Global distribution knowledge-guided cabbage leaf ball section type region division method and device and medium

Abstract

The invention provides a global distribution knowledge guided cabbage leaf ball section type region division method and device and a medium, and belongs to the technical field of agricultural informatization. The method comprises the following steps: collecting spectral information of regions with different color types of a cabbage leaf ball profile, and performing feature analysis; calculating hyperspectral characteristic indexes of different types of regions of the cabbage leaf ball section based on the original wave band reflectivity; the method comprises the following steps: acquiring a hyperspectral image of a cabbage leaf ball section by using a hyperspectral imager, and obtaining initial spectral features of different types of regions of the cabbage leaf ball section by using a hyperspectral feature index calculation method; constructing a Chinese cabbage leaf ball section type region division model; inputting the initial spectral features into a Chinese cabbage leaf ball section type region division model to train the model; and collecting a to-be-detected cabbage leaf ball section hyperspectral image, and inputting the image into the model for region division. According to the invention, rapid and accurate division and identification of different types of areas of the cabbage leaf ball profile can be realized, and the division precision and the working efficiency are greatly improved.

Description

Technical Field

[0001] The invention relates to a method, a device and a medium for dividing cabbage head section type regions guided by global distribution knowledge, and belongs to the technical field of agricultural informatization. Background Art

[0002] Cabbage is one of the important vegetable crops in my country with high nutritional value. The cabbage head is not only the edible organ of cabbage, but also the main organ for photosynthesis and maintaining the growth and development of the body. Whether in production and planting or breeding of new varieties, the cabbage head section phenotype contains key genetic information, and the quantitative acquisition of its phenotypic traits is of great significance. Among them, the color area type of the cabbage head section is one of the most important phenotypic traits of cabbage, which is directly related to the nutritional quality, taste and commerciality of cabbage. Yellow-heart and orange-red-heart cabbage are stronger than traditional white-heart cabbage in terms of nutritional value and market recognition. The inner leaf colors of cabbage heads are mainly yellow, orange-red, white, and purple, and the intermediate transition types are orange-yellow, light yellow, golden yellow, etc. However, the complex structure of the cabbage head section makes it difficult to quickly identify the color area type of the cabbage head. Therefore, the identification and quantification of different color area types of cabbage head sections are difficult problems in the current research on cabbage variety improvement.

[0003] Traditionally, the classification of cabbage head section area types is mainly done by visual identification and ruler measurement, which is time-consuming, laborious, and highly subjective. With the rapid development of remote sensing technology, hyperspectral remote sensing technology can accurately reflect the subtle characteristic differences of the spectrum of ground objects due to its multi-band and narrow characteristics. The different color types of cabbage head areas will cause changes in spectral absorption / reflection due to factors such as cell structure and pigment content, which makes it possible to achieve rapid and accurate monitoring of different types of cabbage head sections.

[0004] Defects of existing technologies: Since the imaging process of hyperspectral imagers is easily affected by the device itself and the external environment, the collected spectral images are usually subject to various noise pollution, which brings great challenges to the spectral distinguishability of different types of cabbage head sections. Therefore, it is necessary to explore a more efficient, fast and accurate method to achieve fine division of different types of cabbage head sections by combining the powerful feature expression learning ability of deep learning. Summary of the invention

[0005] The purpose of the present invention is to provide a method, device and medium for dividing cabbage head section type regions guided by global distributed knowledge, so as to realize fast and accurate division and identification of different types of regions on the scale of cabbage head section.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: A method for dividing cabbage head section type regions guided by global distribution knowledge, comprising: Collect spectral information of different color types of cabbage head sections, and analyze the spectral reflectance characteristics of different color types of areas; The hyperspectral characteristic index of different types of cabbage head profiles was calculated based on the original band reflectance to describe different types of cabbage head profiles. A hyperspectral imager was used to collect hyperspectral images of cabbage head sections, and the initial spectral characteristics of different types of cabbage head sections were obtained using the hyperspectral feature index calculation method. A cabbage head section type regional division model is constructed, wherein the cabbage head section type regional division module uses a UNet network as a basic network, comprises an encoder and a decoder, a downsampling module of the encoder and an upsampling splicing module of the decoder are jump-connected through a global distribution knowledge encoding module, and a feature encoding module with category contrast constraints is added to the last layer of the decoder; The initial spectral features are input into the cabbage head profile type area division model to train the model, and the joint loss function is minimized by gradient back propagation to achieve model optimization; The hyperspectral image of the cabbage head section to be tested is collected and input into the optimized cabbage head section type regional division model for regional division.

