plane. The background board should be clean and free of interference. Use a ruler with a length of 100mm or more. The ruler should have 0cm and 10cm markings and a minimum division of 1mm. S12. The original image without trapezoidal distortion is acquired under the condition that the mobile terminal is parallel to the background plate. The original image to be tested contains the seed to be tested and the ruler. The original image to be tested has a pixel size of at least 3000×4000 pixels. The original image to be tested can contain multiple seeds to be tested.

[0008] Preferably, in step S2, only the detection box with the highest confidence is retained as the output for the reference scale detection box and the preset scale detection box, and non-maximum suppression is used to obtain the output for the root hair seed detection box and the root hairless seed detection box.

[0009] Preferably, the target detection model training phase in S2 employs the following composite loss function: ; in, To achieve a complete intersection and comparison loss, The binary cross-entropy classification loss is used. For confidence loss, They are respectively , and The weighting coefficients.

[0010] Preferably, S5 specifically includes: Based on the center pixel coordinates of the reference scale detection box and the preset scale detection box, respectively (x0, y0) and (x... 10 , y 10 ), calculate the pixel distance between the two: ; The scale factor is: [ ; Where λ is the scale factor between pixels and physical size. This refers to the actual physical dimension between the reference scale and the preset scale.

[0011] Preferably, the loss function of the key point detection model in S6 adopts a weighted Gaussian kernel focus loss: ; in, This represents the keypoint loss, where K represents the total number of keypoints and k represents the keypoint index. It is the predicted heatmap output by the neural network for the k-th key point; the range is between 0 and 1; α is the focusing parameter; It is a positive sample indicator function; During the training of the keypoint detection model, data augmentation techniques are used, including enabling vertical and horizontal flipping of the input seed image.

[0012] Preferably, step S6 further includes updating the key point detection model, specifically by: using the collected user error correction maps monthly, employing the elastic weight solidification regularization method, freezing the first 1 / 3 layers of the backbone network, and fine-tuning the detection head and key point detection head through training.

[0013] Preferably, S8 specifically includes: S81. Extract key point coordinates for each heatmap using sub-pixel interpolation. The key point coordinates are: ; Where i represents the row index of the heatmap pixel, and j represents the column index of the heatmap pixel. This represents the pixel value at position (i, j) in the heatmap. To iterate through the summation sign of all pixel positions on the heatmap, and These are the key points obtained by interpolation, specifically the horizontal and vertical coordinates on the heatmap scale. S82. Connect the key point coordinates of the K heatmaps in the order of the output K channels to form a broken line P; S83, the pixel value for obtaining the total length of polyline P is: ; S84. The physical root length is obtained as follows: ; in, Let λ be the length of a single physical root, and λ be the scale factor between the pixel and the physical size. This represents the pixel value of the total length of the polyline P.

