A crop ear three-dimensional kernel phenotype analysis method and system based on weakly supervised neural radiance field

Abstract

The present application relates to the field of plant phenotype analysis, and provides a crop ear three-dimensional kernel grain phenotype analysis method and system based on weakly supervised neural radiance field. The method comprises: kernel grain surface labeling on an ear image sequence to obtain a labeled data set; inputting the labeled data set into a semantic segmentation model for training to obtain a kernel grain surface segmentation model, so as to infer a multi-view image sequence to obtain a kernel grain surface binary mask set of multiple views; taking the binary mask set as a weakly supervised signal to perform segmentation perception fine-tuning on a neural radiance field model to obtain a kernel grain surface three-dimensional volume representation; extracting a high-density three-dimensional point cloud from the kernel grain surface three-dimensional volume representation, and performing hierarchical clustering segmentation on the high-density three-dimensional point cloud to obtain a kernel grain instance point cloud set; and calculating a total number of kernel grains and a kernel grain row number. The present application realizes sparse labeling driven three-dimensional reconstruction, and realizes automatic counting of kernel grains and analysis of the row number of kernel grains without causing damage.

Description

Technical Field

[0001] This invention relates to the field of plant phenotypic analysis technology, and in particular to a method and system for analyzing the three-dimensional nucleosome phenotype of crop ears based on a weakly supervised neural radiation field. Background Technology

[0002] The number of kernels and the number of kernel rows in maize ears are core phenotypic traits for evaluating yield potential and are of great value in breeding screening, physiological assessment, and agronomic research. Traditional manual counting relies on counting each kernel after shelling, which is cumbersome and difficult to scale. With the development of computer vision technology, kernel counting methods based on two-dimensional images have been proposed, including estimating the number of kernel rows and kernels per row using single-view photographs, image analysis based on cylindrical projection, and object detection and classification methods combined with deep learning. In recent years, neural radiation field technology has provided a new technical approach for reconstructing high-fidelity three-dimensional scenes from ordinary RGB images and has been initially applied in plant phenotypic tasks such as fruit counting.

[0003] However, existing methods still face significant bottlenecks in maize sclerotium phenotypic analysis: sclerotium on the ear surface are densely arranged on curved surfaces, and a large number of sclerotium are obscured from any single viewpoint, making it impossible to observe them. Therefore, two-dimensional imaging methods cannot obtain complete information on the entire ear of sclerotium. While panoramic unfolding can partially alleviate the obstruction, it introduces geometric distortion, and sclerotium in the apex and basal regions are difficult to include in the analysis. In addition, although existing three-dimensional reconstruction methods can obtain the overall morphology, they lack the ability to decompose hundreds of closely contacting sclerotium on a continuous curved surface into independent instances, making instance-level sclerotium segmentation impossible in three-dimensional space. Summary of the Invention

[0004] This invention provides a method and system for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field, in order to overcome the shortcomings of the prior art.

[0005] The first aspect of this invention provides a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field, comprising:

[0006] S1. Perform nucleogranule surface annotation on the multi-view image sequence of the ear of fruit to obtain a sparse nucleogranule surface annotation dataset; S2. Input the sparse nucleogranule surface annotation dataset into the semantic segmentation model for training to obtain the nucleogranule surface segmentation model, so as to perform reasoning on the multi-view image sequence and obtain a set of binary masks for the nucleogranule surface from multiple views. S3. Using the set of binary masks on the nucleogranule surface as a weak supervision signal, the neural radiation field model is segmented and fine-tuned to obtain a three-dimensional volume representation of the nucleogranule surface. S4. Extract high-density three-dimensional point cloud from the three-dimensional volume representation of the nucleoparticle surface, and perform hierarchical clustering and segmentation on the high-density three-dimensional point cloud to obtain a set of nucleoparticle instance point clouds. S5. Based on the nucleus-particle instance point cloud set, calculate the total number of nucleus-particles and the number of nucleus-particle rows respectively.

[0007] According to the present invention, a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field is provided, wherein step S1 further includes: S11. Select representative images from multiple viewing directions from the multi-view image sequence of the ear of fruit as images to be labeled. S12. Perform binary mask annotation on the visible nucleus-particle outer surface region in the image to be annotated to form the original surface annotation sample; S13. Apply geometric transformation, color perturbation and texture distortion operations to the original surface annotation sample to generate an amplified sample. Combine the amplified sample with the original surface annotation sample to form a sparse nucleogranular surface annotation dataset.

[0008] According to the present invention, a method for analyzing the three-dimensional nucleosome phenotype of crop ears based on a weakly supervised neural radiation field is provided, wherein step S2 further includes: S21. Divide the sparse nucleogranule surface labeled dataset into a training subset and a validation subset according to a preset ratio; S22. Based on the segmentation network with encoder-decoder structure as the basic model, the weighted combination of pixel-level classification loss and region overlap loss is used as the joint loss function to optimize the parameters of the basic model, retain the model weight with the best verification performance, and obtain the nucleus-particle surface segmentation model. S23. During the inference phase, the probability map output by the nucleogranular surface segmentation model is upsampled to the original image resolution to obtain a full-resolution nucleogranular surface binary mask, and a set of nucleogranular surface binary masks from multiple perspectives is output.

[0009] According to the present invention, a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field is provided, wherein step S3 further includes: S31. The neural radiation field model is pre-trained using the RGB appearance information of multi-view image sequences to obtain the RGB radiation field. S32. Based on the RGB radiation field, and with the binary mask set on the nucleosome surface replacing RGB as the target, the radiation field parameters of the neural radiation field model are continuously optimized to obtain the three-dimensional volume representation of the nucleosome surface.

[0010] According to the present invention, a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field is provided, wherein step S4 further includes: S41. The three-dimensional volume representation of the nucleus surface is filtered according to a density threshold, and sampling points with a density value not less than a preset threshold are retained to obtain a first-stage filtered point cloud. S42. Remove points in the first-stage filtered point cloud whose RGB brightness is less than the preset brightness to obtain a second-stage filtered point cloud. S43. Calculate the principal axis direction of the secondary screening point cloud through principal component analysis, align the principal axis to the global vertical axis direction, and translate the midpoint of the ear axis to the origin of the coordinate system to obtain the attitude-normalized point cloud. S44. Perform density clustering on the attitude-normalized point cloud using a density clustering algorithm to obtain an initial nucleo-granular cluster set; S45. For excessively large clusters in the initial nucleogranule cluster set that are larger than a preset volume, extract high-brightness subsets and continue to optimize and subdivide them until all clusters meet the single nucleogranule volume constraint, and obtain the nucleogranule instance point cloud set.