[0007] Preferably, the spectral reflectance characteristics of different color type regions are analyzed in the following manner: Collect the specific spectral features of the cabbage leaf section irradiated by light to obtain the original band reflectance; The original band reflectance is transformed into the first-order differential to obtain the first-order differential spectral characteristics; According to the original spectral reflectance characteristics and the first-order differential spectral characteristics, the different color type areas of the cabbage head are analyzed, and the characteristic parameters based on the high-spectral position are extracted, including spectral absorption characteristic parameters and spectral three-edge characteristic parameters; the spectral absorption characteristic parameters are blue valley position, blue valley reflectivity, green peak position, green peak reflectivity, red valley position, and red valley reflectivity; the spectral three-edge characteristic parameters include red edge position, red edge amplitude, red edge area, blue edge position, blue edge amplitude, and blue edge area.

[0008] Preferably, the hyperspectral characteristic index includes: pigment ratio index, Carter index, normalized pigment difference index, vegetation aging reflectance index, structure-insensitive pigment index, overgreen vegetation index; The calculation method is as follows: Pigment ratio index: ; Carter Index: ; Normalized pigment difference index: ; Vegetation senescence reflectance index: ; Structure-insensitive pigment index: ; Green vegetation index: ; in, Indicates wavelength The corresponding spectral reflectance.

[0009] Preferably, the cabbage head section type area division module includes, in sequence, a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a first upsampling splicing module, a second upsampling splicing module, a third upsampling splicing module, a fourth upsampling splicing module, and a feature encoding module of category comparison constraints; The first down-sampling module is jump-connected with the fourth up-sampling splicing module through a global distribution knowledge encoding module, the second down-sampling module is jump-connected with the third up-sampling splicing module through a global distribution knowledge encoding module, the third down-sampling module is jump-connected with the second up-sampling splicing module through a global distribution knowledge encoding module, and the fourth down-sampling module is jump-connected with the first up-sampling splicing module through a global distribution knowledge encoding module.

[0010] Preferably, the output features of the first downsampling module of the cabbage head section type region division module are input into the second downsampling module through maximum pooling, the output features of the second downsampling module are input into the third downsampling module through maximum pooling, the output features of the third downsampling module are input into the fourth downsampling module through maximum pooling, and the output features of the fourth downsampling module are input into the fifth downsampling module through maximum pooling; The output features of the fourth down-sampling module are spliced ​​with the output features of the fifth down-sampling module through the global distribution knowledge encoding module and then used as the input of the first up-sampling splicing module. The output features of the third downsampling module are spliced ​​with the output features of the first upsampling splicing module through the global distribution knowledge encoding module as the input of the second upsampling splicing module. The output features of the second downsampling module are spliced ​​with the output features of the second upsampling splicing module through the global distribution knowledge encoding module and then used as the input of the third upsampling splicing module. The output features of the first downsampling module are spliced ​​with the output features of the third upsampling splicing module through the global distribution knowledge encoding module as the input of the fourth upsampling splicing module. The fourth upsampling concatenation module outputs features as input to the feature encoding module with category contrast constraints.

[0011] Preferably, the global distribution knowledge encoding module processes the feature graph as follows: The input feature map undergoes the first 1×1 convolution, and the output feature map of the first 1×1 convolution is then processed by the softmax function and connected to the output feature map of the first 1×1 convolution. The connected feature map is multiplied by the input feature map through the second 1×1 convolution and serves as the final output of the module.

[0012] Preferably, the feature encoding module of the category contrast constraint is based on the semantic information feature vector of the last layer of the encoder, and calculates the cluster center of each category by a spatial constraint clustering method. During the training process, spatial constraint clustering is performed every 10 epochs to update the cluster center of each category.