[0014] On the other hand, the present invention provides a seed germination index detection system based on a cloud network-based seed root length identification method, comprising: a seed germination kit, a visual data processing unit, and a seed root length identification device. The seed germination kit is used for seed cultivation, the visual data processing unit is used to upload acquired seed images and receive identification results, and the seed root length identification device is used to identify the seed root length based on the uploaded seed images and send the identification results, as detailed below: The seed germination kit adopts a layered composite structure, which includes a ternary integrated system of sterile petri dishes, qualitative filter paper and standard seeds. Qualitative filter paper is placed in sterile petri dishes, and seeds are precisely fixed on the surface of qualitative filter paper in a dotted circle arrangement using biomimetic plant glue through precision micro-dotting technology, forming a seed-carrier composite structure. The visual data processing unit includes an image acquisition module, a terminal transceiver module, and a result display and error correction module. The image acquisition module is used to acquire the original image to be tested with trapezoidal distortion. The image acquisition module obtains the original image to be tested by directly importing or calling the camera to take pictures. The terminal transceiver module is used to transmit data with the cloud transceiver module through a communication network, including uploading the original image to be tested and the image after error correction, and receiving information sent by the cloud transceiver module and then transmitting it to the result display and error correction module. The result display and error correction module is used to display the results returned by the cloud and for users to correct errors. The seed root length recognition device is located in the cloud and includes a preprocessing module, a scale factor determination module, a target detection model module, a key point detection model module, a root length calculation module, and a cloud transceiver module. The preprocessing module processes the original image into a square preprocessed image and sends it to the target detection model module. The target detection model module inputs the preprocessed image and identifies seeds with root hairs, seeds without root hairs, and the regions containing the reference scale and preset scale. The target detection model module then sends all identified regions to the cloud transceiver module and sends the identified reference scale and preset scale regions to... The scale factor determination module sends the root hair seed region to the keypoint detection model module; the scale factor determination module obtains the pixel-to-physical size scale factor based on the reference scale region and the preset scale region, and sends the scale factor to the root length calculation module; the keypoint detection model module inputs the root hair seed region and obtains K heatmaps in channel order, and sends the K heatmaps to the root length calculation module; the root length calculation module obtains the keypoint coordinates and physical root length based on the received heatmaps and scale factor, and sends the keypoint coordinates and root length to the visual data processing unit through the cloud transceiver module.

[0015] Preferably, the standard seeds are white radish seeds with a diameter of 3 to 5 millimeters, and 10 white radish seeds are contained in one sterile petri dish.

[0016] Preferably, the seed germination kit is vacuum-sealed in groups of 10 and stored at 4°C in the dark.

[0017] Preferably, the result display and error correction module specifically performs the following steps: upon receiving the areas containing root hair seeds, rootless seeds, reference scales, and preset scales, it marks and displays the identified areas on the original image using borders, and displays the root length as 0 for the rootless seed area; upon receiving the key point coordinates and root length, it connects the key point coordinates of K heatmaps sequentially according to channel order to form a polyline P, renders the key points and polyline on the root hair seeds based on the original image scale, and displays the root length; when the user zooms in to check the scale alignment, the corrected image data can be obtained by adjusting the key points.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The seed germination kit proposed in this invention overcomes the technical contradiction between transportation fixation and culture release by using biomimetic plant glue; it uses the same seed species, which improves the biological consistency of the experimental seeds to 97.8%; it realizes the standardization and reconstruction of the operation process, compressing the operation time of the traditional method from 25±5min to 2±0.5min, and reducing manual intervention by more than 90%.

[0019] 2. This invention only requires taking a picture and uploading the seed image to identify the seed root length. It does not require professional identification equipment and skills, and its simple operation can reduce costs.

[0020] 3. This invention integrates image acquisition, processing, analysis, and result display by cooperating with the terminal transceiver module and the cloud, thus solving problems such as inconvenience for users and difficulty in data traceability.

[0021] 4. The method proposed in this invention is highly adaptable. When the seed type, germination stage or environmental conditions change, the detection accuracy is not significantly affected. This invention provides a high-precision, high-efficiency, and highly automated intelligent identification system and method for seed germination and root length. Attached Figure Description

[0022] Figure 1 This is a sample display of the original image to be tested in an embodiment of the present invention; Figure 2 This is the loss curve of the training set (train) and the validation set (val) at 300 epochs in an embodiment of the present invention; Figure 3 This is the loss curve of the target detection box during training at 300 epochs in this embodiment of the invention; Figure 4 This is the loss curve of the target detection box at 300 epochs during verification in this embodiment of the invention; Figure 5 This is the loss curve for classification during training at 300 epochs in this embodiment of the invention; Figure 6This is the loss curve for classification at 300 epochs during verification in this embodiment of the invention; Figure 7 This is the loss curve with a confidence level of 300 epochs during training in this embodiment of the invention; Figure 8 This is the loss curve with a confidence level of 300 epochs during verification in this embodiment of the invention; Figure 9 This is the curve showing that mAP@0.5 reaches 98.1% when the epoch is set to 300 in this embodiment of the invention; Figure 10 This is a schematic diagram of the results obtained by the target detection model in an embodiment of the present invention; Figure 11 This is a schematic diagram illustrating the actual recognition effect of 17 key points of root hair in an embodiment of the present invention. Figure 12 This is a flowchart of a seed root length identification method based on a cloud network according to the present invention; Figure 13 This is a schematic diagram of the seed germination kit in an embodiment of the present invention. Detailed Implementation

