Cotton young root image segmentation method, root length and root thickness measurement method and system

Abstract

The application provides a cotton seedling root image segmentation method, a root length and root width measurement method and system. The cotton seedling root image segmentation method comprises: acquiring an image containing cotton seedling roots; inputting the image into a trained improved U-Net network to obtain a cotton seedling root image segmentation result; the improved U-Net network comprises an encoder part, a decoder part and a skip connection part connecting the encoder part and the decoder part, wherein, for each level of feature calculation layer in the encoder part and the decoder part, the standard convolution in the feature calculation layer is replaced with a DSConv; for the encoder part, an SCSE module is embedded between each level of feature calculation layer and a pooling layer; and an ASPP module is integrated in the bottom bottleneck layer. The application can cope with the challenges brought by different target morphologies and variable background textures in the cotton seedling root image, realize accurate and rapid identification and segmentation of the cotton seedling roots in the germination disc environment, and realize calculation of the root length and root width phenotype parameters of the cotton seedling roots based on the identification and segmentation result.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural technology, specifically to a method for segmenting cotton seedling root images, a method and system for measuring root length and diameter. Background Technology

[0002] As an important economic crop, cotton's early growth status has a significant impact on its subsequent growth and development. Among these characteristics, the root system traits during the seedling stage are important phenotypic indicators for evaluating seed vigor and early growth capacity. The root length and root diameter of cotton seedlings are of great reference value in breeding screening and growth analysis.

[0003] Currently, under laboratory conditions, cotton seedlings typically grow in germination trays lined with filter paper. Their slender, easily bent shapes and colors that blend into the background make them prone to approaching and overlapping each other, increasing the difficulty of identifying and measuring them through image processing. Therefore, current cotton seedling root measurements primarily rely on manual methods. Manual measurement of cotton seedlings is inefficient, inconsistent, and lacks stability in complex environments.

[0004] In recent years, semantic segmentation methods based on deep learning have been gradually applied to plant phenotypic analysis. Among them, the U-Net structure has certain advantages in fine-grained target segmentation, but the existing U-Net network is still insufficient in terms of segmentation accuracy and detail preservation for small cotton seedlings in the germination plate environment.

[0005] Chinese invention patent application No. 202410731034.1 discloses a plant root cultivation device, cultivation method, and root image acquisition and segmentation method. The root image segmentation method, based on the disclosed plant root cultivation method, includes the following steps: S5, acquiring images of roots growing on a partial germination board, segmenting large-pixel images into uniform small-pixel images, and labeling the roots; S6, dividing the segmented small-pixel images into training and testing sets, preprocessing the training set images, and inputting them into a U-Net network for root segmentation training; preprocessing the testing set root images, inputting them into the trained U-Net model to detect the segmentation effect, and obtaining the final U-Net network; S7, preprocessing the root images to be segmented, and inputting them into the final U-Net network for segmentation; S8, slicing the segmented root images, extracting the central regions, and assembling them sequentially to the original root image size; S9, binarizing the output images according to a threshold, and outputting a black and white image. This root image segmentation method utilizes the U-Net network to segment the root systems of easily germinating medium-sized seed crops, including cucumber seedlings, such as those in the Cucurbitaceae family. However, the morphology of Cucurbitaceae seedlings differs from that of cotton seedlings, and the culture environment for seedlings in this invention application differs significantly from that of cotton germination trays, resulting in substantial differences in the background of the seedlings. Therefore, the conventionally structured U-Net network used in this invention application is not suitable for image segmentation of cotton seedlings, nor is it suitable for calculating the root thickness and length of cotton seedlings based on this segmentation. Summary of the Invention

[0006] To address at least one of the above technical problems, this invention provides a cotton seedling root image segmentation method, a root length and root thickness measurement method and system. Based on an improved U-Net network, it achieves accurate identification and segmentation of cotton seedling roots in a germination tray environment, and calculates the root thickness and root length phenotypic parameters of cotton seedling roots based on the identification and segmentation results.

