A method for generating rice CT data based on a GAN network

Abstract

The application belongs to the technical field of agricultural data processing, and particularly relates to a rice CT data generation method based on a GAN network, which comprises the following steps: S1, pretreating rice seedlings; S2, collecting two CT images of rice seedling stems; S3, performing standardization processing on the two collected CT images; S4, constructing a rice CT data generation model in combination with a DCLB module and a 2DTo3D module; S5, training the generation model; and S6, inputting the two CT images processed by standardization into the trained generation model, performing feature extraction and spatial information fusion on the input CT images, and generating rice CT image data of an intermediate section. The method overcomes the shortcomings of the prior art, effectively fills the blank between rice slice images, and provides reliable data support for subsequent rice growth analysis and research.

Description

Technical Field

[0001] This invention belongs to the field of agricultural data processing technology, specifically relating to a method for generating rice CT data based on GAN networks. Background Technology

[0002] Rice is one of the most important crops in my country and the world. Especially in my country, improving its yield and quality is crucial for stable economic development. Currently, the main method of rice cultivation in my country is machine transplanting. This method is beneficial for increasing yield and stabilizing output, while also saving labor costs and alleviating labor shortages. However, machine transplanting can easily damage seedlings. For example, damage to the base of the stem can lead to slow seedling recovery and establishment, prolonged growth period, and even seedling death, all of which seriously affect the yield and quality of rice.

[0003] To address these issues, research on the biomechanical properties of rice seedlings is crucial. Generally, this is achieved by establishing a biomechanical model that integrates multi-scale characteristic parameters of large rice seedling stems. This model can accurately simulate the mechanical behavior of rice stems, obtain a quantitative relationship between the degree of stem damage and stress and strain, overcome the limitations of conventional experimental methods, and contribute to the accurate quantitative assessment of mechanical damage to rice seedling stems.

[0004] Obtaining crop microstructure models is fundamental to analyzing mechanical properties. Currently, researchers are employing methods such as homogenization, representative volume element methods, and multi-scale modeling based on X-ray computed tomography (CT) to study the mechanical properties of biomaterials. Current research focuses on using modern technologies such as Micro-CT to obtain three-dimensional microphysical models of rice seedling stems and reconstructing them using high-precision image segmentation techniques.

[0005] This method not only allows for the acquisition of the true geometric features of materials without damaging the sample, but also helps to accurately construct key technologies for growth, damage detection, and variety selection, providing a scientific basis for rice cultivation. However, my country has a large number of rice varieties, and with the development of hybrid rice technology, the number of rice varieties is increasing year by year. Each variety also has multiple morphologies. If each morphology is scanned using Micro-CT technology to construct a three-dimensional microscopic physical model, the cost would be extremely high. Therefore, it is necessary to develop a method that can generate a three-dimensional microscopic physical model of a small segment of rice by analyzing cross-sectional images of both ends. With the continuous development of agricultural automation and intelligent technologies, deep learning technology is gradually being applied to crop reconstruction, but most of these methods reconstruct the overall morphology of crops, failing to accurately generate the microscopic structure.

[0006] The X2CT-GAN network framework based on deep learning generative adversarial networks (GANs) has been used to reconstruct CT scans from lung X-ray images. However, this method is applied in the medical field, and the conversion from two-dimensional features to three-dimensional features is not stable, resulting in the reconstructed CT scans lacking clear boundaries and textures in areas with fine structures. Therefore, this invention proposes a method for reconstructing CT scans using biplane transverse slice images to solve the problems existing in the prior art. Summary of the Invention

[0007] The purpose of this invention is to provide a method for generating rice CT data based on GAN networks, which overcomes the shortcomings of existing technologies, effectively fills the gaps between rice slice images, and provides reliable data support for subsequent rice growth analysis and research.

[0008] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0009] A method for generating CT data of rice based on GAN networks includes the following steps:

[0010] S1. Pre-treat rice seedlings;

[0011] S2. Collect CT images of the cross sections at both ends of the rice seedling stems;

[0012] S3. Standardize the two acquired CT images;

[0013] S4. Construct a rice CT data generation model with x2ct-gan as the backbone network and DCLB (Dilation Convolutional Layer Block) and 2DTo3D modules.

[0014] S5. Train the generative model;

[0015] S6. Input the two standardized CT images into the trained generative model, extract features and fuse spatial information from the input CT images to generate the CT image data of the middle section of rice.

[0016] Furthermore, the rice seedlings mentioned in S1 are 40-day high-yield rice varieties, and the pretreatment involves soaking the rice seedlings in a 15% cesium iodide solution.