[0023] To fully explain the technical content, objectives, and effects of this invention, the embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0024] This invention discloses a seed root length identification method based on cloud networks, such as... Figure 12 As shown, it includes the following steps: S1. Obtain the original image to be tested; The original image to be tested is a clear image without trapezoidal distortion, containing the seed and ruler markings.

[0025] S11. Place the seeds to be tested on the background plate without overlap. Place a ruler on the seedless part of the background plate, ensuring the ruler and the seeds are on the same plane. The background plate should be clean and free of interference. The background plate can typically be the surface of a petri dish or white paper. Use a ruler with a length of 100mm or more, including 0cm and 10cm graduations, with a minimum division of 1mm.

[0026] S12. An original image without trapezoidal distortion acquired under the condition that the mobile device is parallel to the background panel, such as... Figure 1 As shown, the original image to be tested contains the seed to be tested and a ruler, and the original image to be tested has a pixel size of at least 3000×4000 pixels to ensure that the minimum 1 mm scale on the 0 cm and 10 cm rulers in the original image to be tested is clearly distinguishable; the image format is preferably jpg or png; the original image to be tested may contain multiple seeds to be tested; the mobile terminal is preferably a mobile phone or tablet computer.

[0027] S2. Establish a target detection model; In this embodiment, the target detection model is based on YOLOv11. The input to the target detection model is a 1920×1920 square image, and the output is detection boxes of various detection categories: seed boxes with hairs (seeds), seed boxes without hairs (seeds_nohair), baseline scale detection boxes (ruler_0), and preset scale detection boxes (ruler_10). In a preferred embodiment, the baseline scale detection box (ruler_0) is the 0-scale detection box, and the preset scale detection box is the 10-centimeter detection box. The baseline scale (ruler_0) and preset scale (ruler_10) detection boxes are referred to as scale detection boxes, and the seed boxes with hairs (seeds) and seed boxes without hairs (seeds_nohair) are referred to as seed detection boxes. Only the detection boxes with the highest confidence are retained as output for the scale detection boxes. Non-maximum suppression (NMS) is applied to the seed detection boxes to obtain N seed detection boxes {b_i}_{i=1}^N. In this embodiment, the threshold is set to 0.4.

[0028] When training the object detection model, data augmentation strategies are employed. For example, in this embodiment, images in the training set are subjected to mosaic blurring, copy-paste, simulated perspective within 2°, random rotation from 0-360°, scaling by ±30%, shearing, and adjustments to hue, saturation, and brightness (hsv_h, hsv_s, hsv_v). However, this embodiment prohibits vertical or horizontal flipping of the dataset for augmentation to avoid errors caused by reversed scale characters.

[0029] The object detection model uses the following composite loss function during the training phase: ; in, For the Complete Intersection over Union (IoU) loss, The binary cross-entropy classification loss is used. For confidence loss, They are respectively , and The weighting coefficients. Using a training dataset of 692 actual photographs (approximately 10 seeds per image), after 300 epochs of model training, the best model from the training process was selected, achieving mAP@0.5 of 98.1%. mAP@0.5 (mean Average Precision at IoU=0.5) refers to the average precision calculated when the Intersection over Union (IoU) threshold is 0.5. Figure 2 An example is shown showing the loss curves of the training set (train) and the validation set (val) when the epoch is set to 300. Figures 3-8 The example demonstrates the changes in the curves of the object detection box, classification, and confidence score when the epoch is set to 300. Figure 9 An example is shown where mAP@0.5 reaches 98.1% when the epoch is set to 300.

