An intelligent agricultural pest and disease identification system based on deep learning

Abstract

The application discloses an agricultural disease and pest intelligent identification system based on deep learning, which comprises the following steps: performing scale unification and gray scale standardization processing on farmland scene images, extracting maximum stable extreme value regions with area changes meeting stability constraints in a multi-threshold gray scale space, constructing a value region hierarchical structure and calculating a region stability index; extracting hierarchical stability features, area evolution features and boundary geometric distribution features based on candidate regions to form extreme value region structure description vectors; inputting the extreme value region structure description vectors and corresponding region stability indexes into a Kolmogorov-Arnold network for function level mapping to generate function level feature representations; completing disease class or pest class discrimination according to the function level feature representations, and performing spatial correlation with corresponding candidate regions to output agricultural disease and pest identification results. The application belongs to the technical field of computer vision.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an intelligent identification system for agricultural pests and diseases based on deep learning. Background Technology

[0002] With the development of computer vision and deep learning technologies, image-based automatic pest and disease identification methods are gradually being applied in the agricultural field. Existing technologies typically employ convolutional neural networks to extract features and classify the entire image, obtaining high-dimensional feature representations through pixel-level convolution operations. However, these methods generally rely on large-scale labeled samples for training, requiring significant computing resources. In high-resolution farmland image scenarios, intensive computation on the entire image is necessary, resulting in low processing efficiency. Pixel-level feature modeling struggles to directly utilize regional hierarchical structural information, and its ability to express subtle texture differences in early lesions and local morphological changes in insects is limited.

[0003] Some existing methods introduce region detection algorithms to filter candidate regions, but they typically only use the region detection results as a cropping tool or a simple feature source, without coupling the stability and structural information of the regions with the internal expression mechanism of the deep model, resulting in the underutilization of region structural information. In addition, traditional neural network structures use a fixed-weight linear combination method for feature aggregation, lacking a function-level modulation mechanism for the stability of regions, making it difficult to adjust the model response intensity according to the stability differences of different regions.

[0004] Therefore, the existing technology has the following drawbacks: First, it relies on pixel-level dense feature modeling, which has a large computational cost and is difficult to adapt to the fast processing needs of large-scale farmland images; second, it lacks in-depth modeling of the hierarchical structure information of the maximum stable extreme value region and fails to make full use of the regional stability features; third, the network structure is fixed and lacks function-level expression and structural self-evolution mechanism, which limits its ability to express complex textures and morphological changes.

[0005] Therefore, how to provide an intelligent identification system for agricultural pests and diseases based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent agricultural pest and disease identification system based on deep learning. This invention combines maximum stable extremum region extraction with a Kolmogorov-Arnold network to perform structured modeling and function-level feature representation of lesions and insect bodies in farmland scene images. This enables automatic identification and spatial localization of disease and pest categories. This method reduces reliance on pixel-level intensive computation, improves the processing efficiency of large-scale farmland images, and enhances the ability to identify pests and diseases in complex textures, irregular shapes, and scenes with varying lighting conditions.

[0007] An intelligent identification system for agricultural pests and diseases based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] The image acquisition module acquires images of farmland scenes and generates raw image data;

[0009] The image preprocessing module performs scale unification and grayscale standardization on the original image data to obtain standardized image data;

[0010] The stable extreme value region generation module constructs an extreme value region hierarchy in a multi-threshold grayscale space based on standardized image data, selects the largest stable extreme value region whose area change satisfies stability constraints as a candidate region set, and calculates the region stability index for each candidate region.

[0011] The regional structure encoding module takes the candidate region set as input, extracts the hierarchical stability features, area evolution features and boundary geometric distribution features of each candidate region, and fuses them to form an extreme value regional structure description vector. The extreme value regional structure description vector and the corresponding regional stability index are combined to form the regional structure input data.

[0012] The function-level mapping module inputs the regional structure input data into the Kolmogorov-Arnold network, uses the extreme value regional structure description vector as a function variable, embeds the regional stability index into the function mapping process as a modulation parameter, and outputs the function-level feature representation of the candidate region.

