Pumpkin seedling point cloud organ segmentation method based on DWSDENet

Abstract

The invention discloses a pumpkin seedling point cloud organ segmentation method based on DWSDENet, and the method comprises the steps: carrying out the wavelet transformation decoupling of different features of a pumpkin seedling point cloud, and carrying out the enhancement of the features of the point cloud through combining with a dynamic space regularization mechanism; the method comprises the steps of 1, point cloud feature channel coding; step 2, carrying out feature enhancement by using a dynamic wavelet-spatial dual-domain enhancer (DWSDE); step 3, feature decoding; step 4, performing organ segmentation based on the enhanced features; and the DWSDE dynamic wavelet-space double-domain intensifier module is embedded after down-sampling is carried out on the Adapt-Conv. And performing two steps of feature decoding and organ segmentation based on the enhanced features on the features enhanced by the dynamic wavelet-spatial double-domain enhancer to obtain a final segmentation head, thereby completing accurate segmentation of each organ in the whole pumpkin seedling point cloud. The method provided by the invention has the characteristic of high segmentation accuracy, and the organ segmentation intersection-union ratio reaches 90.45%.

Description

Technical Field

[0001] The present invention relates to a pumpkin seedling point cloud organ segmentation method based on DWSDENet. Background Art

[0002] In recent years, intelligent agricultural technology has developed rapidly. As one of the world's most important economic crops, pumpkin is widely used in food, feed, and deep processing. According to the "2023 Global Agricultural Cultivation Report," the area cultivated for pumpkin continues to grow, with global annual production exceeding 30 million tons. Precise management of pumpkin seedlings during the seedling stage is crucial to increasing yields. Efficient organ segmentation (differentiating between organs such as roots, stems, and leaves) can monitor the growth of pumpkin seedlings in real time, assist in disease and pest diagnosis, and optimize water and fertilizer management, thereby improving yield and resource utilization. However, for growers, traditional organ segmentation methods mainly rely on manual visual inspection or two-dimensional image analysis, which is time-consuming, labor-intensive, highly subjective, and has an error rate of over 30%. Especially in complex field environments, light fluctuations, occlusions, and other factors lead to a significant decrease in image recognition accuracy, making large-scale application difficult. On the other hand, the rise of point cloud technology (based on 3D laser scanning) provides a high-precision, non-destructive solution for organ segmentation. Point cloud data can capture the complete spatial structure of pumpkin seedlings, overcoming the limitations of two-dimensional images and enabling full-view, submillimeter-level detail extraction. However, point cloud segmentation faces challenges such as noise interference, uneven point density, and organ overlap. Existing algorithms, such as traditional PointNet-based models, suffer from low processing efficiency and poor generalization, with segmentation accuracy less than 50% in dense seedbeds. Furthermore, point cloud acquisition equipment (such as handheld laser scanners) is low-cost, easy to operate, and can be mounted on agricultural robots or drones, adapting to diverse lighting and terrain conditions.

[0003] In summary, it is necessary to design an efficient pumpkin seedling point cloud organ segmentation method. Summary of the Invention

[0004] There are two main technical problems to be solved by the present invention. The first is to extract feature information that can distinguish different categories from the pumpkin seedling point cloud, so as to accurately segment the organs of the pumpkin seedling point cloud. The second point is that the soil particles, weeds and noise interferences in the pumpkin seedling point cloud, as well as the significant morphological and structural differences between individual plants at different developmental stages, affect the ability of pumpkin seedling organ point cloud segmentation. In response to the first problem, the present invention proposes to use a multi-layer point cloud feature extraction module to extract hierarchical features. This method can extract effective information from the pumpkin seedling point cloud and obtain global and local information. In response to the second problem, the present invention proposes a feature fusion network (DWSDENet). First, wavelet transform decoupling is introduced to obtain the features of different components, and enhancement factors are applied to different components. Then, in order to enhance the learning ability of the morphological differences between individual plants, a dynamic weighting mechanism is combined to strengthen the expression of key features and effectively separate overlapping leaves and subtle stem features. A dynamic wavelet-spatial dual-domain enhancer (DWSDE) is introduced in the backbone network. This method has the characteristics of high segmentation accuracy.