[0007] A first aspect of the present invention provides a method for segmenting cotton seedling images, comprising: acquiring an image containing cotton seedlings; inputting the image into a trained improved U-Net network to obtain a cotton seedling image segmentation result; wherein the improved U-Net network includes an encoder part, a decoder part, and a skip connection part connecting the encoder part and the decoder part; wherein, for each level of the feature calculation layer in the encoder part and the decoder part, the standard convolution within it is replaced with a depthwise separable convolution (DSConv); for the encoder part, a spatial and channel cooperative squeeze excitation (SCSE) module is embedded between the feature calculation layer and the pooling layer at each level; and a dilated spatial convolution pooling pyramid (ASPP) module is integrated in the bottom bottleneck layer.

[0008] Preferably, the image containing cotton seedlings includes a germination disc, filter paper, and at least one cotton seedling placed on the filter paper.

[0009] In any of the above schemes, the cotton seedling image segmentation includes separating the cotton seedling from its background to obtain a binarized result containing only the cotton seedling target.

[0010] In any of the above schemes, a DSConv and a ReLU activation function located after the DSConv are provided in the bottom bottleneck layer, and the ASPP module is integrated in front of the DSConv in the bottleneck layer.

[0011] A second aspect of the present invention provides a method for measuring the root length and root diameter of cotton seedlings, comprising:

[0012] Step 1: Based on the binarization result containing cotton seedling targets obtained by the cotton seedling root image segmentation method, the skeleton of the cotton seedling root is extracted to obtain the skeleton representing the growth path of the cotton seedling root.

[0013] Step 2: For a cotton seedling root, based on its skeleton, traverse the skeleton pixels and calculate the Euclidean distance between two adjacent skeleton pixels. Sum all the Euclidean distances and obtain the root length of the cotton seedling root accordingly.

[0014] Step 3: For a cotton seedling root, based on its skeleton, locate a standard sampling point at a distance T from the top of the cotton seedling root. At the standard sampling point, construct a cross-section along the normal perpendicular to the growth direction of its skeleton, count the number of binarized target pixels covered by the cross-section, and obtain the root thickness of the cotton seedling root accordingly.

[0015] Preferably, in step 1, a morphological refinement algorithm is used to extract the skeleton of cotton seedlings.

[0016] In any of the above schemes, the preferred option is that, in step 2, the root length of the cotton seedling is obtained by combining the sum of all Euclidean distances with the pre-calibrated scale parameters of the camera used to capture images containing cotton seedlings.

[0017] In any of the above schemes, the preferred method is that in step 3, the root thickness of the cotton seedling is obtained based on the number of binarized target pixels covered by the cross section and the actual size corresponding to each pixel.

[0018] In any of the above schemes, the preferred value of T in step 3 is 25% to 35% of the root length.

[0019] A third aspect of the present invention provides a cotton seedling root image segmentation system, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to run the computer program to perform the cotton seedling root image segmentation method.

[0020] A fourth aspect of the present invention provides a cotton seedling root length and root diameter measurement system, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to run the computer program to perform the cotton seedling root length and root diameter measurement method.

[0021] The cotton seedling root image segmentation method, root length and root thickness measurement method and system of the present invention have the following beneficial effects:

[0022] 1. In the U-Net network, by replacing the standard convolutions in the feature calculation layers of each level in the encoder and decoder with depthwise separable convolutions (DSConv), the number of model parameters and redundant operations are significantly reduced without losing feature extraction accuracy. This effectively reduces the demand for computing resources and improves the processing speed of large-scale cotton root images in a laboratory environment.

[0023] 2. In the U-Net network, for the encoder part, a spatial and channel collaborative squeezing excitation (SCSE) module is embedded between the feature calculation layer and the pooling layer at each level. This enhances the expression of fine cotton root texture, can automatically suppress common impurities and light spot interference in laboratory imaging, improves the response intensity of root pixels in low contrast areas, and avoids the problem of missing cotton roots in complex backgrounds.

[0024] 3. In the U-Net network, the Spatial Convolutional Pooling Pyramid (ASPP) module is integrated in the bottom bottleneck layer, which can expand the receptive field without losing spatial resolution. This allows for the simultaneous capture of root targets with different diameters and distribution patterns, compensating for the spatial location loss caused by downsampling mapping and ensuring the coherence and consistency of the root topology in the segmentation results.

[0025] 4. Through the improved U-Net network, it can more robustly cope with the challenges brought by the different target morphologies and varied background textures in cotton seedling root images, achieve accurate identification and segmentation of cotton seedling roots in the germination plate environment, and calculate the root thickness and root length phenotypic parameters of cotton seedling roots based on the identification and segmentation results. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a commonly used standard U-Net network structure.