[0013] Preferably, the joint loss function includes a loss function of a cross entropy loss and a class contrast loss, which is as follows: , , , in, represents the cross entropy loss, represents the class contrast loss, represents the predicted value of the network model, represents the true value of the corresponding point, Represents the semantic information features extracted by the last layer of the network, Indicates that it is constantly updated The average feature vector over the categories, represents the number of categories, is the function representative, Indicates distance, using Euclidean distance, Indicates the similarity between feature vectors.

[0014] A cabbage head section type area division device guided by global distributed knowledge comprises a processor and a memory storing program instructions, wherein the processor is configured to execute the cabbage head section type area division method guided by global distributed knowledge when running the program instructions.

[0015] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the cabbage head section type area division method guided by global distribution knowledge.

[0016] The advantages of the present invention are: The present invention deeply analyzes the spectral reflectance characteristics of different color type areas of the cabbage head section, enhances them using the hyperspectral feature index, and combines them with a deep learning network for fine division. It can achieve fast and accurate division and identification of different types of areas of the cabbage head section, greatly improving the division accuracy and work efficiency.

[0017] The traditional method mainly uses naked eye judgment and ruler measurement, which is time-consuming, labor-intensive and highly subjective. However, the present invention uses automated and intelligent technical means to achieve rapid division of cabbage head section type areas, greatly saving manpower and time costs.

[0018] The research results of the present invention provide a novel and efficient method for high-throughput phenotypic trait research and high-quality gene screening of cabbage crops. By accurately dividing the cabbage head section type area, it can provide important data support and decision-making basis for the breeding of excellent cabbage varieties and gene function research in the agricultural field. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0020] Figure 1 It is a schematic diagram of the process of the present invention.

[0021] Figure 2 Schematic diagram of the module for dividing cabbage head profile types into regions.

[0022] Figure 3 Schematic diagram of the global distributed knowledge encoding module.

[0023] Figure 4 This is the spectral curve of different color types of cabbage head.

[0024] Figure 5 It is the first-order differential spectral curve of different color types of cabbage head.

[0025] Figure 6 These are the spatial distribution characteristics of different types of regions in the cabbage head section. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Example 1 like Figure 1 As shown, a cabbage head section type region division method guided by global distribution knowledge includes: S1: Collect spectral information of different color types of areas in the cabbage head section, and analyze the spectral reflectance characteristics of different color types of areas.

[0028] As a refinement of the above embodiment, the spectral reflectance characteristics of different color type regions are analyzed in the following specific manner: S101: collecting specific spectral features of the light irradiating the cabbage bulb section to obtain the original band reflectance; S102: performing a first-order differential transformation on the original band reflectance to obtain a first-order differential spectral feature; S103: Analyze different color type areas of cabbage head according to the original spectral reflectance characteristics and the first-order differential spectral characteristics, and extract characteristic parameters based on the hyperspectral position, including spectral absorption characteristic parameters and spectral three-edge characteristic parameters; the spectral absorption characteristic parameters are blue valley position, blue valley reflectivity, green peak position, green peak reflectivity, red valley position, and red valley reflectivity; the spectral three-edge characteristic parameters include red edge position, red edge amplitude, red edge area, blue edge position, blue edge amplitude, and blue edge area.

[0029] Specifically, when light is irradiated onto the cabbage head section, it is absorbed and scattered by various pigments in the head section, forming specific spectral characteristics of different color types in the cabbage head section within the 400-2500nm band (such as Figure 4 In order to eliminate the influence of background noise and illumination on the target signal, a first-order differential transformation is performed on the basis of analyzing the reflectivity of the original band to obtain the first-order differential spectral characteristics (as shown in Figure 5 As shown), in order to better discover the law of plant spectrum changes; the first-order differential spectrum characteristic formula is as follows: In the formula, Indicates band The wavelength of Indicates wavelength The spectral reflectance, is the wavelength arrive The difference is determined by the spectral sampling interval.

[0030] By conducting a detailed analysis of the different color types of cabbage head regions from the aspects of original spectral reflectance characteristics and first-order differential spectral characteristics, characteristic parameters based on hyperspectral position were extracted, including spectral absorption characteristic parameters and spectral trilateral characteristic parameters (as shown in Table 1), so as to clarify the differences in spectral characteristics of the four typical types of cabbage head profiles, and provide a theoretical basis for the subsequent construction of spectral characteristic index to identify the regional type of cabbage head profile.