[0030] The object detection model is trained to obtain the trained object detection model.

[0031] S3. Preprocess the original image to be tested; Maintaining the aspect ratio of the original image, the longest side of the original image is shrunk to 1920 pixels, resulting in a 1920×1440 image. The image is then padded to a 1920×1920 square using center padding (fill value R=G=B=114). The padded area will be cropped after subsequent processing, therefore it does not affect the physical scale conversion.

[0032] S4. Use an object detection model to detect preprocessed images; The preprocessed image is input into the trained target detection model to identify the areas containing seeds with hairs, seeds without hairs, the baseline scale, and the preset scale in the preprocessed image.

[0033] If neither the reference scale nor the preset scale is detected simultaneously in the input preprocessed image, the preprocessed image is rotated 90° and the detection is repeated, up to a maximum of 3 times. If neither the reference scale nor the preset scale is detected simultaneously again, the preprocessed image is deemed invalid.

[0034] The seed detection bounding box with root hairs is cropped to obtain a single seed image.

[0035] For the seed detection frame without root hairs, directly execute S9 to return a root length of 0mm.

[0036] The results obtained from the preprocessed image after detection by the object detection model are as follows: Figure 10 As shown.

[0037] S5. Determine the scale factor; Based on the detected scale detection box, the conversion ratio between pixels and physical size is obtained, as follows: Based on the center pixel coordinates (x0, y0) of the reference scale detection box and the preset scale detection box, and (x 10 , y 10 ), calculate the pixel distance between the two: .

[0038] The scale factor is: .

[0039] This λ value will serve as a global scale factor for all subsequent physical length conversions. In this embodiment, the reference scale detection frame is the 0-scale detection frame, and the preset scale detection frame is the 10-centimeter detection frame. The length is 100mm.

[0040] S6. Establish a key point detection model; In this embodiment, the keypoint detection model is based on YOLOv11. The input of the keypoint detection model is a 640×640×3 (RGB) seed image, and the downsampling output is a 96×96×K heatmap (Gaussian image). Each heatmap is 96×96 in size, and K is the number of keypoints. Each keypoint occupies one channel, for a total of K channels. Each channel outputs a Gaussian probability heatmap of the keypoint's location. In this embodiment, K is preferably 17.

[0041] The loss function of the keypoint detection model uses a weighted Gaussian kernel focus loss: ; in, This represents the keypoint loss, where K represents the total number of keypoints and k represents the keypoint index. This is the predicted heatmap (i.e., the k-th layer channel) output by the neural network for the k-th keypoint. Each pixel value in the heatmap represents the confidence probability that the location is a keypoint, ranging from 0 to 1. The closer the value is to 1, the more likely the network considers the point to be a keypoint. α is the focusing parameter, a hyperparameter greater than 0 (usually α=1, but can also be 2 or 3). Its function is to further reduce the loss contribution of samples that have been well predicted by the model (high confidence) and to allow the model to focus more on learning those samples that are difficult to predict and have low confidence. It is part of the standard cross-entropy loss. It measures the difference between the predicted probability distribution and the true distribution. This is the positive sample indicator function, which is a mask and crucial for applying the Gaussian kernel. It's not a simple 0 / 1 label, but typically a Gaussian heatmap.

[0042] In this embodiment, α=3 and the standard deviation σ=1.5 pixels.

[0043] After training is completed, a trained keypoint detection model is obtained.

[0044] During the training of the keypoint detection model, data augmentation techniques can be used: based on the characteristic that the keypoint structure is not sensitive to mirror images, additional vertical flipping (flipud=0.5) and horizontal flipping (fliplr=0.5) can be enabled on the input seed image.