[0005] The technical solutions of the invention are as follows:

[0006] 1. A pumpkin seedling point cloud organ segmentation method based on DWSDENet is characterized by decoupling morphological and detail features through wavelet transform, and combining dynamic spatial regularization mechanism to synergistically optimize frequency domain gradient and spatial statistical features, significantly improving the segmentation ability of high-frequency details such as stem vein texture and leaf edge serrations. The overall process includes the following steps:

[0007] Step 1: Point cloud feature channel encoding;

[0008] The pumpkin seedling point cloud in ShapeNetPart format (format: N×[x, y, z, Nx, Ny, Nz, Tag], where N is the number of points) is input. First, a multi-layer point cloud feature extraction module is used to learn and aggregate local neighborhood information layer by layer to generate point cloud feature channels with different scales and levels of abstraction (64-dimensional, 128-dimensional, 256-dimensional, and 512-dimensional feature vectors).

[0009] Step 2: Apply dynamic wavelet-spatial dual-domain enhancer (DWSDE) for feature enhancement;

[0010] The preprocessed point cloud is input into the Dynamic Wavelet-Spatial Dual-Domain Enhancer (DWSDE). This module first performs a wavelet transform on the point cloud to decouple the low-frequency overall morphology component from the high-frequency detail component. Secondly, a dynamic spatial regularization mechanism is applied to the high-frequency detail component, adaptively adjusting the feature weights based on the geometric statistics of the point's local neighborhood. Finally, a bilinear interaction layer is used to cross-domain fuse the enhanced high-frequency detail features with the low-frequency morphological features, achieving a coordinated optimization of frequency-domain gradient information and spatial statistical features, and outputting the enhanced point cloud feature representation.

[0011] Step 3: Feature decoding

[0012] The multi-scale feature channels enhanced by the Dynamic Wavelet-Spatial Dual Domain Enhancer (DWSDE) are fed into the feature decoding path. The deep, high-semantic information is gradually fused with the shallow, high-resolution position information to restore the spatial details of the point cloud and generate point-by-point features with rich semantics and precise position information.

[0013] Step 4: Organ segmentation based on enhanced features;

[0014] The decoded point-by-point features are passed through the final segmentation head (a small MLP consisting of three lxl convolutional layers (point-by-point convolution)) to predict the organ category (stem, leaf, irrelevant items) to which each point cloud belongs, thereby completing the accurate segmentation of each organ in the entire pumpkin seedling point cloud.

[0015] The specific steps of step one are:

[0016] Input the pumpkin seedling point cloud in ShapeNetPart format into the point cloud feature extraction module;

[0017] The module consists of three processes:

[0018] Farthest Point Sampling (FPS): Iteratively select M key points from the input point cloud (N = 2048 → M1 = 512 → M2 = 128 → M3 = 32);

[0019] Ball Query: With each key point as the center, search for neighboring points within a radius of r;

[0020] Local feature extraction: for each neighborhood point Apply a shared MLP layer (of dimension [64, 128, 256, 512]) and aggregate local features via Max Pooling:

[0021]

[0022] where p j -p i is the relative coordinate, p jis the absolute coordinate;

[0023] Repeat the above process three times to output the multi-scale point cloud feature channel:

[0024] Scale 1:

[0025] Scale 2:

[0026] Scale 3:

[0027] Scale 3:

[0028] In step 2, the multi-scale features F1, F2, F3, and F4 output by the encoder are respectively enhanced by applying the dynamic wavelet-spatial dual-domain enhancer (DWSDE);

[0029] The Dynamic Wavelet-Spatial Dual-Domain Enhancer (DWSDE) module consists of two processes:

[0030] Wavelet transform enhancement: First, the input local point cloud is subjected to wavelet transform. Two-dimensional wavelet decomposition is performed independently for each feature channel, and the original features are decoupled into a low-frequency component cA and a set of high-frequency components (cH, cV, cD). The low-frequency component cA encodes the overall morphological information of the organ (the direction of the stem trunk), and the high-frequency components cH, cV, and cD represent the local detail features in the horizontal, vertical, and diagonal directions (leaf edge curvature, stem-leaf junction mutation). Secondly, an enhancement factor is applied to the high-frequency component, and its contribution intensity is adjusted by parameters α, β, and Y. Finally, the enhanced frequency domain features are reconstructed into spatial domain features F through inverse wavelet transform. enhanced , and perform residual connection with the original features.