[0013] The function-level discrimination module generates discrimination results for disease and pest categories based on function-level feature representation, spatially correlates the discrimination results with the corresponding candidate regions, and outputs agricultural disease and pest identification results.

[0014] Optionally, the raw image data includes color image data of the farmland scene and acquisition attribute data. The color image data is used to characterize the color distribution and spatial morphology information of crop plants and their surface lesions and insects. The color image data includes pixel-level color value information and spatial location information. The acquisition attribute data establishes a unique association with the corresponding color image data. The acquisition attribute data includes image acquisition time information, acquisition plot identification information, acquisition device identification information, and imaging parameter information. The imaging parameter information includes exposure parameters, focal length parameters, and image resolution parameters.

[0015] Optionally, the image preprocessing module includes: reading the original width and original height of the color image data, determining the preset target width and preset target height, scaling the color image data proportionally while maintaining the original aspect ratio to obtain a scaled image, and performing filling processing on the edge areas of the scaled image when the scaled image does not cover the target width and target height to obtain a uniform-scale image, performing color space conversion on the uniform-scale image as input to obtain a grayscale image, calculating the minimum and maximum grayscale values ​​of the grayscale image, performing linear interval mapping on the grayscale image based on the minimum and maximum grayscale values ​​to obtain a standardized grayscale image whose grayscale value distribution falls within a preset standard interval, and establishing the same index association between the standardized grayscale image and the uniform-scale image and combining them to form standardized image data.

[0016] Optional, the stable extreme value region generation module includes:

[0017] The multi-threshold grayscale space construction unit is used to generate a grayscale threshold sequence for the standardized grayscale image in the standardized image data according to a preset grayscale step size, and to perform threshold segmentation processing under each grayscale threshold of the grayscale threshold sequence to obtain a binary image. The grayscale threshold sequence and the corresponding binary image together constitute a multi-threshold grayscale space.

[0018] The value region extraction unit is used to extract the connected regions of pixels in each binary image in the multi-threshold grayscale space as value regions, and record the grayscale threshold and region area corresponding to each value region to form a set of value regions with grayscale labels.

[0019] The value region hierarchical structure building unit is used to establish a parent-child relationship between the value region under the low gray threshold and the value region that is completely contained under the high gray threshold according to the spatial inclusion relationship between value regions under different gray thresholds, forming a value region hierarchical structure, and to aggregate the area of ​​each node in the value region hierarchical structure within the continuous gray threshold range to form a region area sequence.

[0020] The maximum stable extreme value region screening unit is used to calculate the change of region area within a continuous gray threshold interval for the region area sequence corresponding to each node in the value region hierarchy. The node whose region area change meets the stability constraint condition is determined as a stable node. In the same parent-child relationship, the stable node with the smallest region area change and the largest region area is selected, and its corresponding value region is determined as the maximum stable extreme value region, forming a candidate region set.

[0021] The regional stability index calculation unit is used to take the set of candidate regions and their corresponding regional area sequences as input, calculate a stability metric value for each candidate region that represents the degree of change in regional area within the corresponding continuous gray-scale threshold range, and determine the stability metric value as the regional stability index, and output it after establishing a correlation with the corresponding candidate region.

[0022] Optionally, the region structure encoding module includes:

[0023] The hierarchical index construction unit is used to take the candidate region set, region area sequence and region stability index as input, determine the node identifier, parent-child relationship and continuous grayscale threshold range of each candidate region based on the value region hierarchical structure, and establish a unique association between the node identifier and the region area sequence and the region stability index to generate candidate region hierarchical index data.

[0024] The hierarchical stable feature extraction unit is used to extract the hierarchical depth and the number of parent and child nodes from the node identifier and parent-child relationship, and extract the threshold interval length and threshold duration from the continuous gray-scale threshold interval to form hierarchical stable features, taking the candidate region hierarchical index data as input.

[0025] The area evolution feature extraction unit is used to extract the starting and ending areas, the positions of the area's positive and negative values, the direction of area change, and the statistical measures of the area change amplitude under adjacent thresholds from the area sequence in the candidate region hierarchical index data, thereby forming area evolution features.