[0045] Each month, user-corrected images are collected, and the Elastic Weight Fixed (EWC) regularization method is used to freeze the first 1 / 3 of the backbone network, with only the detection head and keypoint detection head being fine-tuned. Typically, no fewer than 500 user-corrected images are used each time, and retraining takes about 2 hours.

[0046] S7. Obtain the heatmap of a single seed image; The single seed image obtained in step S4 through the seed detection box with hair is scaled up to 640×640 and input into the key point detection model trained in step S6; K heatmaps are output.

[0047] S8. Calculate the root length; S81. For each heatmap, extract the key point coordinates through subpixel interpolation. This is equivalent to finding the "centroid" of all pixel positions on the heatmap through weighted averaging, and using the centroid coordinates as the final predicted coordinates of the key points.

[0048] ; Where i represents the row index of the heatmap pixel, and j represents the column index of the heatmap pixel. Represents the pixel value at position (i, j) in the heatmap; this is a probability value ranging from 0 to 1. Indicates the network's confidence that the key point appears in this grid cell (or its vicinity). Iterate through the summation symbols for all pixel positions on the heatmap. and It is the horizontal and vertical coordinate value of the key point obtained by interpolation on the heatmap scale. It is a normalized value with decimals, which is convenient for conversion back to the original map scale later.

[0049] S82. Connect the key point coordinates of the K heatmaps in the order of the output K channels to form a broken line P.

[0050] S83, the pixel value for obtaining the total length of polyline P is: .

[0051] S84. The physical root length is obtained as follows: .

[0052] S9, Display root length; After all the root hair seed detection boxes in S4 have been processed according to steps S6-S8, the physical root length of each root hair seed is obtained. The image within each root hair seed detection box and the obtained root length are displayed for both root hair seed detection boxes and rootless seed detection boxes. Figure 11 As shown.

[0053] In this embodiment, the mean absolute error (MAE) on the test set is 0.05 mm, which meets the accuracy requirements for agronomic measurements (≤0.1 mm).

[0054] This invention also discloses a seed germination index detection system based on a cloud network, such as... Figure 13 As shown, it specifically includes: a seed germination kit, a visual data processing unit, and a seed root length recognition device. The seed germination kit is used for seed cultivation; the visual data processing unit is used to upload acquired seed images and receive recognition results; and the seed root length recognition device is used to identify the seed root length based on the uploaded seed images and send the recognition results, as detailed below: The seed germination kit uses a layered composite structure, such as... Figure 13 As shown, this is a ternary integrated system comprising sterile petri dishes, qualitative filter paper, and standard seeds. Qualitative filter paper is placed in the sterile petri dish, and seeds are precisely fixed to the surface of the filter paper in a dotted circular pattern using a biomimetic plant adhesive through precision micro-dispensing technology, forming a seed-carrier composite structure. The sterile petri dish has a diameter of 90 mm and a height of 15 mm, conforming to GB / T 4548 standard. The biomimetic plant adhesive has excellent adhesion properties, ensuring seed stability during transportation and preventing displacement. During the addition of the extract and transfer from the experimental platform to the incubator, the seeds remain evenly distributed on the filter paper. The adhesive dissolves rapidly upon contact with water, ensuring complete dissolution of the adhesive upon transfer to the incubator, allowing the seeds to detach naturally, thus avoiding interference with the seed germination process and ensuring normal seed development and activity. White radish seeds are selected, and through rigorous screening, plump and high-quality seeds are chosen, with the selection range limited to seeds with a diameter of 3 to 5 mm to ensure seed uniformity and stable germination performance. Typically, each sterile petri dish contains 10 white radish seeds. The seed germination kit is vacuum-sealed in groups of 10 and can be stably stored for 12 months at 4°C in the dark. The seed germination kit uses a biomimetic plant adhesive, overcoming the technical contradiction between transport fixation and culture release; it uses uniformly screened white radish seeds, ensuring a germination rate of ≥80%; furthermore, it enables standardized operating procedures, reducing the traditional 25±5 min operation time to 2±0.5 min, and reducing manual intervention by more than 90%.