[0027] Figure 2 This is a schematic diagram of the improved U-Net network structure in a preferred embodiment of the cotton seedling root image segmentation method according to the present invention.

[0028] Figure 3 For the cotton seedling root image segmentation method according to the present invention, as follows Figure 2 A schematic diagram of the depth convolution process of the DSConv module in the improved U-Net network shown in the embodiment.

[0029] Figure 4 For the cotton seedling root image segmentation method according to the present invention, as follows Figure 2 A schematic diagram of the pointwise convolution process of the DSConv module in the improved U-Net network shown in the embodiment.

[0030] Figure 5 For the cotton seedling root image segmentation method according to the present invention, as follows Figure 2 A schematic diagram of the SCSE module structure in the improved U-Net network shown in the embodiment.

[0031] Figure 6 For the cotton seedling root image segmentation method according to the present invention, as follows Figure 2 A schematic diagram of the ASPP module structure in the improved U-Net network shown in the embodiment.

[0032] Figure 7 For the cotton seedling root image segmentation method according to the present invention, as follows Figure 2 The illustrated embodiment shows the binarized result obtained by segmenting three images containing cotton seedlings.

[0033] Figure 8 This is a flowchart illustrating a preferred embodiment of the cotton young root length and diameter measurement method according to the present invention.

[0034] Figure 9 For the cotton young root length and root diameter measurement method according to the present invention, as follows Figure 8 The illustrated embodiment is based on Figure 7 The third image in the diagram shows the skeleton extraction results and a schematic diagram of the cross-sectional root thickness measurement location. Detailed Implementation

[0035] To better understand the present invention, the present invention will be described in detail below with reference to specific embodiments.

[0036] First, a brief explanation of the commonly used standard U-Net network structure will be given.

[0037] like Figure 1 The diagram shows a commonly used standard U-Net network structure. The core structure of the U-Net network includes the encoder part (also known as the shrinking path) on the left, the decoder part (also known as the expanding path) on the right, and the jump connection part connecting the encoder part and the decoder part.

[0038] The encoder part aims to extract features from the image and understand its content. It consists of several repeating modules, each of which includes two 3x3 convolutional layers and one 2x2 pooling layer. The 3x3 convolutional layers are responsible for learning features from the image to obtain a feature map, while the 2x2 pooling layer is responsible for downsampling, halving the size of the feature map, and doubling the number of feature channels. The input image is processed in the encoder as follows: First-level feature calculation layer (i.e., the 3*3 convolutional layer in the first module, counting from top to bottom) → First-level pooling layer (i.e., the 2*2 pooling layer in the first module, counting from top to bottom) → Second-level feature calculation layer (i.e., the 3*3 convolutional layer in the second module, counting from top to bottom) → Second-level pooling layer (i.e., the 2*2 pooling layer in the second module, counting from top to bottom) → ... → Last-level feature calculation layer (i.e., the 3*3 convolutional layer in the last module, counting from top to bottom) → Last-level pooling layer (i.e., the 2*2 pooling layer in the last module, counting from top to bottom).

[0039] The decoder aims to restore the high-level but coarse feature maps extracted by the encoder to the size of the original image, and achieve pixel-level precise localization in the process. It also includes several repeating modules, each of which includes a 2*2 upsampling convolution and two 3*3 convolutional layers. The 2*2 upsampling convolution is responsible for doubling the size of the feature map and halving the number of channels, while the 3*3 convolutional layers are responsible for feature calculation. The input feature map is processed in the decoder as follows: the last upsampled convolution (i.e., the 2*2 upsampled convolution in the last module from top to bottom) → concatenated with the feature map from the last feature calculation layer of the encoder → the last feature calculation layer (i.e., the 3*3 convolutional layer in the last module from top to bottom) → the penultimate upsampled convolution (i.e., the 2*2 upsampled convolution in the second-to-last module from top to bottom) → concatenated with the feature map from the penultimate feature calculation layer of the encoder → the penultimate feature calculation layer (i.e., the 3*3 convolutional layer in the second-to-last module from top to bottom) → ... → the first upsampled convolution (i.e., the 2*2 upsampled convolution in the first module from top to bottom) → concatenated with the feature map from the first feature calculation layer of the encoder → the first feature calculation layer (i.e., the 3*3 convolutional layer in the first module from top to bottom).