[0017] Furthermore, the normalization process described in S3 includes adjusting the image resolution and grayscale range.

[0018] Furthermore, the generative model includes a 3D generator and a 3D discriminator, and the encoder of the 3D generator uses a DCLB module for feature extraction; a 2DTo3D module is used to convert the two-dimensional features into three-dimensional features.

[0019] Furthermore, the DCLB module utilizes four dilated convolutional kernels with different dilation rates (r=1, r=2, r=3, r=4) to process input features in parallel, capturing spatial features through multiple receptive fields at various scales, thereby enhancing the modeling capability of fine structures in rice CT data. This module operates through the following steps:

[0020] (1) The input features are first reduced in size by a downsampling module, which halves the width and height of the original features, thereby achieving feature compression;

[0021] (2) Then, a convolutional layer is used for dimensionality reduction to reduce the number of channels to 1 / 4 of the original number of channels, so as to reduce computational complexity;

[0022] (3) The dimensionality-reduced features are convolved in parallel through four dilated convolutional layers, using different dilation rates r=1, r=2, r=3 and r=4 respectively, to capture spatial features at different scales;

[0023] (4) The outputs of the four dilated convolutional layers are added element by element to fuse multi-scale information;

[0024] (5) Then, the fused features are restored to the original number of channels through a full-channel convolutional layer, and the output feature map is added to the features before dimensionality reduction in step (2).

[0025] (6) Repeat steps (2) to (5) N times to output the feature map.

[0026] (7) Finally, the feature map is concatenated with the downsampled features from step (1) along the channel number dimension to generate the final feature map.

[0027] Furthermore, the 2DTo3D module is used to extend two-dimensional features to three-dimensional features, and its specific steps are as follows:

[0028] (1) Two-dimensional feature broadcasting expansion: The input two-dimensional features are expanded along the third dimension (depth direction) through a broadcasting mechanism to generate a three-dimensional feature map with four layers of depth.

[0029] (2) Stitching operation: The expanded two-dimensional features are stitched together with the downsampled three-dimensional features so that the spatial information of the two-dimensional features and the three-dimensional features remains consistent during the information fusion process.

[0030] (3) Three-dimensional convolution processing: The spliced ​​features are processed through multi-layer three-dimensional convolution operations to further extract structural information in three-dimensional space and improve the model's ability to model three-dimensional data.

[0031] Furthermore, in S5, the generative model is trained, specifically through the following methods:

[0032] S51. Acquire pre-processed CT images of rice seedling stems;

[0033] S52. Create a CT image dataset from the collected CT images of rice seedling stems;

[0034] S53. Divide the CT image dataset into a training set and a test set. Use the training set to input the generative model for data modeling, and use the test set to test the generative model.

[0035] Furthermore, the steps described in S52 for creating a CT image dataset from the collected rice seedling stem CT images are as follows:

[0036] (1) The collected CT images of rice seedling stems were cropped at different sizes, and the images cropped at the same size were grouped into the same batch;

[0037] (2) Scale the cropped image to 256*256;

[0038] (3) Take 256 images from the same batch and stack them in order to form a three-dimensional CT image data. After rotating the images, a CT image dataset consisting of 1180 images is obtained.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] The present invention discloses a method for generating rice CT data based on GAN network, which constructs a rice CT data generation model with x2ct-gan as the backbone network and combines DCLB module and 2DTo3D module;

[0041] Through an encoder-decoder structure, the DCLB module extracts features and fuses spatial information from the input sliced ​​images. The 2DTo3D module expands the two-dimensional features into three-dimensional features through a broadcast mechanism, then concatenates them with the downsampled three-dimensional features, and further extracts spatial structure information through multi-layer three-dimensional convolution, thereby enhancing the model's ability to model three-dimensional data and generating rice CT image data of the middle section, which can accurately reflect the internal structural details between the two sliced ​​images. Attached Figure Description

[0042] Figure 1 This is a CT image of rice seedling stems from a method for generating rice CT data based on GAN networks.

[0043] Figure 2 This is a network model architecture for the generative model in a rice CT data generation method based on GAN networks.

[0044] Figure 3This refers to rice CT image data generated by a model in a rice CT data generation method based on GAN networks. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The present invention discloses a method for generating rice CT data based on GAN networks. In this embodiment, the method operates on an Ubuntu 22.04.2 operating system. The server is configured with 2 Intel XRONGold6133 CPUs at 2.5GHz, 4 NVIDIA GeForce RTX3090 24g TURBO graphics cards, 8 Samsung 32GB DDR4 / 2933recc memory, PyCharm 2023 development tools, PyTorch development framework, and Python 3.6 programming language.