[0055] The visual data processing unit includes an image acquisition module, a terminal transceiver module, and a result display and error correction module. The visual data processing unit is located on a mobile phone, tablet, or computer. The image acquisition module acquires the original image of the trapezoidal distortion. This can be done by directly importing or capturing the original image using a camera. The original image must be a clear image without trapezoidal distortion, containing the seeds cultured using the seed germination kit and the scale markings on a ruler. Specifically, the seeds cultured from the seed germination kit are placed on a background plate, with the ruler and the seeds on the same plane. The background plate must be clean and free of interference. The ruler must include 0cm and 10cm markings, with a minimum division of 1mm. The original image must have a minimum resolution of 3000×4000 pixels. The image acquisition module sends the acquired original image to the terminal transceiver module. The terminal transceiver module transmits data with the cloud transceiver module via a communication network. The terminal transceiver module preferably uses the WeChat mini-program wx.uploadFile to upload the original image in segments. Each image in the WeChat mini-program wx.uploadFile is 512KB, and the average size of the original image is 3.2MB, so the upload time on a 4G network is approximately 0.8 seconds. The terminal transceiver module receives the areas containing root-hair seeds, root-hairless seeds, the reference scale, and the preset scale, and then sends them to the result display and error correction module. The terminal transceiver module receives the keypoint coordinates and root length sent by the cloud transceiver module and sends them to the result display and error correction module. The result display and error correction module displays the results returned from the cloud and allows for user error correction. After receiving the areas containing root-hair seeds, root-hairless seeds, the reference scale, and the preset scale, the result display and error correction module marks the identified areas on the original image with borders and displays them, showing a root length of 0 for the root-hairless seed area. After receiving the keypoint coordinates and root length, the result display and error correction module connects the keypoint coordinates of K heatmaps sequentially according to channel order to form a polyline P, renders the keypoints and polyline on the root-hair seed based on the original image scale, and displays the root length. Rendering can be implemented using canvas. Users can zoom in to check the scale alignment and adjust key points to send back the corrected image data.

[0056] The seed root length recognition device is located in the cloud and includes a preprocessing module, a scale factor determination module, a target detection model module, a key point detection model module, a root length calculation module, and a cloud transceiver module. The preprocessing module processes the original image into a square preprocessed image and sends it to the target detection model module. The preprocessing method for the original image in this embodiment is as described in step S3. The target detection model module inputs the preprocessed image and identifies seeds with root hairs, seeds without root hairs, the reference scale, and the area containing the preset scale. The target detection model module sends all identified areas to the cloud transceiver module, and sends the identified reference scale and preset scale areas to the scale factor determination module, and sends the areas of seeds with root hairs to the key point detection model module. The specific recognition process of the target detection model module is as described in step S4. The scale factor determination module obtains the pixel-to-physical-size conversion ratio based on the reference scale area and the preset scale area, and sends the conversion ratio to the root length calculation module. The specific process is as described in step S5. The key point detection model module takes the root hair seed region as input and obtains K heatmaps in channel order. These K heatmaps are then sent to the root length calculation module; in this embodiment, K is 17. The root length calculation module obtains the key point coordinates and physical root length based on the received heatmaps and the conversion ratio, as detailed in step S8. The key point coordinates and root length are then sent to the visual data processing unit via the cloud transceiver module.

[0057] Using the cloud-based seed root length identification method and seed germination index detection system of this application, the following performance was achieved on 10,000 test images collected in real shooting scenarios: scale detection success rate: 100%; seed classification accuracy with / without root hairs: 99.2%; average absolute error of root length measurement: 0.05mm, maximum error: 0.18mm; end-to-end processing time per image: average 112ms.