[0031] Dynamic multi-scale fusion: First, the input feature F is reduced in dimension and activated by two parallel 1×1 convolutions to obtain the reduced-dimensional feature F (1) and F (2) Then, from F (1) Perform global average pooling, global maximum pooling, and local average pooling (window size w, step size s) to obtain three sets of multi-scale channel features, and map the local pooling results to the same size through interpolation. Then, these three sets of channel features are sent to lightweight MLP+Softmax to generate dynamic weights The three sets of features are weighted and fused to obtain the fused channel description f c At the same time, the F (2) Perform channel average pooling to obtain spatial features f space Then, using the channel feature f c With the spatial feature f spaceDo bilinear interaction, add the original dimensionality reduction features, and get F through square root nonlinear transformation fused Finally, the number of channels is restored through 1×1 upsampling convolution + BN + ReLU, and the input feature F is subtracted from F fused Output in the form of residual.

[0032] In step 3, the enhanced multi-scale point cloud features are put into the feature decoding module;

[0033] The module consists of three processes:

[0034] Upsampling path: Use the Feature Propagation (FP) layer to gradually restore the spatial resolution (512→1024); each step propagates deep features to shallow points through inverse distance weighted interpolation;

[0035] Jump connection: concatenate the encoder's same-scale features with the upsampled features by channel (torch.cat([F enc , F up ], dim = -1);

[0036] Feature refinement: Apply MLP layer to the concatenated features to generate decoding features

[0037] In step 4, the point-by-point features output by the decoder are input into the segmentation head. The segmentation head structure consists of two layers of 1×1 convolutional layers:

[0038] First layer: Convld(256,128)+BatchNorm+ReLU

[0039] Second layer: Convld(128,C cls ), where C cls =3 (corresponding to the three categories of stem, leaf, and irrelevant items);

[0040] Finally, the output vector for each point i Apply the Softmax function:

[0041]

[0042] Pick as the predicted organ category label of point i.

[0043] Beneficial effects:

[0044] The pumpkin seedling point cloud organ segmentation method based on DWSDENet in this invention is a feature fusion method that aims to capture local detail features by using wavelet transform, and at the same time combine with dynamic weighting mechanism to strengthen the expression of key features and reduce the influence of pumpkin seedling point cloud noise information on segmentation.

[0045] First, to make the data more realistic and the model designed in this invention more adaptable, we cultivated pumpkin seedlings in an artificial climate chamber at the Agricultural Research Institute and used a 3D laser scanner to obtain point cloud data of pumpkin seedlings in various shapes. This data included interfering objects such as soil particles and weeds. Five agricultural experts then annotated the point clouds organ by organ. Ultimately, we created a dataset in the ShapeNetPart format containing three categories: roots, leaves, and irrelevant items.

[0046] Secondly, to reduce the loss of fine-structure features extracted by traditional point cloud convolution, the present invention proposes a wavelet transform module to perform a wavelet transform on the input local point cloud. Two-dimensional wavelet decomposition is performed independently on each feature channel, decoupling the original features into a set of low-frequency and high-frequency components. Enhancement factors are applied to each component, and their contribution strengths are adjusted through parameters. This proposed feature fusion can improve the accuracy of pumpkin seedling point cloud organ segmentation.

[0047] Thirdly, the present invention adaptively processes organ scale differences and improves the ability to distinguish overlapping leaves and subtle stem features. The present invention embeds a dynamic weighting module after wavelet transform, optimizes the ability to distinguish complex morphological details of pumpkin seedlings, and improves the accuracy of pumpkin seedling point cloud organ segmentation.

[0048] To better extract features from pumpkin seedling point clouds and fully absorb the timbre and pitch information of audio signals, this paper proposes a feature fusion method (DWSDE). It combines the ability of wavelet transform to capture local high-frequency details with the advantages of dynamic weighting mechanism in enhancing key features. By integrating wavelet transform and dynamic weighting operation, the DWSDE module not only optimizes the fusion of local details and global context information, but also significantly improves the model's ability to distinguish complex morphological details in pumpkin seedling point cloud segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the network framework diagram of pumpkin seedling point cloud organ segmentation based on DWSDENet in the present invention;

[0050] Figure 2 This is the Adapt-Conv network framework diagram;

[0051] Figure 3 Pumpkin seedling point cloud dataset DETAILED DESCRIPTION

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0053] Figure 1 This is the network framework diagram of the present invention, which includes two main modules:

[0054] 1. Wavelet Feature Enhancer: First, the input local point cloud is subjected to wavelet transform. Two-dimensional wavelet decomposition is performed independently for each feature channel, and the original features are decoupled into a low-frequency component cA and a set of high-frequency components (cH, cV, cD). The low-frequency component cA encodes the overall morphological information of the organ (such as the direction of the stem trunk), and the high-frequency components cH, cV, and cD represent the local detail features in the horizontal, vertical, and diagonal directions (such as the curvature of the leaf edge and the mutation of the stem-leaf junction). Secondly, an enhancement factor is applied to the high-frequency component, and its contribution intensity is adjusted by parameters α, β, and γ. Finally, the enhanced frequency domain features are reconstructed into spatial domain features F_enhanced through inverse wavelet transform, and residual connection is performed with the original features.