[0026] The boundary geometric distribution feature extraction unit is used to extract the boundary pixel set and generate the boundary point sequence by taking the boundary pixel set of each candidate region in the candidate region set as input. Based on the boundary point sequence, the boundary perimeter, the size of the circumscribed rectangle and the aspect ratio are calculated, and the direction change of the boundary point sequence is statistically analyzed to generate the boundary direction distribution feature, thus forming the boundary geometric distribution feature.

[0027] The structural description vector construction unit is used to concatenate hierarchical stability features, area evolution features, and boundary geometric distribution features according to a preset feature arrangement order to generate an extreme value region structural description vector corresponding to the node identifier.

[0028] The regional structure input data construction unit is used to call the regional stability index corresponding to the extreme value regional structure description vector in the candidate regional level index data with the extreme value regional structure description vector as the primary key index, and combine the extreme value regional structure description vector and the corresponding regional stability index to form the regional structure input data.

[0029] Optional, stability coupling function-level mapping modules include:

[0030] The index alignment unit is used to take the regional structure input data as input, read the extreme value regional structure description vector and regional stability index corresponding to the same candidate region in each regional structure input data, and align them according to the candidate region index to form an aligned input pair;

[0031] The function variable construction unit is used to take the extreme region structure description vector in the aligned input pair as input, write the features of each dimension of the extreme region structure description vector into the function variable position according to the preset feature arrangement order, and generate the function variable input corresponding to the candidate region index;

[0032] The modulation parameter construction unit is used to generate modulation parameters corresponding to candidate region indices by taking the region stability index in the aligned input pair as input, and to establish the same index association between the modulation parameters and the function variable input.

[0033] The modulation embedding unit is used to take the function variable input and modulation parameters as inputs, and perform modulation processing on the function variable input after the function variable input is formed to obtain the modulated function variable input. The modulation processing is achieved by proportional modulation or offset modulation of the features of each dimension of the function variable input, so that the modulated function variable input is jointly determined by the function variable input and the modulation parameters.

[0034] The network input / output unit is used to take the modulated function variable as input to the Kolmogorov-Arnold network to obtain the network output, and establish a correspondence between the network output and the candidate region index to obtain the function-level feature representation of the candidate region.

[0035] Optional, the Kolmogorov-Arnold network includes:

[0036] The input receiving unit is used to take the modulated function variable input as input, read the corresponding modulated function variable input according to the candidate region index, and generate the input variable group corresponding to the candidate region index;

[0037] The one-dimensional variable expansion unit is used to split the input variable group into a one-dimensional variable sequence according to the preset feature arrangement order, while maintaining the correspondence between the one-dimensional variable sequence and the candidate region index.

[0038] The interaction discovery unit is used to calculate the interaction metric value of each pair of one-dimensional variables as input, and generate a variable interaction set based on the interaction metric value and the preset interaction filtering rules. The variable interaction set includes index information of several one-dimensional variable pairs, which is used to indicate the one-dimensional variable pairs that enter the binary join function mapping.

[0039] The one-dimensional connection function mapping unit is used to take a one-dimensional variable sequence as input, apply a one-dimensional connection function mapping to each one-dimensional variable in the one-dimensional variable sequence, and generate a one-dimensional connection function output sequence corresponding to the candidate region index.

[0040] The binary link function mapping unit is used to apply binary link function mapping to the one-dimensional variable pairs indicated by the variable interaction set and the one-dimensional variable sequence as input, generate the binary link function output sequence corresponding to the candidate region index, and align and merge the binary link function output sequence with the one-dimensional link function output sequence to form the extended function output set.

[0041] The function aggregation unit is used to take the output set of the extended function as input, perform summation and function composition operations on the output set of the extended function according to the preset aggregation topology, generate intermediate aggregation representations, and maintain the correspondence between the intermediate aggregation representations and the candidate region indices;

[0042] The outer aggregation function unit is used to take the intermediate aggregation representation as input, apply the outer aggregation function transformation to the intermediate aggregation representation, and generate the network output corresponding to the candidate region index;

[0043] The contribution evaluation unit is used to calculate the contribution index of each connection function channel in the one-dimensional connection function mapping unit and the binary connection function mapping unit to the loss, based on the candidate region index, and to generate channel contribution data, with the network output as input.