[0011] According to the present invention, a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field is provided. In step S45, the method for extracting the high-brightness subset further includes: S451. Select the points whose brightness values ​​within the current large cluster are located in the first preset percentile interval to form a high-brightness subset. The points in the high-brightness subset correspond to the nucleus-particle ridge line and the strongly irradiated surface region. S452. Using the high-brightness subset as input, perform secondary density clustering, map the obtained sub-cluster labels back to all points of the current oversized cluster, complete the current sub-subdivision, and obtain the high-brightness subset.

[0012] According to the present invention, a method for analyzing the three-dimensional nucleosome phenotype of crop ears based on a weakly supervised neural radiation field is provided. In step S5, the total number of nucleosomes is obtained by counting the total number of clusters in the nucleosome instance point cloud set. In step S5, the method for calculating the nucleoparticle row count further includes: S51. Extract the middle segment point cloud from the nucleogranule instance point cloud set, where the vertical axis direction is located below the second preset percentile interval, as the middle segment analysis region of the spikelet; S52. Divide the mid-segment analysis region of the spike into multiple thin slices along the longitudinal axis. Construct a two-dimensional occupancy histogram for the cross-sectional coordinates of each point in each slice. After Gaussian smoothing and peak ratio threshold binarization, calculate the radial nucleogranule row number of the current thin slice by connecting component labeling. S53. Take the median of the radial nut row count of all thin-layer cuts as the nut row count of the ear.

[0013] A second aspect of this invention provides a three-dimensional nucleogranular phenotypic analysis system for crop ears based on a weakly supervised neural radiation field, comprising: Annotation module: used to annotate the nucleoplast surface of multi-view image sequences of ear of fruit to obtain a sparse nucleoplast surface annotation dataset; Training module: used to input the sparse nucleogranular surface annotation dataset into the semantic segmentation model to train the semantic segmentation model and obtain the nucleogranular surface segmentation model; The inference module is configured as the nucleogranular surface segmentation model obtained by the training module, and is used to infer the multi-view image sequence to obtain a set of binary masks for the nucleogranular surface from multiple views. Fine-tuning module: used to perform segmentation and fine-tuning of the neural radiation field model using the set of binary masks on the nucleosome surface as a weak supervision signal, to obtain a three-dimensional volume representation of the nucleosome surface; Segmentation module: used to extract high-density three-dimensional point cloud from the three-dimensional volume representation of the nucleogranule surface, and to perform hierarchical clustering segmentation on the high-density three-dimensional point cloud to obtain a set of nucleogranule instance point clouds; Calculation module: used to calculate the total number of kernels and the number of kernel rows based on the kernel instance point cloud set.

[0014] A third aspect of this invention provides a three-dimensional nucleosome phenotypic analysis device for crop ears based on a weakly supervised neural radiation field, comprising: A memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause a crop ear three-dimensional kurtoid phenotype analysis device based on a weakly supervised neural radiation field to perform a crop ear three-dimensional kurtoid phenotype analysis method based on a weakly supervised neural radiation field as described in any of the preceding claims.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field as described in any of the preceding claims.

[0016] This invention employs a method of selecting only a small number of representative images from a multi-view image sequence for binary mask annotation of the nucleoplasmic outer surface. Combined with data augmentation operations such as geometric transformation, color perturbation, and texture distortion, a sparsely labeled dataset is constructed. This significantly reduces the manual annotation workload, fundamentally avoiding the burden of dense per-view annotation or voxel-level labeling in traditional 3D segmentation processes. This makes the method feasible for low-cost and rapid deployment in practical breeding and phenotypic screening scenarios. Secondly, the nucleoplasmic surface segmentation model trained with pixel-level classification loss and region overlap joint loss can stably generalize on diverse samples with significant differences in brightness conditions, shooting angles, and nucleoplasmic morphology. The output full-resolution binary mask provides a reliable weak supervision signal for subsequent 3D reconstruction. Furthermore, during the segmentation-aware fine-tuning stage, this invention continuously accumulates high density in the nucleoplasmic region of the neural radiation field while simultaneously suppressing background density. This allows for the formation of a coherent nucleoplasmic volume representation in 3D space based solely on sparse 2D supervision of the visible surface, overcoming the inherent limitation of traditional methods where incomplete whole-ear nucleoplasmic information is caused by occlusion. The high-density point cloud obtained by density threshold screening and brightness filtering is normalized by principal component analysis. Then, a hierarchical density clustering strategy is adopted to extract the high-brightness subsets of excessively large clusters and continuously optimize the subdivision. This strategy enables reliable separation of closely contacting adjacent nucleoids that are highly overlapping in two-dimensional projection. The resulting nucleoid instance point cloud set directly supports the automatic counting of the number of nucleoids in the whole ear. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A schematic flowchart of a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field is provided in this embodiment of the invention. Figure 2 This is a schematic diagram illustrating an example of an amplified image generated from an original labeled sample through a data amplification operation, provided in an embodiment of the present invention. Figure 3 A schematic diagram of a three-dimensional nucleogranular phenotype analysis system for crop ears based on a weakly supervised neural radiation field is provided for an embodiment of the present invention. Figure 4 A schematic diagram of the three-dimensional reconstruction result of COB-9 provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the visualization result of a corn ear after three-dimensional nucleoid instance segmentation provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the visualization results of radial nucleus row estimation from a cross-sectional slice of the middle section of the ear provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] The embodiments of the present invention are described below with reference to the figures.

[0021] like Figure 1 As shown, this invention provides a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field, comprising: S1. Perform nucleoplast surface annotation on the multi-view image sequence of the ear of fruit to obtain a sparse nucleoplast surface annotation dataset.

[0022] Step S1 further includes: S11. Select representative images from multiple viewing directions from the multi-view image sequence of the ear of fruit as images to be labeled.

[0023] In step S1, the present invention first performs a handheld 360° surround shot of each corn ear, covering the complete field of view from the top to the base. Then, image frames are extracted from the video at a rate of 4 frames per second using ffmpeg, resulting in 90 to 190 images per ear. Each image frame is then subjected to 2x spatial downsampling to form a multi-view image sequence.