[0040] The output of the first-level feature calculation layer of the decoder is passed through a 1*1 convolutional layer to output the final result.

[0041] Example 1

[0042] Currently, under laboratory conditions, cotton seedlings are typically grown in germination trays lined with filter paper. Multiple cotton seeds germinate and root on the filter paper, which is generally light-colored (e.g., white, light yellow, light blue, light gray), similar to the color of the cotton seedlings. Furthermore, cotton seedlings are slender and easily bend, leading to overlap or partial occlusion between different seedlings. Therefore, identifying and measuring cotton seedlings in this scenario is challenging. Additionally, laboratory experiments involve various environments to observe the impact of these environments on cotton germination and rooting, and the germination and rooting status is monitored at different time points. This necessitates processing large-scale images containing cotton seedlings for identification and measurement of root length and thickness. Moreover, during the photographing of cotton seedlings in the laboratory, interference from light spots and impurities is unavoidable, affecting the accuracy of seedling identification. When the standard U-Net network is used to identify cotton seedlings in the above scenario, it usually consumes more computing resources and cannot meet the speed requirements of large-scale image processing, and the recognition accuracy also needs to be improved.

[0043] Based on the above, this invention improves the standard U-Net network and provides a cotton seedling root image segmentation method based on the improved U-Net network.

[0044] A method for segmenting cotton seedling images includes: acquiring an image containing cotton seedlings; inputting the image into a trained improved U-Net network to obtain cotton seedling image segmentation results. Figure 2 As shown, the improved U-Net network includes an encoder part, a decoder part, and a skip connection part connecting the encoder part and the decoder part; the improvements include: (1) replacing the standard convolutions in the feature computation layer of each level in the encoder part and the decoder part with depthwise separable convolutions (DSConv); (2) embedding a spatial and channel co-extrusion excitation (SCSE) module between the feature computation layer and the pooling layer of each level in the encoder part; (3) integrating a dilated spatial convolution pooling pyramid (ASPP) module in the bottom bottleneck layer.

[0045] To balance the accuracy of cotton root segmentation with hardware computational efficiency, standard convolutions in each feature computation layer of both the encoder and decoder are replaced with depthwise separable convolutions (DSConv). Figure 3 and Figure 4As shown, DSConv extracts single-channel spatial features from the input feature map through deep convolutional layers, and then combines pointwise convolutional layers to achieve cross-channel information integration and feature map dimensionality transformation. During the downsampling feature extraction (implemented in the encoder part) and upsampling detail recovery (implemented in the decoder part), DSConv utilizes its unique weight and output probability distribution preservation properties to ensure the complete transmission of multi-scale geometric features of cotton seedlings while reducing network redundancy parameters and computational overhead. DSConv is integrated throughout the overall feature mapping process of the improved U-Net network, enabling the network to accurately capture fine seedling textures while significantly improving data processing speed. This meets the performance requirements for efficient, real-time analysis of large batches of cotton seedling images in a laboratory environment, achieving a balance. It should be noted that DSConv uses a 3*3 convolutional kernel followed by a ReLU activation function. Based on the previous explanation of the standard U-Net network, it should be understood that the feature calculation layer refers to the convolutional layer containing the two DSConv instances in each stage of the encoder and decoder parts.

[0046] Cotton seedlings exhibit diversity in size, texture, and distribution under different experimental conditions. To accommodate this diversity and suppress the impact of laboratory light spots and impurities on recognition accuracy, a Spatial and Channel Cooperative Squeezing Excitation (SCSE) module is embedded between the feature calculation layer and the pooling layer at each level of the encoder. The structure of the SCSE module is as follows: Figure 5 As shown, the SCSE module includes parallel spatial compression and channel activation (cSE) branches and channel compression and spatial activation (sSE) branches. The cSE branch compresses spatial information into channel descriptors through global average pooling, learns the interdependencies between channels, and achieves adaptive enhancement of key semantic feature channels. The sSE branch compresses the channel dimension into a spatial projection through 1*1 convolution and generates pixel-level spatial weight maps using activation functions to accurately locate target regions and suppress background noise interference. Finally, the SCSE module element-wise adds the outputs of the cSE and sSE branches. While maintaining efficient inference speed, the SCSE module enables simultaneous calibration of feature maps in both spatial and channel dimensions, enhancing the network's ability to represent fine root morphology. It also automatically and effectively suppresses common impurities and light spots in laboratory imaging, improves the response intensity of root pixels in low-contrast regions, and solves the problem of missed detection of cotton roots in complex backgrounds.