[0047] The specific steps are as follows:

[0048] S1. Pre-treat rice seedlings.

[0049] The rice seedlings used were high-quality rice varieties that had been cultivated for 40 days. The pretreatment involved soaking the rice seedlings in a 15% cesium iodide solution.

[0050] S2. Acquire two CT images of rice seedling stems.

[0051] Data was acquired using a SkyScan 1272 scanner at an accelerating voltage of 50 kV, a current of 80 μA, and a detector exposure time of 2800 ms.

[0052] Figure 1 CT images of rice seedling stems collected.

[0053] S3. Standardize the two acquired CT images.

[0054] The two input rice slice images are standardized to ensure their size and format conform to the model requirements. Standardization involves adjusting the image resolution and grayscale range to meet the model's input specifications.

[0055] S4. Construct a rice CT data generation model with x2ct-gan as the backbone network and DCLB (Dilation Convolutional Layer Block) and 2DTo3D modules.

[0056] The generative model includes a 3D generator and a 3D discriminator, with the 3D generator employing a DCLB module for feature extraction; a 2DTo3D module is used to convert two-dimensional features into three-dimensional features.

[0057] The DCLB module utilizes four dilated convolutional kernels with different dilation rates (r=1, r=2, r=3, r=4) to process input features in parallel. By capturing spatial features through receptive fields at multiple scales, it enhances the ability to model fine structures in rice CT data. This module operates through the following steps:

[0058] (1) The input features are first reduced in size by a downsampling module, which halves the width and height of the original features, thereby achieving feature compression;

[0059] (2) Then, a convolutional layer is used for dimensionality reduction to reduce the number of channels to 1 / 4 of the original number of channels, so as to reduce computational complexity;

[0060] (3) The dimensionality-reduced features are convolved in parallel through four dilated convolutional layers, using different dilation rates r=1, r=2, r=3 and r=4 respectively, to capture spatial features at different scales;

[0061] (4) The outputs of the four dilated convolutional layers are added element by element to fuse multi-scale information;

[0062] (5) Then, the fused features are restored to the original number of channels through a full-channel convolutional layer, and the output feature map is added to the features before dimensionality reduction in step (2).

[0063] (6) Repeat steps (2) to (5) N times to output the feature map.

[0064] (7) Finally, the feature map is concatenated with the downsampled features from step (1) along the channel number dimension to generate the final feature map.

[0065] The 2DTo3D module is used to extend two-dimensional features to three-dimensional features. The specific steps are as follows:

[0066] (1) Two-dimensional feature broadcasting expansion: The input two-dimensional features are expanded along the third dimension (depth direction) through a broadcasting mechanism to generate a three-dimensional feature map with four layers of depth.

[0067] (2) Stitching operation: The expanded two-dimensional features are stitched together with the downsampled three-dimensional features so that the spatial information of the two-dimensional features and the three-dimensional features remains consistent during the information fusion process.

[0068] (3) Three-dimensional convolution processing: The spliced ​​features are processed through multi-layer three-dimensional convolution operations to further extract structural information in three-dimensional space and improve the model's ability to model three-dimensional data.

[0069] Figure 2 The network model architecture for generating the model.

[0070] S5. Train the generative model.

[0071] Specific methods include:

[0072] S51. Acquire pre-processed CT images of rice seedling stems;

[0073] S52. Create a CT image dataset from the collected CT images of rice seedling stems;

[0074] S53. Divide the CT image dataset into a training set and a test set. Use the training set to input the generative model for data modeling, and use the test set to test the generative model.

[0075] The optimal results were obtained by adjusting the parameters, as shown in the table below, which compares the experimental results of different models under the same parameters.

[0076] Model CosineSimilarity↑ PSNR↑ SSIM↑ x2ct-gan+S 0.9109536 49.2100563 0.6921427 x2ct-gan+B 0.9777877 53.7586789 0.7880505 DAGct-gan 0.9809820 54.7150993 0.8211958

[0077] Note: +s indicates a single view; +B indicates a two-view view; DAGct-gan is the name of the improved model; CosineSimilarity: cosine similarity; PSNR: peak signal-to-noise ratio; SSIM: structural similarity.

[0078] S6. Input the two standardized CT images into the trained generative model, extract features and fuse spatial information from the input CT images to generate the CT image data of the middle section of rice.