[0024] In step S11, the present invention randomly selects three representative images covering different observation directions from all the images of the above-mentioned ear of fruit, corresponding to the top view, horizontal view and bottom view respectively, as images to be labeled. When selecting, it is ensured that the three images have obvious differences in the shooting direction, so that the labeling information can cover different areas of the ear of fruit surface and lighting conditions.

[0025] S12. Perform binary mask annotation on the visible nucleus-particle outer surface region in the image to be annotated to form the original surface annotation sample.

[0026] In step S12, for the above three images to be labeled, the present invention only manually outlines the visible nucleus outer surface area in each image. The labeling result is a binary mask image with the same size as the original image. The pixel where the nucleus outer surface is located is assigned a value of 1, and the background and non-nucleus areas such as the rachis are assigned a value of 0. In addition, the internal area of ​​the nucleus that is hidden is not inferred during labeling. Only the visually visible nucleus surface outline is recorded. After the three images are labeled, three sets of image-mask pairs are formed, which are the original surface labeling samples.

[0027] S13. Apply geometric transformation, color perturbation and texture distortion operations to the original surface annotation sample to generate an amplified sample. Combine the amplified sample with the original surface annotation sample to form a sparse nucleogranular surface annotation dataset.

[0028] In step S13, using the three sets of original surface annotation samples as input, data amplification operations are applied to each set of samples. Specifically, the geometric transformations include random rotation, flipping, scaling, and cropping, which are applied synchronously to the image and the corresponding mask to ensure that the pixel-level correspondence is not destroyed; color perturbation randomly adjusts the brightness, contrast, saturation, and hue of the image, and the mask does not participate in this operation; texture distortion changes the local texture structure of the image through elastic deformation or mesh perturbation, and the same deformation field is applied to the corresponding mask. After amplification, each set of original samples generates 40 amplified variants, and a total of 120 amplified images and corresponding masks are generated from the three sets. Subsequently, the 120 amplified samples are merged with the original three sets of samples to form a total of 123 image-mask pairs, i.e., the sparse nucleogranular surface annotation dataset.

[0029] Specifically, Figure 2 An example of an amplified image generated from the original labeled samples through a data amplification operation is shown. Figure 2 The top row contains six amplified RGB images of the ear of fruit, while the bottom row contains the corresponding binary mask images of the nucleus surface. The white areas correspond to the marked areas on the outer surface of the nucleus, and the black areas represent the background. These six examples cover the appearance of the ear of fruit under different colors, scales, and poses, intuitively demonstrating the effects of color perturbation and geometric transformation on the image appearance. At the same time, the mask and the image maintain a pixel-by-pixel correspondence.

[0030] S2. Input the sparse nucleogranule surface labeled dataset into the semantic segmentation model for training to obtain the nucleogranule surface segmentation model, so as to perform inference on the multi-view image sequence and obtain a set of binary masks for the nucleogranule surface from multiple views.

[0031] Step S2 further includes: S21. Divide the sparse nucleogranule surface labeled dataset into a training subset and a validation subset according to a preset ratio.

[0032] In step S21, the present invention randomly divides the above 123 sets of image-mask pairs into two groups at a ratio of 80% and 20%, with about 98 groups constituting the training subset and about 25 groups constituting the verification subset. There is no image overlap between the two subsets during the division.

[0033] S22. Based on the segmentation network with encoder-decoder structure as the basic model, the model is optimized by using a weighted combination of pixel-level classification loss and region overlap loss as the joint loss function. The model weights with the best verification performance are retained to obtain the nucleo-particle surface segmentation model.

[0034] Furthermore, the basic model adopts the U-Net architecture, with the encoder being a ResNet-34 pre-trained on ImageNet, responsible for extracting multi-scale feature maps of the image layer by layer. The decoder fuses the feature maps of each layer of the encoder with the upsampled features through skip connections, gradually restoring the spatial resolution, and finally outputting a pixel-level prediction probability map of the same size as the input image.

[0035] During training, the joint loss function is a weighted sum of binary cross-entropy loss and Dice loss with the same weight coefficient of 1: the binary cross-entropy loss calculates the logarithmic error between the predicted probability and the ground truth label pixel by pixel, while the Dice loss calculates the deviation of the region overlap ratio between the predicted mask and the ground truth mask. The optimizer used for training is Adam, with an initial learning rate of 1×10⁻⁶. -4 When the validation loss stops decreasing for several consecutive epochs, the learning rate is automatically reduced by 10%; when the validation loss does not improve for 100 consecutive epochs, training terminates and the model weights corresponding to the lowest validation loss are saved as the final nucleogranular surface segmentation model.

[0036] S23. During the inference phase, the probability map output by the nucleogranular surface segmentation model is upsampled to the original image resolution to obtain a full-resolution nucleogranular surface binary mask, and a set of nucleogranular surface binary masks from multiple perspectives is output.

[0037] Furthermore, during inference, this invention inputs each multi-view image of each ear of fruit into the nucleus surface segmentation model. The model outputs a pixel-level probability map corresponding to the downsampling resolution for each image, where the value of each pixel represents the predicted probability that the location belongs to the outer surface of the nucleus. Subsequently, bilinear interpolation upsampling is performed on the probability map to restore the resolution to the original image size at the time of shooting. Then, the upsampled probability map is binarized with a fixed threshold, and pixels with probabilities higher than the threshold are assigned a value of 1, while the rest are assigned a value of 0, resulting in a full-resolution binary mask for the nucleus surface. After performing the above operations sequentially on all image frames of each ear of fruit, the binary masks of all frames are summarized to form a set of binary masks for the nucleus surface of the ear of fruit.

[0038] S3. Using the set of binary masks on the nucleogranule surface as a weak supervision signal, the neural radiation field model is segmented and fine-tuned to obtain a three-dimensional volume representation of the nucleogranule surface.

[0039] Step S3 further includes: S31. The neural radiation field model is pre-trained using the RGB appearance information of multi-view image sequences to obtain the RGB radiation field.

[0040] Furthermore, the neural radiation field model adopts the Nerfacto architecture, with the hidden layer dimension expanded from the default 32 to 128. Before training, this invention first performs motion recovery structure calculation on the multi-view image sequence of each ear of fruit using COLMAP. COLMAP extracts feature points from the images and performs cross-view matching. Through triangulation and bundle adjustment, it jointly solves the camera intrinsic and extrinsic parameters corresponding to each image, and outputs the camera pose matrix of each image.