[0031] Table 1 Characteristic parameters based on hyperspectral position Optionally, in this embodiment, an ASD FieldSpec HandHeldTM ground spectrometer manufactured by ASD (Analytical Spectral Devices Inc) of the United States is used, and its spectral wavelength range is 350-2500nm, and the sampling interval is 1nm, and the spectrum of different color types of cabbage head areas is measured. In order to reduce the influence of the change of the solar altitude angle on the spectral measurement results during measurement, the field measurement time is selected between 10:00 and 14:00 when it is sunny and windless. At the same time, the standard plate is calibrated every 15 minutes or so, and the instrument probe is kept vertically downward during measurement, and the spectrum is measured 10 times at each sampling point.

[0032] The spectral resolution of the spectrometer for raw data at 700nm is 3.5nm, and the actual data obtained is the sampling data with 1nm intervals exported by ASDViewSpecPro software. Since the spectrum has large noise after 950nm, the present invention selects the reflection spectrum band in the range of 350-950nm as the effective spectral data for subsequent analysis; S2: The hyperspectral characteristic index of different types of areas in the cabbage head profile is calculated based on the original band reflectance to describe different types of areas in the cabbage head profile.

[0033] As a refinement of the above embodiment, most of the sensitive bands of different color types of cabbage head areas are within the range of "green peak" and "red edge", and their spectral absorption characteristics within the corresponding range have different values. For the better difference band characteristic parameters between different color types of cabbage head areas, the existing spectral parameter calculation method is referred to and different bands are processed and combined to calculate 6 hyperspectral characteristic indexes, which can enhance the spectral characteristics of the target type area and weaken the influence of other factors, thereby improving the expression ability and distinguishability of the spectral characteristics of different types of areas.

[0034] The hyperspectral characteristic indexes include: pigment ratio index, Carter index, normalized pigment difference index, vegetation aging reflectance index, structure-insensitive pigment index, and overgreen vegetation index; The calculation method is as follows: Pigment ratio index: ; Carter Index: ; Normalized pigment difference index: ; Vegetation senescence reflectance index: ; Structure-insensitive pigment index: ; Green vegetation index: ; in, Indicates wavelength The corresponding spectral reflectance.

[0035] S3: using a hyperspectral imager to collect a hyperspectral image of the cabbage head section, and using a hyperspectral feature index calculation method to obtain initial spectral features of different types of regions of the cabbage head section. The hyperspectral imager is an S185 hyperspectral imager.

[0036] As a refinement of the above embodiment, after defining the hyperspectral characteristic index of cabbage head, the initial characteristic map of cabbage head section is calculated using hyperspectral image. However, since the collected hyperspectral image itself contains noise, the phenomenon of "different objects with the same spectrum" or "same object with different spectrum" is serious. It is difficult to obtain satisfactory segmentation quality by simply using the initial characteristic map of cabbage head section for training or prediction of recognition model. At the same time, by analyzing the spatial distribution characteristics between different types of regions of cabbage head section (such as Figure 6 As shown in the figure, it can be concluded that different types of regions show a regular hierarchical spatial distribution in the cabbage head section. For example, the roots are generally located in the bottom local area, and the green leaves are usually distributed around the head. These regular hierarchical spatial distribution characteristics provide prior knowledge for the fine division of type regions. By utilizing the prior knowledge of hierarchical spatial distribution coupled with the powerful feature learning ability of deep learning to optimize the spectral features, the fine division of different types of regions in the cabbage head section can be achieved.

[0037] S4: Construct a cabbage head section type regional division model. The cabbage head section type regional division module uses the UNet network as the basic network, including an encoder and a decoder. The downsampling module of the encoder and the upsampling splicing module of the decoder are jump-connected through a global distributed knowledge encoding module. The last layer of the decoder adds a feature encoding module with category contrast constraints.

[0038] As a refinement of the above embodiment, the entire deep learning network framework uses the initial spectral features of the cabbage head section as input information. The encoder part can gradually reduce the spatial resolution and extract high-level abstract features by stacking convolution and pooling layers, while the decoder part restores these abstract features to the same resolution as the input image by upsampling and skip connections. In view of the hierarchical spatial distribution characteristics of different types of regions in the cabbage head section, a global distribution knowledge encoding module is added on the basis of the traditional UNet model. This module effectively encodes the spatial distribution of different types of regions in the overall cabbage section by calculating the global feature map of the correlation between all points, realizing the fusion of local features and global features, thereby improving the model detection accuracy.