[0058] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the invention. For example, in the target detection stage, other target detection algorithms such as Yolov8 and Faster R-CNN can be used instead of the Yolov11 algorithm. Although there may be differences in detection speed or accuracy, the detection functions of seed and scale bar can still be achieved. In the keypoint detection stage, CNN-based or Transformer-based keypoint detection algorithms can be used instead of Yolov11-based keypoint detection algorithms to adapt to different detection needs and computational resource conditions. Therefore, without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by those skilled in the art should fall within the protection scope defined by the claims of the present invention.

Claims

1. A seed root length identification method based on cloud networks, characterized in that, It includes the following steps: S1. Obtain the original image to be tested: The original image to be tested is an image without trapezoidal distortion that includes the seed and the scale markings; S2. Establish a target detection model: The input of the target detection model is a square image, and the output is a detection box of various detection categories. The detection box includes a seed detection box with a root hair, a seed detection box without a root hair, a reference scale detection box, and a preset scale detection box. Data augmentation strategies are used to train the model, resulting in a trained target detection model. S3. Preprocess the original image to be tested: Maintain the aspect ratio of the original image to be tested and process it into a square image required by the target detection model. S4. Use the target detection model to detect the preprocessed image: Input the preprocessed image into the trained target detection model to identify the areas of seeds with hairs, seeds without hairs, the reference scale, and the preset scale in the preprocessed image; crop the detection boxes of seeds with hairs to obtain individual seed images; for the detection boxes of seeds without hairs, directly give the root length as 0mm. S5. Determine the scale factor: Based on the detected detection box, obtain the scale factor λ between the pixel and the physical size; S6. Establish a key point detection model: The keypoint detection model takes a seed image as input and outputs a heatmap. K is the number of keypoints set, each keypoint occupies one channel, and each channel outputs a Gaussian probability heatmap of the keypoint's location. After training, the trained keypoint detection model is obtained. S7. Obtain the heatmap of a single seed image: Input the single seed image obtained in step S4 through the seed detection box with root hair into the key point detection model trained in step S6; output K heatmaps. S8. Calculate the root length. Extract key point coordinates for each heatmap using subpixel interpolation. Connect the key point coordinates of the K heatmaps sequentially according to the K output channels to form a polyline P. Obtain the pixel value of the total length of polyline P. Calculate the physical root length based on the pixel value and the scale factor λ. S9, Display root length.

2. The seed root length identification method based on cloud network according to claim 1, characterized in that, Specifically, S1 is: S11. Place the seed to be tested flat on the background board, and place a ruler on the background board with the ruler and the seed to be tested on the same plane. S12. Acquire an original image without trapezoidal distortion under the condition that the mobile terminal is parallel to the background. The original image to be tested contains the seed to be tested and the ruler, and the original image to be tested has a pixel size of at least 3000×4000 pixels; the original image to be tested contains multiple seeds to be tested.

3. The seed root length identification method based on cloud network according to claim 1, characterized in that, In step S2, only the detection box with the highest confidence is retained as the output for the reference scale detection box and the preset scale detection box, and non-maximum suppression is used to obtain the output for the root hair seed detection box and the root hairless seed detection box.

4. The seed root length identification method based on cloud network according to claim 1, characterized in that, The target detection model in S2 uses a composite loss function during the training phase. ； in, To achieve a complete intersection and comparison loss, The binary cross-entropy classification loss is used. For confidence loss, They are respectively , and The weighting coefficients.

5. The seed root length identification method based on cloud network according to claim 1, characterized in that, Specifically, S5 is: Based on the center pixel coordinates of the reference scale detection box and the preset scale detection box, respectively (x0, y0) and (x... 10 ,y 10 ), calculate the pixel distance between the reference scale detection box and the preset scale detection box: ； The scale factor is: [ ； Where λ is the scale factor between pixels and physical size. This refers to the actual physical dimension between the reference scale and the preset scale.