[0041] Subsequently, Nerfstudio's coordinate normalization tool unified all camera poses to a standardized coordinate system. Following this, the present invention used the normalized camera poses and corresponding images as input to pre-train the Nerfacto model in RGB for 2000 epochs. Each iteration used 5120 rays as a batch, and the learning rate for both the radiance network and the proposal network was set to 1×10⁻⁶. - ², the camera parameter fine-tuning learning rate is set to 4×10. -4 Each of the three parameter sets uses an independent Adam optimizer, with ε set to 1×10. -8 The weights decay to 0, and finally, the RGB radiation field is obtained after training.

[0042] S32. Based on the RGB radiation field, and with the binary mask set on the nucleosome surface replacing RGB as the target, the radiation field parameters of the neural radiation field model are continuously optimized to obtain the three-dimensional volume representation of the nucleosome surface.

[0043] In step S32, based on the RGB radiation field weights, this invention replaces the training target from RGB pixel values ​​to a nucleus surface binary mask. Specifically, for each sampled ray, the model integrates the volume density along the ray direction and outputs the predicted probability that the pixel corresponding to that ray belongs to the nucleus surface. Then, it calculates the segmentation loss with the corresponding label of that pixel in the nucleus surface binary mask set.

[0044] During the fine-tuning phase, the learning rate of each optimizer is reduced by two orders of magnitude compared to the RGB pre-training phase, with the learning rate of the radiation field network and the proposal network set to 1×10. -4 The camera parameter learning rate is set to 4×10. -6The other optimizer hyperparameters remain unchanged. During the fine-tuning process, the model continuously accumulates high volume density for the light corresponding to the nucleus region and continuously reduces the volume density for the light corresponding to the background region. In the final output three-dimensional volume representation, the high-density region spatially corresponds to the nucleus distribution on the surface of the ear, i.e., the three-dimensional volume representation of the nucleus surface.

[0045] S4. Extract high-density three-dimensional point cloud from the three-dimensional volume representation of the nucleoparticle surface, and perform hierarchical clustering and segmentation on the high-density three-dimensional point cloud to obtain a set of nucleoparticle instance point clouds.

[0046] Step S4 further includes: S41. The three-dimensional volume representation of the nucleus surface is filtered according to a density threshold, and sampling points with a density value not less than a preset threshold are retained to obtain a first-stage filtered point cloud.

[0047] Furthermore, the three-dimensional volume representation of the nucleoparticle surface is stored in the form of discrete sampling points, each sampling point containing three types of attributes: three-dimensional spatial coordinates, volume density value, and RGB color value. In step S41, during point cloud export, the present invention reads the density value of each volume sampling point along each sampling ray, discarding sampling points with a density value lower than 70 (NeRF units) and retaining only those with a density value not lower than this threshold. Simultaneously, the three-dimensional spatial range is defined by an axis-aligned bounding box, and sampling points in the background area outside the bounding box are also removed. After the above two filtering processes, the remaining sampling points constitute a filtered point cloud, where each point still carries three-dimensional coordinates, density value, and RGB color value.

[0048] S42. Points with RGB brightness less than a preset brightness in the first-stage filtered point cloud are removed to obtain the second-stage filtered point cloud.

[0049] In step S41, each point in the primary point cloud carries an RGB color value. In step S42, the present invention compares the three channels of this color value (R, G, B) with a brightness threshold of 0.65. Points with all three channel values ​​below 0.65 are identified as low-brightness points and are discarded. Low-brightness points spatially correspond to shaded areas on the ear surface or areas with poor reconstruction quality. The remaining points after discarding these points constitute a secondary point cloud, where each point still retains its three-dimensional coordinates and RGB color value attributes.

[0050] S43. Calculate the principal axis direction of the secondary screened point cloud through principal component analysis, align the principal axis to the global vertical axis direction, and translate the midpoint of the ear axis to the origin of the coordinate system to obtain the attitude-normalized point cloud.

[0051] In step S43, the present invention constructs a point coordinate matrix from the three-dimensional coordinates of all points in the secondary-selected point cloud, calculates the covariance matrix of this matrix, and then performs eigenvalue decomposition on the covariance matrix to obtain three eigenvectors and their corresponding eigenvalues. The eigenvector with the largest eigenvalue is the principal axis direction of the point cloud, corresponding to the longitudinal extension direction of the ear of fruit. Subsequently, the present invention performs axis permutation on the point cloud coordinates, rotating and aligning the principal axis direction to the global z-axis direction.

[0052] After completing axis alignment, calculate the maximum and minimum values ​​of the point cloud in the z-axis direction, take the average of the two as the z-coordinate of the midpoint of the ear axis, and translate the midpoint to the origin of the coordinate system. That is, apply the same translation amount to the coordinates of all points so that z=0 corresponds to the longitudinal center position of the ear axis, thus obtaining the attitude-normalized point cloud.

[0053] S44. Perform density clustering on the attitude-normalized point cloud using a density clustering algorithm to obtain an initial nucleo-granular cluster set.

[0054] For the DBSCAN spatial clustering algorithm, the core parameters include the neighborhood radius and the minimum number of points, MinPts: For each point in the point cloud, if its neighborhood radius contains no fewer than the minimum number of points, MinPts, then the point is determined to be a core point; all core points and their neighboring points that are density-reachable within each other's neighborhood radius are grouped into the same cluster; points that do not meet the density conditions are marked as noise points and discarded. In step S44, the present invention inputs the pose-normalized point cloud into DBSCAN, sets the neighborhood radius to 0.0035, and sets MinPts to 25. After input, DBSCAN performs density clustering on all points in the point cloud, outputs the cluster number to which each point belongs, and summarizes them to obtain the initial nucleo-granular cluster set.

[0055] S45. For excessively large clusters in the initial nucleogranule cluster set that are larger than a preset volume, extract high-brightness subsets and continue to optimize and subdivide them until all clusters meet the single nucleogranule volume constraint, and obtain the nucleogranule instance point cloud set.