[0079] Two standardized rice slice images were input into a rice CT data generation model. The model's 3D generator, using an encoder-decoder structure and a DCLB module, extracted features and fused spatial information from the input slice images. A 2D-to-3D module then converted the two-dimensional features into three-dimensional features, ultimately generating the CT image data of the middle segment of the rice. The generated CT data exhibited high precision and quality, accurately reflecting the internal structural details between the two slice images. During the generation process, the model continuously optimized parameters and compared experimental results to achieve the best possible generation effect.

[0080] Figure 3 The paper presents rice CT image data generated by the model, demonstrating excellent detail and overall structure. In this way, the present invention effectively fills in the gaps between rice slice images, providing reliable data support for subsequent rice growth analysis and research.

[0081] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for generating CT data of rice based on a GAN network, characterized in that: The method comprises the following steps: S1, pretreating rice seedlings; S2, collecting CT images of two stems of the pretreated rice seedlings; S3, standardizing the collected two CT images; S4, constructing a CT data generation model for rice by taking x2ct-gan as a backbone network, combining a DCLB module and a 2DTo3D module; S5, training the generation model; S6, inputting the two standardized CT images into the trained generation model, extracting features of the input CT images and fusing spatial information, and generating CT image data of the middle section of the rice; The generation model comprises a 3D generator and a 3D discriminator, and the DCLB module is used for feature extraction in the encoder of the 3D generator; the 2DTo3D module is used for converting two-dimensional features into three-dimensional features; The DCLB module uses four dilated convolution kernels with different expansion rates to process input features in parallel, captures spatial features through a multi-scale receptive field, and enhances the modeling capability of fine structures in the CT data of the rice; the DCLB module can extract textures, edges and complex internal structures in the CT data of the rice through its multi-scale feature extraction mechanism, improve the restoration capability of the generation model for the fine structures of the stems of the rice seedlings, and ensure that the generated CT images of the rice have high resolution and accurate details; The specific operation steps of the DCLB module are as follows: (1) first, the size of the input feature is reduced through a downsampling module, and the width and height of the input feature are halved for feature compression; (2) a convolution layer is used for dimension reduction processing, and the number of channels is reduced to 1 / 4 of the original number of channels to reduce the calculation complexity; (3) the reduced features are processed through four dilated convolution layers in parallel, and different expansion rates r=1, r=2, r=3 and r=4 are used to capture spatial features of different scales; (4) the outputs of the four dilated convolution layers are added element by element to fuse multi-scale spatial features; (5) the fused multi-scale spatial features are restored to the original number of channels through a full-channel convolution layer, and then added to the features before dimension reduction in step (2) to obtain the output feature map; (6) repeat steps (2) to (5) N times to output the feature map; (7) the feature map output in step (6) is spliced with the feature after downsampling in step (1) in the channel dimension to generate the final feature map; The 2DTo3D module is used to expand two-dimensional features to three-dimensional features, and the specific steps are as follows: (1) two-dimensional feature broadcast expansion: the input CT image is expanded along the third dimension through a broadcast mechanism to obtain the expanded two-dimensional feature; (2) splicing operation: the expanded two-dimensional feature is spliced with the three-dimensional feature after downsampling processing, so that the spatial information of the two-dimensional feature and the three-dimensional feature is consistent during information fusion; (3) three-dimensional convolution processing: the spliced feature is processed through multi-layer three-dimensional convolution operation to further extract structural information in the three-dimensional space. 2.The method of claim 1, wherein the method is characterized by: The rice seedlings in S1 are 40-day-old rice seedlings of Liangyou 2, and the pretreatment is soaking the rice seedlings in a 15% cesium iodide solution. 3.The method of claim 1, wherein the method further comprises: determining a CT value of the water in the water droplet based on the CT value of the water in the water droplet and the CT value of the water in the water droplet in the CT image. The standardization processing in S3 includes adjusting the resolution and gray value range of the image.

4. The method of claim 1, wherein the method is based on a GAN network. In S5, the generated model is trained, and the specific method includes: S51, collecting the CT image of the pretreated rice seedling stem; S52, making the CT image of the rice seedling stem into a CT image data set; S53, dividing the CT image data set into a training set and a test set, inputting the training set into the generated model for data modeling, and testing the generated model with the test set.

5. The method of claim 4, wherein the method is based on a GAN network. In S52, the CT image of the rice seedling stem is made into a CT image data set, and the specific steps are: (1) The collected CT image of the rice seedling stem is center cropped with different sizes, and images with the same size are collected as the same batch; (2) The cropped image is scaled to 256*256; (3) 256 images in the same batch are stacked in order to form a three-dimensional CT image data, and the three-dimensional CT image data is rotated to obtain a CT image data set composed of 1180 images.

Images