[0039] Specifically, Figure 2As shown, the cabbage head section type area division module includes, in sequence, a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a first upsampling splicing module, a second upsampling splicing module, a third upsampling splicing module, a fourth upsampling splicing module, and a feature encoding module of category comparison constraints.

[0040] The first down-sampling module is jump-connected with the fourth up-sampling splicing module through a global distribution knowledge encoding module, the second down-sampling module is jump-connected with the third up-sampling splicing module through a global distribution knowledge encoding module, the third down-sampling module is jump-connected with the second up-sampling splicing module through a global distribution knowledge encoding module, and the fourth down-sampling module is jump-connected with the first up-sampling splicing module through a global distribution knowledge encoding module.

[0041] The input and output relationship of each module is as follows: The output features of the first downsampling module of the cabbage head section type area division module are input into the second downsampling module through maximum pooling, the output features of the second downsampling module are input into the third downsampling module through maximum pooling, the output features of the third downsampling module are input into the fourth downsampling module through maximum pooling, and the output features of the fourth downsampling module are input into the fifth downsampling module through maximum pooling; the output features of the fourth downsampling module are spliced ​​with the output features of the fifth downsampling module through the global distribution knowledge encoding module as the input of the first upsampling splicing module, the output features of the third downsampling module are spliced ​​with the output features of the first upsampling splicing module through the global distribution knowledge encoding module and the output features of the first upsampling splicing module as the input of the second upsampling splicing module, the output features of the second downsampling module are spliced ​​with the output features of the second upsampling splicing module through the global distribution knowledge encoding module and the output features of the second upsampling splicing module as the input of the third upsampling splicing module, the output features of the first downsampling module are spliced ​​with the output features of the third upsampling splicing module as the input of the fourth upsampling splicing module, and the output features of the fourth upsampling splicing module are input into the feature encoding module of category comparison constraints.

[0042] The downsampling module and the upsampling splicing module both include two 3×3 convolutions. The feature map input to the downsampling module is processed by two 3×3 convolutions in sequence as the output feature map of the downsampling module; the upsampling splicing module obtains the spliced ​​feature map and then processes it by two 3×3 convolutions in sequence as the output feature map of the upsampling splicing module. The output feature map of the fourth upsampling splicing module is input to the feature encoding module through 1×1 convolution.

[0043] As a refinement of the above embodiment, the global distribution knowledge encoding module processes the feature graph as follows: like Figure 3As shown in the figure, the input feature map undergoes the first 1×1 convolution, the output feature map of the first 1×1 convolution is then processed by the softmax function and connected with the output feature map of the first 1×1 convolution, and the connected feature map is multiplied with the input feature map by the second 1×1 convolution as the final output of the module.

[0044] As a refinement of the above embodiment, the feature encoding module of the category contrast constraint is based on the semantic information feature vector of the last layer of the encoder, and calculates the cluster center of each category through the spatial constraint clustering method. During the training process, spatial constraint clustering is performed every 10 epochs to update the cluster center of each category.

[0045] S5: The initial spectral features are input into the cabbage head profile type area division model to train the model, and the joint loss function is minimized by gradient back propagation to achieve model optimization.

[0046] As a refinement of the above embodiment, a loss function of the joint cross entropy loss and the class contrast loss is used to measure the difference between the model prediction result and the true value. During the model training process, the model error is calculated and the gradient back propagation is used to minimize the joint loss function, thereby optimizing the model.

[0047] The cross entropy loss function can effectively guide the model to learn to approximate the probability distribution of the true label by minimizing the negative log-likelihood loss, better deal with the importance differences of different categories and categories that need to be ignored, thereby significantly improving the performance of the model in classification tasks.