[0047] To eliminate the impact of morphological and dimensional differences in cotton seedlings caused by growth cycles and environments on image segmentation, and to improve the consistency of network representation of cotton seedlings with different diameters and distribution states in a laboratory environment, a Spatial Convolutional Pooling Pyramid (ASPP) module is integrated into the bottom bottleneck layer. Specifically, a DSConv activation function and a ReLU activation function following DSConv are set in the bottom bottleneck layer, and the ASPP module is integrated before DSConv in the bottleneck layer. The structure of the ASPP module is as follows. Figure 6 As shown, the ASPP module consists of a set of parallel feature extraction branches, including a standard convolutional path, multiple dilated convolutional paths with different sampling rates, and a global average pooling path. Dilated convolution, by inserting gaps of a specific proportion between convolutional kernel elements, significantly expands the receptive field without increasing the number of parameters or computational complexity, enabling the network to simultaneously cover the local texture of fine young roots and the extended morphology of thick main roots. The ASPP module performs channel-level concatenation and fusion of feature maps extracted by each parallel feature extraction branch, while using dilated convolutional paths with different sampling rates to generate multi-granularity receptive fields, which can compensate for the spatial location information lost during downsampling. The global average pooling path provides semantic constraints on local features for the entire image, ensuring that the coherence of the cotton young root structure is maintained even under complex background interference. The multi-scale feature extraction mechanism of the ASSP module can eliminate the influence of morphological and size differences of cotton seedlings caused by growth cycle and growth environment on image segmentation, significantly improve the network's representation consistency of cotton seedlings with different diameters and distribution states in a laboratory environment, and provide technical support for achieving efficient and accurate segmentation of large batches of cotton seedling images.

[0048] The improvements made to the standard U-Net network enable the improved U-Net network to efficiently represent the features of cotton seedlings in images, ensuring the accuracy of identification and segmentation of cotton seedlings of different morphologies in complex backgrounds, as well as the processing speed when processing large batches of images.

[0049] It should be noted that the structure of other parts of the improved U-Net network is the same as that of the standard U-Net network, such as using 2*2 convolution kernels for upsampling convolutions and using max pooling or average pooling layers; the training of the improved U-Net network can be carried out with reference to existing technologies, and will not be described in detail in this application.

[0050] It should be noted that the improved U-Net network in this application is proposed for image segmentation of cotton seedlings in images. It should be understood that the image should contain cotton seedlings. Furthermore, given that cotton seedlings grow in a germination tray culture environment lined with filter paper under laboratory experimental conditions, the image input to the improved U-Net network typically includes a germination tray, filter paper, and at least one cotton seedling placed on the filter paper. It should also be noted that the improved U-Net network in this application exhibits high segmentation accuracy and processing speed for the identification and segmentation of other plant seedlings in complex backgrounds.

[0051] When an image containing cotton seedlings is input into the improved U-Net network, the cotton seedlings are separated from their background, resulting in a binarized image containing only the cotton seedlings. It should be understood that obtaining the cotton seedling image segmentation result includes obtaining a binarized image containing only the cotton seedlings.

[0052] Figure 7 Three images containing cotton seedlings are shown, along with the cotton seedling image segmentation results obtained after inputting these three images into the improved U-Net network. Based on the three images containing cotton seedlings, it can be seen that: the color of the cotton seedlings is similar to that of the background (filter paper) they are in, and also similar to the color of the cotton seedlings at their tips; the images contain shadows of the cotton seedlings, light spots, and water bubbles formed below the filter paper, indicating that the background texture information of the cotton seedlings is varied; the morphology of the cotton seedlings varies, and the longer the cotton seedlings, the thinner and more easily bent they become; there are instances where cotton seedlings are close to each other, or even touching, with cotton seed hulls. The image segmentation results show that the image segmentation method of this application is not affected by changes in the background texture information of the cotton seedlings, the different morphologies of the cotton seedlings, or the positional relationship between the cotton seedlings, and can accurately segment the cotton seedlings from their background, obtaining a binary result containing only the cotton seedlings.