[0055] 2. Dynamic Multi Scale Fusion (DMSF): First, the input feature F is reduced in dimension and activated by two parallel 1×1 convolutions to obtain the reduced dimension feature F ( 1 ) and F( 2 Then, from F( 1 ) are respectively subjected to global average pooling, global maximum pooling and local average pooling (window size w, step size s) to obtain three sets of multi-scale channel features, and the local pooling results are mapped to the same size by interpolation. Then, these three sets of channel features are sent to the lightweight MLP+Softmax to generate dynamic weights The three sets of features are weighted and fused to obtain the fused channel description f c At the same time, the F (2) Perform channel average pooling to obtain spatial features f space Then, using the channel feature f c With the spatial feature f space Do bilinear interaction, add the original dimensionality reduction features, and get F through square root nonlinear transformation fused Finally, the number of channels is restored through 1×1 upsampling convolution + BN + ReLU, and the input feature F is subtracted from F fused Output in the form of residual.

[0056] Figure 2 Adapt-Conv network was first proposed by six Chinese researchers including Haoran Zhou from Nanjing University in 2021. Its core innovation lies in addressing the geometric adaptability flaw of traditional convolution. By dynamically generating input-dependent convolution kernels, it breaks through the representation bottleneck of fixed-weight kernels in irregular point cloud structures. The generation method of convolution kernels can be expressed as:

[0057]

[0058] in is the kernel generating function, P geo are local neighborhood coordinates, is the residual connection.

[0059] This convolution generates an adaptive kernel based on its dynamically learned features. Compared to using fixed / isotropic kernels, AdaptConv improves the flexibility of point cloud convolution, effectively and accurately capturing the different relationships between points in different semantic parts. Unlike the popular attention weight scheme, the proposed AdaptConv implements adaptability in the convolution operation instead of simply assigning different weights to adjacent points.

[0060] Figure 3 This is a point cloud dataset of pumpkin seedlings. The "Zhuangshi" pumpkin (Cucurbita moschata D.) was used as the experimental material, which was provided by Fujian Nongyou Seedling Co., Ltd. (Fuzhou, China). The experimental materials were cultured in an artificial climate chamber at the Institute of Facility Agriculture, Guangdong Academy of Agricultural Sciences from July 2023 to August 2023. The pumpkin seeds were first soaked in 55°C hot water, and then soaked for 4-6 hours after natural cooling. After soaking, the seeds were repeatedly washed and finally wrapped with sterilized wet gauze to keep them moist. The seeds were placed in a constant temperature incubator at 28°C for germination. When the seeds grew a 0.5 cm white radicle, they were neatly sown in a nutrient pot filled with a substrate. The pumpkin material was cultured in an artificial climate chamber (temperature 25°C, 14 hours of daylight / 10 hours of nightlight, average relative humidity 60-85%, and photosynthetically active radiation of 160 μmol m-2s-1). After seed germination, when the pumpkin seedlings had grown steadily to a plant height of approximately 13 cm to 18 cm, 450 plants with good growth were selected for 3D data acquisition.

[0061] In terms of data acquisition, we use RVC-X-mini structured light camera to collect point cloud data. The specific parameters of the camera are shown in Table 1. Figure 1 As shown in the figure, the camera is firmly mounted on the end of the ELITE ROBOT six-axis robotic arm. The base of the robotic arm and the placement of the pumpkin seedlings are fixed to ensure stability during imaging. By precisely controlling the movement of the robotic arm, the camera can capture images of the pumpkin seedlings from multiple angles, thereby realizing multi-viewpoint point cloud data acquisition. These point cloud data are then processed through manual coarse alignment and spatial coordinate matrix transformation to unify the data from different perspectives into a single initial viewpoint. This step is crucial for subsequent 3D reconstruction. Representative point clouds of pumpkin seedlings of different morphologies were selected, and it can be seen that different samples have obvious differences in organ morphology and local point cloud structure. Some samples have clear and evenly distributed leaf edges, while others have more complex geometric details due to local noise or uneven sampling.