[0048] After the cross entropy loss function gradually tends to converge, in order to further improve the expressiveness and distinguishability of the learned feature vectors, a feature encoding module with category contrast constraints is added. Specifically, based on the semantic information feature vector of the last layer of the network, the spatial constraint clustering method is used to calculate the cluster center of each category (i.e., the average feature vector), and then the spatial constraint clustering is performed every 10 epochs during the training process to update the cluster center of each category. The spatial constraint clustering method takes into account the similarity between spatial position information and spectral information at the same time, so that the clustering results are more in line with the spatial distribution characteristics of different types of regions in the cabbage profile. The purpose of this module is to explicitly constrain the descriptive power of feature vectors in high-dimensional feature space, and by dynamically constructing and maintaining cluster centers of different types of regions, the feature vectors corresponding to the same type of regions are closer to their cluster centers and farther away from other cluster centers, thereby enhancing the similarity of feature vectors between the same categories and the difference of feature vectors between different categories. The joint loss function includes the loss functions of cross entropy loss and category contrast loss, which are as follows: , , , in, represents the cross entropy loss, represents the class contrast loss, represents the predicted value of the network model, represents the true value of the corresponding point, Represents the semantic information features extracted by the last layer of the network, Indicates that it is constantly updated The average feature vector over the categories, represents the number of categories, is the function representative, Indicates distance, using Euclidean distance, Indicates the similarity between feature vectors.

[0049] S6: Collect the hyperspectral image of the cabbage head section to be detected, and input the optimized cabbage head section type regional division model to perform regional division.

[0050] Example 2 The embodiment of the present disclosure provides a cabbage head section type area division device guided by global distributed knowledge, including a processor and a memory. Optionally, the device may also include a communication interface and a bus. The processor, the communication interface, and the memory may communicate with each other through the bus. The communication interface may be used for information transmission. The processor may call the logic instructions in the memory to execute the cabbage head section type area division method guided by global distributed knowledge of the above embodiment.

[0051] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0052] The memory, as a computer-readable storage medium, can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor 304 executes the function application and data processing by running the program instructions / modules stored in the memory, that is, the cabbage head section type area division method guided by global distribution knowledge in the above embodiment is implemented.

[0053] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory.

[0054] The embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above-mentioned method for dividing cabbage head section type regions guided by global distributed knowledge.

[0055] The computer-readable storage medium mentioned above may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0056] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for partitioning cabbage head section types based on global distribution knowledge, characterized in that: include: Collect spectral information of different color types of cabbage head sections, and analyze the spectral reflectance characteristics of different color types of areas; The hyperspectral characteristic index of different types of cabbage head profiles was calculated based on the original band reflectance to describe different types of cabbage head profiles. A hyperspectral imager was used to collect hyperspectral images of cabbage head sections, and the initial spectral characteristics of different types of cabbage head sections were obtained using the hyperspectral feature index calculation method. A cabbage head section type regional division model is constructed, wherein the cabbage head section type regional division module uses a UNet network as a basic network, comprises an encoder and a decoder, a downsampling module of the encoder and an upsampling splicing module of the decoder are jump-connected through a global distribution knowledge encoding module, and a feature encoding module with category contrast constraints is added to the last layer of the decoder; The initial spectral features are input into the cabbage head profile type area division model to train the model, and the joint loss function is minimized by gradient back propagation to achieve model optimization; The hyperspectral image of the cabbage head section to be tested is collected and input into the optimized cabbage head section type regional division model for regional division.

2. The method for dividing cabbage head section type regions guided by global distribution knowledge according to claim 1 is characterized in that: The spectral reflectance characteristics of different color types are analyzed as follows: Collect the specific spectral features of the cabbage leaf section irradiated by light to obtain the original band reflectance; The original band reflectance is transformed into the first-order differential to obtain the first-order differential spectral characteristics; According to the original spectral reflectance characteristics and the first-order differential spectral characteristics, the different color type areas of the cabbage head are analyzed, and the characteristic parameters based on the high-spectral position are extracted, including spectral absorption characteristic parameters and spectral three-edge characteristic parameters; the spectral absorption characteristic parameters are blue valley position, blue valley reflectivity, green peak position, green peak reflectivity, red valley position, and red valley reflectivity; the spectral three-edge characteristic parameters include red edge position, red edge amplitude, red edge area, blue edge position, blue edge amplitude, and blue edge area.