[0053] Example 2

[0054] like Figure 8 As shown, a method for measuring the root length and root diameter of cotton seedlings includes:

[0055] Step 1: Based on the binarization result containing cotton seedling targets obtained by the cotton seedling root image segmentation method, the skeleton of the cotton seedling root is extracted to obtain the skeleton representing the growth path of the cotton seedling root.

[0056] Step 2: For a cotton seedling root, based on its skeleton, traverse the skeleton pixels and calculate the Euclidean distance between two adjacent skeleton pixels. Sum all the Euclidean distances and obtain the root length of the cotton seedling root accordingly.

[0057] Step 3: For a cotton seedling root, based on its skeleton, locate a standard sampling point at a distance T from the top of the cotton seedling root. At the standard sampling point, construct a cross-section along the normal perpendicular to the growth direction of its skeleton, count the number of binarized target pixels covered by the cross-section, and obtain the root thickness of the cotton seedling root accordingly.

[0058] Preferably, in step 1, a morphological thinning algorithm is used to extract the skeleton of the cotton seedling roots; in step 2, the root length of the cotton seedling roots is obtained based on the sum of all Euclidean distances and the pre-calibrated scale parameter of the camera used to capture images containing cotton seedling roots; in step 3, the value of T is 25% to 35% of the root length, and the root thickness of the cotton seedling roots is obtained based on the number of binarized target pixels covered by the cross section and the actual size corresponding to each pixel.

[0059] like Figure 9 As shown, with Figure 7 Taking the third image containing cotton seedlings as an example, image segmentation is performed to obtain a binarized result containing cotton seedling targets; then, based on this binarized result, the cotton seedling root length and root diameter measurement method is executed for each cotton seedling target; after step 1, a skeleton characterizing the growth path of cotton seedlings is obtained; through step 2, the root length of each cotton seedling is obtained; based on the root length of the cotton seedlings, with T being 30% of the root length, a cross-section for measuring the root diameter of the cotton seedlings is constructed in step 3. Figure 9 (as shown by the red line), and the root thickness of the cotton seedling is obtained based on the cross-section. For Figure 7 The time taken to segment the third image containing cotton seedlings was 1.02s; the time taken to obtain the root length and root diameter measurement results was 5.93s.

[0060] Meanwhile, the image segmentation method and the root length and root thickness measurement method were tested using no less than 100 images containing cotton seedlings. Each image contained no less than 30 cotton seedlings. The average time for image segmentation was 0.801s, and the average time for root length and thickness measurement was 4.55s.

[0061] The above results demonstrate that the cotton seedling root image segmentation method described in this application can meet the processing requirements of large-scale cotton seedling root images in a laboratory environment.

[0062] Example 3

[0063] This embodiment provides a cotton seedling root image segmentation system, including a processor and a memory. The memory is used to store a computer program, and the processor is used to run the computer program to execute the cotton seedling root image segmentation method.

[0064] This embodiment also provides a cotton seedling root length and root diameter measurement system, including a processor and a memory. The memory is used to store a computer program, and the processor is used to run the computer program to execute the cotton seedling root length and root diameter measurement method.