[0062] Detailed description of steps

[0063] like Figure 1-3 The pumpkin seedling point cloud organ segmentation method based on DWSDENet of the present invention includes the following steps:

[0064] Step 1: Point cloud feature channel encoding: Input the pumpkin seedling point cloud in ShapeNetPart format into the point cloud feature extraction module;

[0065] The module consists of three processes:

[0066] Farthest Point Sampling (FPS): Iteratively select M key points from the input point cloud (N = 2048 → M1 = 512 → M2 = 128 → M3 = 32);

[0067] Ball Query: With each key point as the center, search for neighboring points within a radius of r;

[0068] Local feature extraction: for each neighborhood point Apply a shared MLP layer (of dimension [64, 128, 256, 512]) and aggregate local features via Max Pooling:

[0069]

[0070] where p j -p i is the relative coordinate, p j is the absolute coordinate;

[0071] Repeat the above process three times to output the multi-scale point cloud feature channel:

[0072] Scale 1:

[0073] Scale 2:

[0074] Scale 3:

[0075] Scale 3:

[0076] Step 2: Apply the dynamic wavelet-spatial dual-domain enhancer (DWSDE) for feature enhancement: The multi-scale features F1, F2, F3, and F4 output by the encoder are respectively enhanced by the dynamic wavelet-spatial dual-domain enhancer (DWSDE). The dynamic wavelet-spatial dual-domain enhancer (DWSDE) module consists of two processes:

[0077] Wavelet transform enhancement: First, the input local point cloud is subjected to wavelet transform. Two-dimensional wavelet decomposition is performed independently for each feature channel, and the original features are decoupled into a low-frequency component cA and a set of high-frequency components (cH, cV, cD). The low-frequency component cA encodes the overall morphological information of the organ (the direction of the stem trunk), and the high-frequency components cH, cV, and cD represent the local detail features in the horizontal, vertical, and diagonal directions (leaf edge curvature, stem-leaf junction mutation). Secondly, an enhancement factor is applied to the high-frequency component, and its contribution intensity is adjusted by parameters α, β, and Y. Finally, the enhanced frequency domain features are reconstructed into spatial domain features F through inverse wavelet transform. enhanced , and perform residual connection with the original features. The following equation can express the above process:

[0078]

[0079] Here, cA encodes the main organ shape, while cH, cV, and cD correspond to detailed features in the horizontal, vertical, and diagonal directions (such as changes in leaf edge curvature), respectively. α, β, and Y∈[0,1] are learnable parameters optimized via backpropagation. The stem-leaf junction is sensitive to the vertical high-frequency component (β), while the overlapping leaf region is more dependent on the horizontal component (α). A truncation operation [1:N] is used to ensure feature dimension alignment and avoid geometric distortion caused by edge effects.

[0080] Dynamic multi-scale fusion: First, the input feature F is reduced in dimension and activated by two parallel 1×1 convolutions to obtain the reduced-dimensional feature F( 1 ) and F( 2 Then, from F( 1 ) are respectively subjected to global average pooling, global maximum pooling and local average pooling (window size w, step size s) to obtain three sets of multi-scale channel features, and the local pooling results are mapped to the same size by interpolation. Then, these three sets of channel features are sent to the lightweight MLP+Softmax to generate dynamic weights The three sets of features are weighted and fused to obtain the fused channel description f c At the same time, the F (2) Perform channel average pooling to obtain spatial features f space Then, using the channel feature f c With the spatial feature f space Do bilinear interaction, add the original dimensionality reduction features, and get F through square root nonlinear transformation fused Finally, the number of channels is restored through 1×1 upsampling convolution + BN + ReLU, and the input feature F is subtracted from F fused Output in the form of residual.

[0081] The following equation can express the above process:

[0082] F(1) =ReLU(ConvDown(F))

[0083] F (2) =ReLU(ConvDown(F))

[0084] in, is the original input, ConvDown is a 1×1 convolution, and ReLU is used for activation. Through two parallel 1×1 convolutions, we reduce the number of channels from C to C / r and get F (1) and F (2) , respectively used for subsequent channel and spatial feature extraction.

[0085] After completing the dimensionality reduction, the next step is to (1) Multi-scale channel features are extracted to capture global and local information.

[0086]

[0087] Among them, AvgPool1D(F (1) ; w, s) represents one-dimensional average pooling with window w and step size s on length N, and the output length is approximately ″Interpolate″(·) means interpolating / scaling the pooled segmentation result to a final single value (or fixed size) to align with other feature branches. avg Capture global average information; f max Capture the most significant global activation; f local It reflects the local context changes; the three together realize the multi-scale feature extraction of the channel dimension.