[0056] Furthermore, in the initial nucleogranule cluster set, adjacent nucleogranules that are in contact with each other often merge into a single oversized cluster after one density clustering. In step S45, the present invention counts the number of points in each cluster in the initial nucleogranule cluster set, and determines clusters with more than 200 points as oversized clusters, and performs the following subdivision process on them: first, execute S451 to extract a high-brightness subset, then execute S452 to perform a second density clustering on the subset and map the sub-cluster labels back to all points of the oversized cluster; after the subdivision is completed, check the number of points again for all newly generated sub-clusters. If there are still sub-clusters with more than 200 points, repeat the above subdivision process for the sub-cluster, for a maximum of three recursive levels. After all clusters meet the point count constraint, all clusters are summarized to form a nucleogranule instance point cloud set.

[0057] In step S45, the method for extracting the high-brightness subset further includes: S451. Select the points whose brightness values ​​within the current large cluster are located in the first preset percentile interval to form a high-brightness subset. The points in the high-brightness subset correspond to the nucleus-particle ridge line and the strongly irradiated surface region.

[0058] In step S451, the brightness values ​​of all points within the current large cluster are sorted, and the brightness values ​​are calculated from the average of the RGB three channels. This invention selects points whose brightness values ​​fall within the 25th to 60th percentile range, i.e., the remaining mid-to-high brightness points after removing the darkest 25% and brightest 40% points, forming a mid-to-high brightness subset. Points within this range are spatially concentrated at the ridge boundaries between adjacent nuclei and on the surface of nuclei receiving stronger light. These regions geometrically correspond to the boundary positions of different nuclei, and the spatial spacing of the point distribution is relatively obvious.

[0059] S452. Using the high-brightness subset as input, perform secondary density clustering, map the obtained sub-cluster labels back to all points of the current oversized cluster, complete the current sub-subdivision, and obtain the high-brightness subset.

[0060] In step S452, the present invention uses the three-dimensional coordinates of all points in the high-brightness segment subset obtained in S451 as input, and performs DBSCAN clustering again, with parameters consistent with the first density clustering in S44. During execution, DBSCAN outputs the sub-cluster number for each point in the high-brightness segment subset. Subsequently, the present invention maps the sub-cluster numbers back to all points in the current oversized cluster. Each point in the high-brightness segment subset has obtained a sub-cluster number, and the remaining points not in the subset are assigned to the nearest sub-cluster according to the spatial nearest neighbor principle, so that all points in the oversized cluster obtain the sub-cluster affiliation label after subdivision, thus completing this subdivision.

[0061] S5. Based on the nucleus-particle instance point cloud set, calculate the total number of nucleus-particles and the number of nucleus-particle rows respectively.

[0062] In step S5, the total number of kernel particles is obtained by counting the total number of clusters in the kernel particle instance point cloud set.

[0063] The total number of nuclei obtained by this invention is directly calculated from the total number of clusters in the nuclei instance point cloud set, that is, the sum of the number of numbers of all valid clusters in the set is used as the predicted value of the number of nuclei of the ear.

[0064] In step S5, the method for calculating the nucleoparticle row count further includes: S51. Extract the middle segment point cloud from the nucleogranule instance point cloud set, where the vertical axis is located below the second preset percentile interval, as the middle segment analysis region of the spikelet.

[0065] For step S51, this invention statistically analyzes the z-axis coordinates of all points in the nucleus-granule instance point cloud set, calculating the 25th and 75th percentile values ​​of the z-axis coordinates. Subsequently, this invention extracts all points whose z-axis coordinates fall within the interval between the 25th and 75th percentiles, that is, removing the top 25% and bottom 25% of points along the z-axis direction. The remaining points constitute the mid-ear analysis region, which corresponds to the 50% range of the longitudinal center of the ear.

[0066] S52. Divide the mid-segment analysis region of the spike into multiple thin slices along the longitudinal axis. Construct a two-dimensional occupancy histogram for the cross-sectional coordinates of each point in each slice. After Gaussian smoothing and peak ratio threshold binarization, calculate the radial nucleogranule row number of the current thin slice by connecting component labeling.

[0067] In step S52, the present invention divides the z-axis range of the mid-segment analysis region into 10 thin-layer slices, each slice being a z-axis interval. Subsequently, for all points within each slice, the x-axis and y-axis coordinates are extracted, and a two-dimensional occupancy histogram is constructed on a 256×256 two-dimensional grid. Specifically, the present invention divides the xy-plane into 256×256 grid cells, counts the number of points falling into each cell, and forms a two-dimensional density distribution map. Then, a Gaussian smoothing with a standard deviation of 2 is applied to this two-dimensional density distribution map to eliminate local noise, resulting in a smoothed density map. Next, using 30% of the peak value in the smoothed density map as a binarization threshold, grid cells with density values ​​higher than this threshold are assigned a value of 1, and those lower are assigned a value of 0, resulting in a binary occupancy map. Finally, connected component labeling is performed on the binary occupancy map, that is, spatially connected regions with a value of 1 are grouped into the same connected component, and the total number of connected components is counted as the radial nucleogranular row number of the current slice.

[0068] S53. Take the median of the radial nut row count of all thin-layer cuts as the nut row count of the ear.

[0069] In step S53, the present invention summarizes the radial kernel row values ​​calculated by each of the 10 thin-layer slices, arranges the 10 values ​​in ascending order, and takes the average of the middle value, that is, the 5th and 6th values, as the final kernel row estimate of the ear.

[0070] like Figure 2 As shown, the present invention also provides a three-dimensional nucleogranular phenotypic analysis system for crop ears based on a weakly supervised neural radiation field, comprising: Annotation module 100: used to annotate the nucleolus surface of multi-view image sequences of ear of fruit to obtain a sparse nucleolus surface annotation dataset; Training module 200: used to input the sparse nucleogranule surface annotation dataset into the semantic segmentation model to train the semantic segmentation model and obtain the nucleogranule surface segmentation model; The inference module 300 is configured as the nucleogranular surface segmentation model obtained by the training module, and is used to infer the multi-view image sequence to obtain a set of binary masks for the nucleogranular surface from multiple views. Fine-tuning module 400: Used to perform segmentation sensing fine-tuning on the neural radiation field model using the set of binary masks on the nucleosome surface as a weak supervision signal, so as to obtain a three-dimensional volume representation of the nucleosome surface; Segmentation module 500: used to extract high-density three-dimensional point cloud from the three-dimensional volume representation of the nucleus surface, and to perform hierarchical clustering segmentation on the high-density three-dimensional point cloud to obtain a set of nucleus instance point clouds; Calculation module 600: used to calculate the total number of kernel particles and the number of kernel particle rows based on the kernel particle instance point cloud set.