6. The seed root length identification method based on cloud network according to claim 1, characterized in that, The loss function of the key point detection model in S6 is a weighted Gaussian kernel focus loss: ； in, This represents the keypoint loss, where K represents the total number of keypoints and k represents the keypoint index. It is the predicted heatmap output by the neural network for the k-th key point; the range is between 0 and 1; α is the focusing parameter; It is a positive sample indicator function; During the training of the keypoint detection model, data augmentation strategies are employed, including enabling vertical and horizontal flipping of the input seed image.

7. The seed root length identification method based on cloud network according to claim 1, characterized in that, Specifically, S8 is: S81. Extract key point coordinates for each heatmap using sub-pixel interpolation. The key point coordinates are: ； Where i represents the row index of the heatmap pixel, and j represents the column index of the heatmap pixel. This represents the pixel value at position (i, j) in the heatmap. To iterate through the summation sign of all pixel positions on the heatmap, and These are the horizontal and vertical coordinates of the key points obtained by interpolation on the heatmap scale; S82. Connect the key point coordinates of the K heatmaps in the order of the output K channels to form a broken line P; S83, the pixel value for obtaining the total length of polyline P is: ； S84. The physical root length is obtained as follows: ； in, Let λ be the length of a single physical root, and λ be the scale factor between the pixel and the physical size. This represents the pixel value of the total length of the polyline P.

8. A seed germination index detection system utilizing the cloud network-based seed root length identification method according to any one of claims 1-7, characterized in that, It includes: The system comprises a seed germination kit, a visual data processing unit, and a seed root length recognition device. The seed germination kit is used for seed culture; the visual data processing unit uploads acquired seed images and receives recognition results; and the seed root length recognition device identifies the seed root length based on the uploaded seed images and sends the recognition results. Specifically: The seed germination kit adopts a layered composite structure, which includes a ternary integrated system of sterile petri dishes, qualitative filter paper and standard seeds. Qualitative filter paper is placed in sterile petri dishes, and seeds are precisely fixed on the surface of qualitative filter paper in a dotted circle arrangement using biomimetic plant glue through precision micro-dotting technology, forming a seed-carrier composite structure. The visual data processing unit includes an image acquisition module, a terminal transceiver module, and a result display and error correction module. The image acquisition module is used to acquire the original test image with trapezoidal distortion. The image acquisition module obtains the original test image by directly importing or calling the camera to take pictures. The terminal transceiver module is used to transmit data with the cloud transceiver module through a communication network, including uploading the original image to be tested and the corrected image, receiving information sent by the cloud transceiver module and then transmitting it to the result display and error correction module; the result display and error correction module is used to display the results returned by the cloud and for users to correct errors. The seed root length recognition device is located in the cloud and includes a preprocessing module, a scale factor determination module, a target detection model module, a key point detection model module, a root length calculation module, and a cloud transceiver module. The preprocessing module processes the original image into a square preprocessed image and sends it to the target detection model module. The target detection model module inputs the preprocessed image and identifies seeds with root hairs, seeds without root hairs, and the regions containing the reference scale and preset scale. The target detection model module then sends all identified regions to the cloud transceiver module and sends the identified reference scale and preset scale regions to... The scale factor determination module sends the root hair seed region to the keypoint detection model module; the scale factor determination module obtains the pixel-to-physical size scale factor based on the reference scale region and the preset scale region, and sends the scale factor to the root length calculation module; the keypoint detection model module inputs the root hair seed region and obtains K heatmaps in channel order, and sends the K heatmaps to the root length calculation module; the root length calculation module obtains the keypoint coordinates and physical root length based on the received heatmaps and scale factor, and sends the keypoint coordinates and root length to the visual data processing unit through the cloud transceiver module.

9. The seed germination index detection system according to claim 8, characterized in that, The standard seeds are white radish seeds with a diameter of 3 to 5 millimeters, and 10 white radish seeds are contained in one sterile petri dish.

10. The seed germination index detection system according to claim 8, characterized in that, A ruler is a ruler with a length of 100 mm or more, with 0 cm and 10 cm markings on the surface, and the smallest division is 1 mm.

Images