3. The method for dividing cabbage head section type regions guided by global distribution knowledge according to claim 2 is characterized in that: The hyperspectral characteristic indexes include: pigment ratio index, Carter index, normalized pigment difference index, vegetation aging reflectance index, structure-insensitive pigment index, and overgreen vegetation index; The calculation method is as follows: Pigment ratio index: ; Carter Index: ; Normalized pigment difference index: ; Vegetation senescence reflectance index: ; Structure-insensitive pigment index: ; Green vegetation index: ; in, Indicates wavelength The corresponding spectral reflectance.

4. The method for partitioning cabbage head section types based on global distribution knowledge according to claim 1, characterized in that: The cabbage head section type area division module sequentially includes a first downsampling module, a second downsampling module, a third downsampling module, a fourth downsampling module, a fifth downsampling module, a first upsampling splicing module, a second upsampling splicing module, a third upsampling splicing module, a fourth upsampling splicing module, and a feature encoding module of category comparison constraints; The first down-sampling module is jump-connected with the fourth up-sampling splicing module through a global distribution knowledge encoding module, the second down-sampling module is jump-connected with the third up-sampling splicing module through a global distribution knowledge encoding module, the third down-sampling module is jump-connected with the second up-sampling splicing module through a global distribution knowledge encoding module, and the fourth down-sampling module is jump-connected with the first up-sampling splicing module through a global distribution knowledge encoding module.

5. The method for partitioning cabbage head section types based on global distribution knowledge according to claim 1, characterized in that: The output features of the first downsampling module of the cabbage head section type area division module are input to the second downsampling module through maximum pooling, the output features of the second downsampling module are input to the third downsampling module through maximum pooling, the output features of the third downsampling module are input to the fourth downsampling module through maximum pooling, and the output features of the fourth downsampling module are input to the fifth downsampling module through maximum pooling; The output features of the fourth down-sampling module are spliced ​​with the output features of the fifth down-sampling module through the global distribution knowledge encoding module and then used as the input of the first up-sampling splicing module. The output features of the third downsampling module are spliced ​​with the output features of the first upsampling splicing module through the global distribution knowledge encoding module as the input of the second upsampling splicing module. The output features of the second downsampling module are spliced ​​with the output features of the second upsampling splicing module through the global distribution knowledge encoding module and then used as the input of the third upsampling splicing module. The output features of the first downsampling module are spliced ​​with the output features of the third upsampling splicing module through the global distribution knowledge encoding module as the input of the fourth upsampling splicing module. The fourth upsampling splicing module outputs the feature 1×1 convolution input category contrast constrained feature encoding module.

6. The method for dividing cabbage head section type regions guided by global distribution knowledge according to any one of claims 1 to 5, characterized in that: The global distribution knowledge encoding module processes the feature graph as follows: The input feature map undergoes the first 1×1 convolution, and the output feature map of the first 1×1 convolution is then processed by the softmax function and connected to the output feature map of the first 1×1 convolution. The connected feature map is multiplied by the input feature map through the second 1×1 convolution and serves as the final output of the module.

7. The method for partitioning cabbage head section types based on global distribution knowledge according to claim 1, characterized in that: The feature encoding module of the category contrast constraint is based on the semantic information feature vector of the last layer of the encoder, and calculates the cluster center of each category through the spatial constraint clustering method. During the training process, the spatial constraint clustering is performed every 10 epochs to update the cluster center of each category.

8. The method for partitioning cabbage head section types based on global distribution knowledge according to claim 7, characterized in that: The joint loss function includes the loss functions of cross entropy loss and category contrast loss, which are as follows: ， ， ， in, represents the cross entropy loss, represents the class contrast loss, represents the predicted value of the network model, represents the true value of the corresponding point, Represents the semantic information features extracted by the last layer of the network, Indicates that it is constantly updated The average feature vector over the categories, represents the number of categories, is the function representative, Indicates distance, using Euclidean distance, Indicates the similarity between feature vectors.

9. A cabbage head section type area division device guided by global distributed knowledge, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the cabbage head section type area division method guided by global distributed knowledge as described in any one of claims 1-8 when running the program instructions.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method for dividing cabbage head section type regions guided by global distributed knowledge as described in any one of claims 1 to 8 above is implemented.

Images