[0065] 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 foregoing embodiments have described the present invention in detail, those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein, and these substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for segmenting cotton seedling root images, comprising: Obtain an image containing cotton seedlings, input the image into a trained improved U-Net network, and obtain the cotton seedling image segmentation result; The improved U-Net network is an improvement upon the standard U-Net network, comprising an encoder section, a decoder section, and a skip connection section connecting the encoder and decoder sections. Its key features include: improvements to the standard U-Net such that, for each level of the feature computation layer in the encoder and decoder sections, the standard convolutions are replaced with depthwise separable convolutions; for the encoder section, a spatial and channel-coordinated excitation module is embedded between the feature computation layer and the pooling layer at each level; and a dilated spatial convolution pooling pyramid module is integrated within the bottom bottleneck layer. In other words: The encoder section includes four repeating and sequentially arranged encoding modules. Each encoding module includes a feature calculation layer, a spatial and channel co-extrusion excitation module, and a max pooling layer arranged sequentially from front to back. The feature calculation layer in each encoding module includes two sequentially connected 3*3 depthwise separable convolutions + ReLU activation functions. That is, the data input to the feature calculation layer of a certain encoding module is processed by two sequentially connected 3*3 depthwise separable convolutions + ReLU activation functions to obtain the feature map of that encoding module. The feature map of that encoding module is then input into the spatial and channel co-extrusion excitation module of the same encoding module for processing and then input into the max pooling layer of the same encoding module for downsampling. Two 3*3 depth separable convolution + ReLU activation functions are set in the bottom bottleneck layer. The dilated spatial convolution pooling pyramid module is set between the two 3*3 depth separable convolution + ReLU activation functions. That is, the output of the encoder part is processed by the first 3*3 depth separable convolution + ReLU activation function in the bottom bottleneck layer and then enters the dilated spatial convolution pooling pyramid module for processing. The output of the dilated spatial convolution pooling pyramid module is processed by the second 3*3 depth separable convolution + ReLU activation function in the bottom bottleneck layer and then input to the decoder part. The decoder section includes four repeating and sequentially arranged decoding modules. Each module includes a 2*2 upsampled convolution and a feature calculation layer arranged sequentially from front to back. The feature calculation layer in each decoding module includes two sequentially connected 3*3 depthwise separable convolutions with ReLU activation functions. That is, the data input to a decoding module is first processed by a 2*2 upsampled convolution. The output of the 2*2 upsampled convolution in the decoding module is concatenated with the feature map output from the feature calculation layer of the same level encoding module. Then, it is input into the two sequentially connected 3*3 depthwise separable convolutions with ReLU activation functions in the decoding module to obtain the feature map of the decoding module.

2. The cotton seedling root image segmentation method as described in claim 1, characterized in that: The image containing cotton seedlings includes a germination disc, filter paper, and at least one cotton seedling placed on the filter paper.

3. The cotton seedling root image segmentation method as described in claim 1, characterized in that: Cotton seedling root image segmentation involves separating the cotton seedling roots from their surrounding background to obtain a binarized result containing only the cotton seedling root target.

4. A method for measuring the root length and diameter of cotton seedlings, characterized in that: include: Step 1: Based on the binarization result containing the cotton seedling target obtained by the cotton seedling root image segmentation method as described in any one of claims 1-3, the skeleton of the cotton seedling root is extracted to obtain the skeleton characterizing the growth path of the cotton seedling root. Step 2: For a cotton seedling root, based on its skeleton, traverse the skeleton pixels and calculate the Euclidean distance between two adjacent skeleton pixels. Sum all the Euclidean distances and obtain the root length of the cotton seedling root accordingly. Step 3: For a cotton seedling root, based on its skeleton, locate a standard sampling point at a distance T from the top of the cotton seedling root. At the standard sampling point, construct a cross-section along the normal perpendicular to the growth direction of its skeleton, count the number of binarized target pixels covered by the cross-section, and obtain the root thickness of the cotton seedling root accordingly.

5. The method for measuring the root length and root diameter of cotton seedlings as described in claim 4, characterized in that: In step 1, a morphological refinement algorithm is used to extract the skeleton of cotton seedlings.

6. The method for measuring the root length and root diameter of cotton seedlings as described in claim 4, characterized in that: In step 2, the root length of the cotton seedling is obtained by combining the sum of all Euclidean distances with the pre-calibrated scale parameters of the camera used to capture images containing the cotton seedling roots.

7. The method for measuring the root length and root diameter of cotton seedlings as described in claim 4, characterized in that: In step 3, the root thickness of the cotton seedling is obtained based on the number of binarized target pixels covered by the cross section and the actual size corresponding to each pixel.

8. The method for measuring the root length and root diameter of cotton seedlings as described in claim 4, characterized in that: In step 3, the value of T is 25% to 35% of the root length.

9. A cotton seedling root image segmentation system, comprising a processor and a memory, wherein the memory is used to store a computer program, characterized in that: The processor is used to run the computer program to perform the cotton seedling root image segmentation method as described in any one of claims 1-3.

10. A system for measuring the length and thickness of cotton seedling roots, comprising a processor and a memory, wherein the memory is used to store a computer program, characterized in that: The processor is used to run the computer program to perform the method for measuring the root length and root diameter of cotton seedlings as described in any one of claims 4-8.

Images