[0071] The present invention also provides a crop ear three-dimensional kurtosis phenotype analysis device based on a weakly supervised neural radiation field, comprising: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the crop ear three-dimensional kurtosis phenotype analysis device based on a weakly supervised neural radiation field to perform a crop ear three-dimensional kurtosis phenotype analysis method based on a weakly supervised neural radiation field as described in any of the above.

[0072] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement a method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field as described in any of the preceding claims.

[0073] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0074] The following describes the method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field, using specific implementation results.

[0075] In one specific embodiment, the RGB neural radiation field model produced high-fidelity three-dimensional reconstruction results for all nine maize ears, with an average peak signal-to-noise ratio of 23.7 dB. As shown in Table 1, the SSIM and LPIPS indices also remained stable. Even under natural and uncontrolled lighting conditions used during data acquisition, the overall ear morphology and fine kernel surface texture were reliably reproduced, such as... Figure 4 As shown in Figure a, the method of the present invention can perform accurate NeRF reconstruction from a simple handheld image.

[0076] Table 1. Quantitative evaluation results of the RGB neural radiation field reconstruction quality of maize ears.

[0077] Table 1 summarizes the quantitative evaluation results of the RGB neural radiation field reconstruction quality of nine maize ears. The evaluation metrics included Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Patch Similarity (LPIPS). The evaluation sample size N was 117 images, derived from approximately 10% of the randomized views of each ear. In Table 1, PSNR measures the pixel-level error between the reconstructed image and the real image, expressed in decibels (dB). A higher PSNR value indicates a smaller pixel difference between the reconstructed and real images. The mean PSNR in Table 1 is 23.72 dB, the median is 23.81 dB, the minimum is 19.80 dB, and the maximum is 28.64 dB. dB; SSIM (Structural Similarity Index) comprehensively measures the perceptual similarity between two images from three dimensions: brightness, contrast, and structure. The value ranges from 0 to 1. The closer the value is to 1, the more similar the two images are in visual structure. In Table 1, the mean SSIM is 0.87, the median is 0.87, the minimum is 0.77, and the maximum is 0.92. LPIPS (Learned Perceptual Patch Similarity) calculates the perceptual distance between two images based on the intermediate layer features of a deep neural network. The lower the value, the closer the two images are in perceptual sense. In the table, the mean LPIPS is 0.13, the median is 0.13, the minimum is 0.082, and the maximum is 0.334.

[0078] Given the inherent limitations of two-dimensional phenotypic analysis, the ability to reconstruct a complete three-dimensional surface is particularly important. The granules on the ear of fruit are densely arranged on a curved surface, and from any single viewpoint, a large number of granules are occluded. Purely two-dimensional methods, including deep learning detectors, cannot resolve this occlusion problem, and therefore cannot completely count the granules of the entire ear. While panoramic projection can partially alleviate occlusion, it introduces image distortion and cannot include the apical and basal regions (where the granule geometry deviates from the cylindrical assumption). The method of this invention, employing a three-dimensional radiation field representation, fundamentally overcomes these limitations by reconstructing the volume of the entire ear surface.

[0079] In addition, in this embodiment, the present invention generates high-quality three-dimensional nucleoid surface representations on all nine ears of fruit by fine-tuning the segmented perceptual neural radiation field, achieving an average cross-union ratio of 0.70, a Dice coefficient of 0.80, and a pixel accuracy higher than 0.95, as shown in Table 2. It is worth noting that the segmentation mask used for training the neural radiation field originates from a single U-Net model trained with annotations on only three representative ears of fruit. These sparse surface annotations, after extensive amplification, enable U-Net to effectively generalize under conditions of significant variations in viewpoint, illumination, ear size, and nucleoid morphology. The prediction mask provides sufficient supervision signals for the neural radiation field, enabling it to learn a coherent three-dimensional nucleoid surface, even though each mask only reveals the visible portion of each nucleoid, such as... Figure 4 As shown in b.

[0080] The annotation efficiency contrasts sharply with traditional supervised 3D segmentation workflows, which typically require dense multi-view annotations or voxel-level labels. In the framework proposed in this invention, the volumetric representation of neural radiation fields allows sparse 2D signals to propagate across viewpoints, significantly reducing the annotation workload while avoiding the subjectivity and efficiency issues inherent in manual annotation.

[0081] Table 2. Quantitative assessment results of segmented sensory neural radiation fields on maize ears.

[0082] Table 2 summarizes the quantitative evaluation results of segmented sensory neural radiation fields on 9 corn ears. The evaluation indicators include crossover ratio (IoU), Dice coefficient and pixel accuracy. The evaluation sample consisted of 117 images, which were also derived from approximately 10% of the random sampled views of each ear.

[0083] In Table 2, IoU (Intersection over Union) is calculated as the ratio of the intersection area to the union area between the predicted mask and the real mask. The value ranges from 0 to 1. The higher the value, the higher the degree of overlap between the predicted mask and the real mask. In Table 2, the mean IoU is 0.6986, the median is 0.7096, the minimum is 0.3884, and the maximum is 0.8092.

[0084] The Dice coefficient, like IoU, measures the degree of overlap between the predicted and ground truth masks. It is calculated by dividing twice the intersection area by the sum of the predicted and ground truth mask areas, and its value ranges from 0 to 1, with higher values ​​indicating better overlap. Compared to IoU, the Dice coefficient is more sensitive to the weight of overlapping regions. In Table 2, the mean Dice value is 0.8033, the median is 0.8296, the minimum is 0.5595, and the maximum is 0.8945.

[0085] Pixel accuracy is calculated as the proportion of correctly classified pixels in the predicted mask out of the total number of pixels, with a value ranging from 0 to 1. In Table 2, the mean pixel accuracy is 0.9528, the median is 0.9705, the minimum is 0.8191, and the maximum is 0.9869.

[0086] Figure 4 The 3D reconstruction results of COB-9 are shown, comprising two sub-images. Sub-image (a) presents the RGB neural radiation field reconstruction results, showcasing the overall 3D morphology of the ear and the surface texture of the kernels. The color, arrangement, and surface relief structure of the kernels are clearly reproduced. The background area is presented as a blue-purple gradient, and the ear is displayed vertically in the center. The support rod at the bottom of the ear axis is also visible in the reconstruction results. Sub-image (b) shows the kernel surface point cloud extracted after segmentation-aware fine-tuning and density threshold filtering. The point cloud is presented as a white point set against a black background. The overall outline of the point cloud highly matches the ear morphology in (a). The point cloud distribution is concentrated on the outer surface area of ​​the kernels, and background noise has been effectively removed. The uneven structure of the kernel surface is reflected in the density distribution of the point cloud.