[0088] After obtaining the three sets of channel features, they need to be adaptively assigned weights and fused to enter the dynamic weight calculation stage.

[0089]

[0090] w=Softmax(MLP(v))

[0091]

[0092] in, is the statistical vector after averaging the three sets of features across channels; MLP (Shared Multilayer Perceptron) consists of a series of 1×1 convolution, BN, and ReLU; Softmax is normalized in the last dimension to obtain w = (w1, w2, w3); The network automatically adjusts the importance of the three scales based on the input, avoiding the fixed weight limitations of simple concatenation or averaging.

[0093] After obtaining the adaptive channel description, another parallel branch will be used to 2 ) to extract spatial features.

[0094]

[0095] After dimensionality reduction, F( 2 ) performs average pooling in the channel dimension to obtain a one-dimensional spatial context feature, which reflects the overall response of each position in the entire channel.

[0096] At this point, we have the channel description f c and the spatial description f space , bilinear interaction is possible.

[0097] F bilinear =f c ×f space

[0098]

[0099] Among them, matrix multiplication “×” realizes the global interaction between channels and spaces; adding ∈ ensures The value of is stable;

[0100] After completing the feature fusion, the last step is to restore the number of channels to the original C and output it in the form of residual.

[0101]

[0102] Among them, ConvUp is a 1×1 upsampling convolution, which restores the number of channels from C / r to C; BN+ReLU enhances nonlinearity and training stability; F enhanced is the feature enhanced by the wavelet transform module.

[0103] In step 3, feature decoding: the enhanced multi-scale point cloud features are put into the feature decoding module; the module consists of three processes:

[0104] Upsampling path: Use Feature Propagation (FP) layer to gradually restore the spatial resolution (512→1024): each step propagates the deep features to the shallow points through inverse distance weighted interpolation; jump connection: concatenate the encoder's same-scale features with the upsampled features by channel (torch.cat([F enc , F up ], dim=-1)):

[0105] Feature refinement: Apply MLP layer to the concatenated features to generate decoding features

[0106] Step 4: Organ segmentation based on enhanced features: Input the point-by-point features output by the decoder into the segmentation head. The segmentation head structure consists of two layers of 1×1 convolutional layers:

[0107] First layer: Convld(256,128)+BatchNorm+ReLU

[0108] Second layer: Convld(128,C cls ), where C cls =3 (corresponding to the three categories of stem, leaf, and irrelevant items);

[0109] Finally, the output vector for each point i Apply the Softmax function:

[0110]

[0111] Pick as the predicted organ category label of point i.

[0112] In order to prove the validity of the patent of this invention, the following experiments were conducted.

[0113] In this experiment, the "Zhuangshi" pumpkin (Cucurbita moschata D.) was used as the experimental material, which was provided by Fujian Nongyou Seedling Co., Ltd. (Fuzhou, China). The experimental materials were cultured in an artificial climate chamber at the Institute of Facility Agriculture, Guangdong Academy of Agricultural Sciences from July 2023 to August 2023. The pumpkin seeds were first soaked in 55°C hot water, and then soaked for 4-6 hours after natural cooling. After soaking, the seeds were repeatedly washed and finally wrapped with sterilized wet gauze to keep them moist. The seeds were placed in a constant temperature incubator at 28°C for germination. When the seeds grew a 0.5 cm white radicle, they were neatly sown in a nutrient pot filled with a substrate. The pumpkin material was cultured in an artificial climate chamber (temperature 25°C, 14 hours of daylight / 10 hours of nightlight, average relative humidity 60-85%, and photosynthetically active radiation of 160 μmol m-2 s-1). After seed germination, when the pumpkin seedlings had grown steadily to a plant height of approximately 13 cm to 18 cm, 450 plants with good growth were selected for 3D data acquisition. Figure 2 This image shows representative point clouds of pumpkin seedlings of varying shapes. Clear differences in organ morphology and local point cloud structure are evident across samples. Some samples have clear, evenly distributed leaf edges, while others exhibit more complex geometric details due to local noise or uneven sampling.