[0087] Furthermore, the challenges of instance-level nucleosome segmentation differ fundamentally from those of conventional neural radiation field semantic segmentation. Nucleosomes share boundaries, highly overlap in two-dimensional projections, and are only visible locally from most perspectives. Nucleosomes are densely packed along curved surfaces, with extremely small spacing between adjacent nucleosomes, making it impossible for any single image to clearly represent the boundaries of all nucleosomes. These factors make instance-level decomposition far more difficult than target-level segmentation, which has been the primary focus of previous neural radiation field frameworks.

[0088] Despite the aforementioned challenges, in this embodiment, the multi-stage clustering process applied to the surface point cloud derived from the neural radiation field successfully achieved the separation of individual nucleoli on all nine ears of fruit. Figure 5 As shown in Table 3, this process relies on geometric and photometric refinement of the extracted surface, enabling the density clustering algorithm to operate on the pre-processed nucleus shape rather than directly on the original point cloud. This allows physically contact nuclei, often indistinguishable in two-dimensional views, to be reliably segmented into independent clusters. The predicted nucleus count is in high agreement with manual counting, with a mean absolute error of 10.3 nuclei, as shown in Table 3.

[0089] Table 3. Predicted values ​​of kernel number and kernel row number in maize ears.

[0090] Table 3 summarizes the predicted number of kernels and the predicted number of kernel rows for the nine corn ears, compares them with the actual values ​​counted manually, and lists the prediction error for each ear.

[0091] Table 3 lists the predicted and actual values ​​(GT) for kernel number, with the error being the signed difference between the predicted and actual values. The prediction errors for the kernel number of the nine ears are: +9 for COB-a, and [missing value] for COB-b. 5. COB-c is 6. COB-d is 11. COB-e is +4, COB-f is 17. COB-g is +8, COB-h is 8. With COB-i of +25, the mean absolute error (MAE) for predicting the number of kernels on 9 ears is 10.3.

[0092] Regarding the number of nucleus rows, the predicted and actual values ​​are listed one by one, with the row count error being the signed difference. The predicted row count errors for the 9 ear clusters are as follows: COB-a = 0, COB-b = +1, COB-c = 0. 1. COB-d is +1, COB-e is 0, COB-f is 0, COB-g is 0, COB-h is +1, and COB-i is 0. The mean absolute error (MAE) of the row number prediction for the 9 ears is 0.33 rows, and the absolute error of the row number prediction for all ears does not exceed 1 row.

[0093] Figure 5 This displays the visualization results of nine corn ears after three-dimensional nucleogranule instance segmentation. Figure 5 From middle a to Figure 5 In the figure, i corresponds to COB-1 to COB-9 respectively. In each subfigure, the ear point cloud is presented in a horizontal position after PCA pose normalization, and each independent nucleosome instance is assigned a different color for distinction, with a dark blue background. The predicted nucleosome numbers for each subfigure are: (a) 701, (b) 335, (c) 729, (d) 319, (e) 531, (f) 471, (g) 649, (h) 491, and (i) 677. It can be observed from the figure that there are significant differences in the number of nuclei, ear length, nucleosome arrangement density, and morphology among different ears. Some ears exhibit a strong columnar morphology but lack obvious taper. Hierarchical clustering segmentation successfully separated adjacent nuclei into independent color blocks under all the above morphological conditions, and the color boundaries between adjacent nuclei are clearly distinguishable.

[0094] Figure 6 This presentation displays visualizations of radial nucleus row estimation from cross-sectional slices of the midsection of nine ear clusters. Figure 6 From middle a to Figure 6In the image, i corresponds to COB-1 to COB-9 respectively. Each sub-image presents a binary occupancy map of a representative slice from 10 thin-layer slices of the ear, corresponding to the median row number. The image background is black, and the white areas are grid cells identified as nuclei after Gaussian smoothing and peak ratio thresholding. As can be seen from the image, the white areas are discretely distributed along the circumferential direction, forming an approximately ring-shaped arrangement. Each independent white connected region corresponds to the cross-sectional projection of a column of radial nuclei rows. The predicted nuclei row numbers corresponding to each sub-image are: (a) 20, (b) 15, (c) 17, (d) 18, (e) 16, (f) 18, (g) 18, (h) 17, and (i) 18. The number and distribution density of connected regions differ among different ears, reflecting the actual changes in the number of nuclei rows and ear diameter among ears.

[0095] This capability stands in stark contrast to previous neural radiation field segmentation frameworks. Previous frameworks operated at the target level, typically requiring a full target mask or post-processing segmentation based on 2D projection; they were not designed for decomposing densely stacked units or performing sub-organ-level decomposition. The results of this invention demonstrate for the first time that a neural radiation field model finely tuned with an incomplete 2D mask can support hundreds of instance-level segmentations along continuous biological structures.

[0096] The overall performance of the reconstruction, segmentation and clustering stages of this invention demonstrates that fine-scale nucleosome traits can be automatically and non-destructively extracted from simple handheld video images. For whole-ear nucleosome counting, which traditionally relies on manual counting or destructive sampling, the method based on this invention only requires sparse surface labeling and low-cost imaging equipment. This process can maintain stable generalization among samples with significant differences in ear size, morphology, nucleosome morphology and light conditions, indicating that it is suitable for deployment in diverse phenotypic analysis environments.

[0097] These capabilities are of great significance for plant phenotypic analysis. This invention eliminates the need for dense annotation, bypasses the inherent occlusion limitations of two-dimensional imaging, and can acquire sub-organ-level geometric information at a resolution compatible with high-throughput analysis. Because this invention relies on general geometric and photometric cues, rather than morphological features specific to maize, it is easily generalized to other densely packed biological structures, such as wheat spikelets, sunflower heads, grape bunches, or fruit seeds, structures that are difficult to quantitatively assess under traditional two-dimensional imaging conditions.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field, characterized in that, include: S1. Perform nucleogranule surface annotation on the multi-view image sequence of the ear of fruit to obtain a sparse nucleogranule surface annotation dataset; S2. Input the sparse nucleogranule surface annotation dataset into the semantic segmentation model for training to obtain the nucleogranule surface segmentation model, so as to perform reasoning on the multi-view image sequence and obtain a set of binary masks for the nucleogranule surface from multiple views. S3. Using the set of binary masks on the nucleogranule surface as a weak supervision signal, the neural radiation field model is segmented and fine-tuned to obtain a three-dimensional volume representation of the nucleogranule surface. S4. Extract high-density three-dimensional point cloud from the three-dimensional volume representation of the nucleoparticle surface, and perform hierarchical clustering and segmentation on the high-density three-dimensional point cloud to obtain a set of nucleoparticle instance point clouds. S5. Based on the nucleus-particle instance point cloud set, calculate the total number of nucleus-particles and the number of nucleus-particle rows respectively.