[0114] The experiments were conducted using the following configuration: Intel(R) Xeon(R) Platinum 8352V CPU, 48GB memory, and NVIDIA GeForce RTX 2080Ti x2 (22GB) graphics card. The deep learning framework PyTorch installed in Ubuntu was used to build the 3D segmentation network. Although the versions of Python and CUDA do not affect the experimental results, they must be compatible with the specific software and hardware used. PRNet was implemented using PyTorch 1.11.0 and Python 3.8. During the training phase, the batch size was set to 16 to balance computational efficiency and model convergence. The initial learning rate was set to 0.1 to promote stable gradient descent and learning through practice. The training process was carried out for 200 epochs, allowing the model to have enough iterations to learn from the data. The maximum and minimum learning rates were 1×10^(-1) and 1×10^(-3), respectively. This patent proposes an efficient feature fusion network DWSDENet to improve the ability to accurately segment pumpkin seedling organ point clouds. This experiment uses Adapt-Conv as the backbone network to verify the effectiveness of feature fusion.

[0115] The experimental results are shown in Table 1. Using the Adapt-Conv backbone network, we performed organ segmentation on a pumpkin seedling point cloud. The original network achieved an Intersection-over-Union (IoU) of 90.19% for organ segmentation. By adding the Dynamic Wavelet-Spatial Dual-Domain Enhancer (DWSDE) for feature enhancement while maintaining the backbone network, we achieved an IoU of 90.45%, a 6.20% improvement over the baseline.

[0116] Table 1

[0117] Network Model Organ segmentation intersection-union ratio Adapt-Conv 84.65％ DWSDENet 90.45％

[0118] The network's recognition rate for a dataset is related to the data used to divide the training and test sets. Therefore, the self-built dataset is divided into three parts: Split 1, Split 2, and Split 3. The experimental results are shown in Table 2:

[0119] Table 2

[0120] Number of divisions Adapt-Conv DWSDENet Split1 84.01％ 90.19％ Split2 84.19％ 90.28％ Split3 84.33％ 90.45％

[0121] In the three-partition dataset, embedding the proposed model into Adapt-Conv achieved the highest classification accuracy improvement of 6.18%, followed by 6.12%, and the lowest improvement of 6.09% over the backbone network. Overall, it steadily improved the performance of the backbone network. The experiments in Table 1 were all conducted under the Split2 partition.

[0122] This paper discloses a DWSDENet-based method for organ segmentation in pumpkin seedling point clouds. This method first extracts multi-scale features from point clouds using a hierarchical encoder and then performs a frequency-spatial joint optimization of these features using a dynamic wavelet-spatial dual-domain enhancer (DWSDE). Wavelet transforms are used in the frequency domain to decouple organ geometry and texture details, while dynamic multi-scale fusion is used in the spatial domain to enhance multi-scale context perception. A feature decoder is then used to restore organ boundary details, and a lightweight segmentation head is used to achieve accurate organ classification. This method achieves a 90.45% intersection-over-union (mIoU) on a greenhouse pumpkin seedling point cloud dataset.

Claims

1. A pumpkin seedling point cloud organ segmentation method based on DWSDENet, characterized in that: By decoupling morphological and detail features through wavelet transform and combining it with a dynamic spatial regularization mechanism to synergistically optimize frequency domain gradients and spatial statistical features, the segmentation capability of high-frequency details such as stem vein texture and leaf edge serrations is significantly improved. The overall process includes the following steps: Step 1: Point cloud feature channel encoding; The pumpkin seedling point cloud in ShapeNetPart format (format: N×[x, y, z, Nx, Ny, Nz, Tag], where N is the number of points) is input. First, a multi-layer point cloud feature extraction module is used to learn and aggregate local neighborhood information layer by layer to generate point cloud feature channels with different scales and levels of abstraction (64-dimensional, 128-dimensional, 256-dimensional, and 512-dimensional feature vectors). Step 2: Apply dynamic wavelet-spatial dual-domain enhancer (DWSDE) for feature enhancement; The preprocessed point cloud is input into the Dynamic Wavelet-Spatial Dual-Domain Enhancer (DWSDE). This module first performs a wavelet transform on the point cloud to decouple the low-frequency overall morphology component from the high-frequency detail component. Secondly, a dynamic spatial regularization mechanism is applied to the high-frequency detail component, adaptively adjusting the feature weights based on the geometric statistics of the point's local neighborhood. Finally, a bilinear interaction layer is used to cross-domain fuse the enhanced high-frequency detail features with the low-frequency morphological features, achieving a coordinated optimization of frequency-domain gradient information and spatial statistical features, and outputting the enhanced point cloud feature representation. Step 3: Feature decoding The multi-scale feature channels enhanced by the Dynamic Wavelet-Spatial Dual Domain Enhancer (DWSDE) are fed into the feature decoding path. The deep, high-semantic information is gradually fused with the shallow, high-resolution position information to restore the spatial details of the point cloud and generate point-by-point features with rich semantics and precise position information. Step 4: Organ segmentation based on enhanced features; The decoded point-by-point features are passed through the final segmentation head (a small MLP consisting of three 1x1 convolutional layers (point-by-point convolution)) to predict the organ category (stem, leaf, irrelevant items) to which each point cloud belongs, thereby completing the accurate segmentation of each organ in the entire pumpkin seedling point cloud.

2. The pumpkin seedling point cloud organ segmentation method based on DWSDENet according to claim 1 is characterized in that: The specific steps of step one are: Input the pumpkin seedling point cloud in ShapeNetPart format into the point cloud feature extraction module; The module consists of three processes: Farthest Point Sampling (FPS): Iteratively select M key points from the input point cloud (N = 2048 → M1 = 512 → M2 = 128 → M3 = 32); Ball Query: With each key point as the center, search for neighboring points within a radius of r; Local feature extraction: for each neighborhood point Apply a shared MLP layer (of dimension [64, 128, 256, 512]) and aggregate local features via MaxPooling: where p j -p i is the relative coordinate, p j is the absolute coordinate; Repeat the above process three times to output the multi-scale point cloud feature channel: Scale 1: Scale 2: Scale 3: Scale 3:

3. The pumpkin seedling point cloud organ segmentation method based on DWSDENet according to claim 2 is characterized in that: In step 2, the multi-scale features F1, F2, F3, and F4 output by the encoder are respectively enhanced by applying the dynamic wavelet-spatial dual-domain enhancer (DWSDE); The Dynamic Wavelet-Spatial Dual-Domain Enhancer (DWSDE) module consists of two processes: Wavelet transform enhancement: First, the input local point cloud is subjected to wavelet transform. Two-dimensional wavelet decomposition is performed independently for each feature channel, and the original features are decoupled into a low-frequency component cA and a set of high-frequency components (cH, cV, cD). The low-frequency component cA encodes the overall morphological information of the organ (the direction of the stem trunk), and the high-frequency components cH, cV, and cD represent the local detail features in the horizontal, vertical, and diagonal directions (leaf edge curvature, stem-leaf junction mutation). Secondly, an enhancement factor is applied to the high-frequency component, and its contribution intensity is adjusted by parameters α, β, and Y. Finally, the enhanced frequency domain features are reconstructed into spatial domain features F through inverse wavelet transform. enhanced , and perform residual connection with the original features. Dynamic multi-scale fusion: First, the input feature F is reduced in dimension and activated by two parallel 1×1 convolutions to obtain the reduced-dimensional feature F (1) and F (2) Then, from F (1) Perform global average pooling, global maximum pooling, and local average pooling (window size w, step size s) to obtain three sets of multi-scale channel features, and map the local pooling results to the same size through interpolation. Then, these three sets of channel features are sent to lightweight MLP+Softmax to generate dynamic weights The three sets of features are weighted and fused to obtain the fused channel description f c At the same time, the F (2) Perform channel average pooling to obtain spatial features f space Then, using the channel feature f c With the spatial feature f space Do bilinear interaction, add the original dimensionality reduction features, and get F through square root nonlinear transformation fused Finally, the number of channels is restored through 1×1 upsampling convolution + BN + ReLU, and the input feature F is subtracted from F fused Output in the form of residual.

4. The pumpkin seedling point cloud organ segmentation method based on DWSDENet according to claim 3 is characterized in that: In step 3, the enhanced multi-scale point cloud features are put into the feature decoding module; The module consists of three processes: Upsampling path: Use the Feature Propagation (FP) layer to gradually restore the spatial resolution (512→1024); each step propagates deep features to shallow points through inverse distance weighted interpolation; Jump connection: concatenate the encoder's same-scale features with the upsampled features by channel (torch.cat([F enc , F up ], dim = -1); Feature refinement: Apply MLP layer to the concatenated features to generate decoding features 5. The pumpkin seedling organ segmentation method according to claim 4, characterized in that: In step 4, the point-by-point features output by the decoder are input into the segmentation head. The segmentation head structure consists of two layers of 1×1 convolutional layers: First layer: Convld(256,128)+BatchNorm+ReLU Second layer: Convld(128,C cls ), where C cls =3 (corresponding to the three categories of stem, leaf, and irrelevant items); Finally, the output vector for each point i Apply the Softmax function: Pick as the predicted organ category label of point i.

Images