2. The method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field according to claim 1, characterized in that, Step S1 further includes: S11. Select representative images from multiple viewing directions from the multi-view image sequence of the ear of fruit as images to be labeled. S12. Perform binary mask annotation on the visible nucleus-particle outer surface region in the image to be annotated to form the original surface annotation sample; S13. Apply geometric transformation, color perturbation and texture distortion operations to the original surface annotation sample to generate an amplified sample, and combine the amplified sample with the original surface annotation sample to form a sparse nucleogranular surface annotation dataset.

3. The method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field according to claim 1, characterized in that, Step S2 further includes: S21. Divide the sparse nucleogranule surface labeled dataset into a training subset and a validation subset according to a preset ratio; S22. Based on the segmentation network with encoder-decoder structure as the basic model, the weighted combination of pixel-level classification loss and region overlap loss is used as the joint loss function to optimize the parameters of the basic model, retain the model weight with the best verification performance, and obtain the nucleus-particle surface segmentation model. S23. During the inference phase, the probability map output by the nucleogranular surface segmentation model is upsampled to the original image resolution to obtain a full-resolution nucleogranular surface binary mask, and a set of nucleogranular surface binary masks from multiple perspectives is output.

4. The method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field according to claim 1, characterized in that, Step S3 further includes: S31. The neural radiation field model is pre-trained using the RGB appearance information of multi-view image sequences to obtain the RGB radiation field. S32. Based on the RGB radiation field, and with the binary mask set on the nucleogranule surface replacing RGB as the target, the radiation field parameters of the neural radiation field model are continuously optimized to obtain the three-dimensional volume representation of the nucleogranule surface.

5. The method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field according to claim 1, characterized in that, Step S4 further includes: S41. The three-dimensional volume representation of the nucleus surface is filtered according to a density threshold, and sampling points with a density value not less than a preset threshold are retained to obtain a first-stage filtered point cloud. S42. Remove points in the first-stage filtered point cloud whose RGB brightness is less than the preset brightness to obtain a second-stage filtered point cloud. S43. Calculate the principal axis direction of the secondary screening point cloud through principal component analysis, align the principal axis to the global vertical axis direction, and translate the midpoint of the ear axis to the origin of the coordinate system to obtain the attitude-normalized point cloud. S44. Perform density clustering on the attitude-normalized point cloud using a density clustering algorithm to obtain an initial nucleo-granular cluster set; S45. For excessively large clusters in the initial nucleogranule cluster set that are larger than a preset volume, extract high-brightness subsets and continue to optimize and subdivide them until all clusters meet the single nucleogranule volume constraint, and obtain the nucleogranule instance point cloud set.

6. The method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field according to claim 5, characterized in that, In step S45, the method for extracting the high-brightness subset further includes: S451. Select the points whose brightness values ​​within the current large cluster are located in the first preset percentile interval to form a high-brightness subset. The points in the high-brightness subset correspond to the nucleus-particle ridge line and the strongly irradiated surface region. S452. Using the high-brightness subset as input, perform secondary density clustering, map the obtained sub-cluster labels back to all points of the current oversized cluster, complete the current sub-subdivision, and obtain the high-brightness subset.

7. The method for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field according to claim 1, characterized in that, In step S5, the total number of nuclei is obtained by counting the total number of clusters in the nuclei instance point cloud set; In step S5, the method for calculating the nucleoparticle row count further includes: S51. Extract the middle segment point cloud from the nucleogranule instance point cloud set, where the vertical axis direction is located below the second preset percentile interval, as the middle segment analysis region of the spikelet; S52. Divide the mid-segment analysis region of the spike into multiple thin slices along the longitudinal axis. Construct a two-dimensional occupancy histogram for the cross-sectional coordinates of each point in each slice. After Gaussian smoothing and peak ratio threshold binarization, calculate the radial nucleogranule row number of the current thin slice by connecting component labeling. S53. Take the median of the radial nut row count of all thin-layer cuts as the nut row count of the ear.

8. A three-dimensional nucleogranular phenotypic analysis system for crop ears based on weakly supervised neural radiation fields, characterized in that, include: Annotation module: used to annotate the nucleoplast surface of multi-view image sequences of ear of fruit to obtain a sparse nucleoplast surface annotation dataset; Training module: used to input the sparse nucleogranular surface annotation dataset into the semantic segmentation model to train the semantic segmentation model and obtain the nucleogranular surface segmentation model; The inference module is configured as the nucleogranular surface segmentation model obtained by the training module, and is used to infer the multi-view image sequence to obtain a set of binary masks for the nucleogranular surface from multiple views. Fine-tuning module: used to perform segmentation and fine-tuning of the neural radiation field model using the set of binary masks on the nucleosome surface as a weak supervision signal, to obtain a three-dimensional volume representation of the nucleosome surface; Segmentation module: used to extract high-density three-dimensional point cloud from the three-dimensional volume representation of the nucleogranule surface, and to perform hierarchical clustering segmentation on the high-density three-dimensional point cloud to obtain a set of nucleogranule instance point clouds; Calculation module: used to calculate the total number of kernels and the number of kernel rows based on the kernel instance point cloud set.

9. A device for analyzing the three-dimensional nucleogranular phenotype of crop ears based on a weakly supervised neural radiation field, characterized in that, include: A memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause a crop ear three-dimensional kurtoid phenotype analysis device based on a weakly supervised neural radiation field to perform a crop ear three-dimensional kurtoid phenotype analysis method based on a weakly supervised neural radiation field as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement a method for analyzing the three-dimensional nucleosome phenotype of crop ears based on a weakly supervised neural radiation field, as described in any one of claims 1 to 7.

Images