[0044] The structure self-evolution unit is used to perform a structure update operation on the connection function channels with channel contribution data as input, and generate an updated set of connection function channels. The structure update operation includes deleting connection function channels with channel contribution data lower than a preset threshold, adding connection function channels to connection function channels with contribution data that meet preset conditions, or performing function refinement replacement on connection function channels. The updated set of connection function channels is then written back to the one-dimensional connection function mapping unit and the binary connection function mapping unit.

[0045] The output alignment unit is used to establish a correspondence between the network output and the candidate region index, generating a function-level feature representation of the candidate region.

[0046] Optional, the function-level discrimination module includes:

[0047] The index alignment receiving unit is used to take the function-level feature representation of the candidate region as input, read the corresponding function-level feature representation according to the candidate region index, and establish the same candidate region index alignment relationship for the candidate region set.

[0048] The discriminative input building unit is used to take the function-level feature representation as input, perform dimension-consistency processing on the function-level feature representation to obtain the discriminative feature vector, and establish a binding relationship between the discriminative feature vector and the candidate region index.

[0049] The disease category discrimination unit is used to take the discrimination feature vector as input, calculate the category of the discrimination feature vector according to the preset disease category mapping relationship, generate the disease category discrimination result corresponding to the candidate region, and establish a correlation between the disease category discrimination result and the candidate region index;

[0050] The pest category discrimination unit is used to take the discrimination feature vector as input, calculate the category of the discrimination feature vector according to the preset pest category mapping relationship, generate the pest category discrimination result corresponding to the candidate region, and establish a correlation between the pest category discrimination result and the candidate region index;

[0051] The spatial information reading unit is used to extract the pixel set and bounding rectangle position information of the corresponding candidate region according to the candidate region index, taking the candidate region set as input, and obtain the spatial information of the candidate region corresponding to the candidate region index.

[0052] The spatial association output unit is used to combine the disease category discrimination results and pest category discrimination results with the corresponding candidate region spatial information, using the candidate region index as the primary key, to generate regional-level identification results, and to aggregate all regional-level identification results to form agricultural disease and pest identification results.

[0053] The beneficial effects of this invention are:

[0054] This invention combines the maximum stable extreme value region extraction mechanism with the function-level modeling method of the Kolmogorov-Arnold network to construct a unified recognition framework of "stable region extraction - structural feature encoding - function-level mapping - category discrimination". This framework enables the region-level stability information, area evolution information and boundary geometric information to directly participate in the function expression process, realizing the mapping from the region structure space to the function expression space. This improves the ability to express the subtle texture differences of early lesions and the local morphological changes of the insect body, and enhances the recognition stability under complex lighting and background interference conditions.

[0055] This invention achieves rapid screening of candidate regions and suppression of background noise in large-scale farmland images by extracting the maximum stable extreme value region in a multi-threshold grayscale space and constructing a value region hierarchy. This avoids intensive pixel-level calculations on the entire image, reduces computational resource consumption, and improves the system's operating efficiency and coverage in edge devices and large-scale farmland inspection scenarios.

[0056] This invention introduces a variable interaction discovery mechanism and a structure self-evolution mechanism during the function-level mapping process, enabling the Kolmogorov-Arnold network to jointly model key variable pairs and dynamically optimize the connection function structure during training. This enhances the modeling ability for complex and irregular forms, while reducing redundant function channels and improving model expression accuracy and generalization performance, thereby achieving efficient, accurate, and automated identification of agricultural pests and diseases. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of an intelligent agricultural pest and disease identification system based on deep learning proposed in this invention;

[0059] Figure 2 This is a schematic diagram of the Kolmogorov-Arnold network structure in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figures 1-2 A deep learning-based intelligent identification system for agricultural pests and diseases includes the following steps:

[0062] The image acquisition module acquires images of farmland scenes and generates raw image data;

[0063] The image preprocessing module performs scale unification and grayscale standardization on the original image data to obtain standardized image data;

[0064] The stable extreme value region generation module constructs an extreme value region hierarchy in a multi-threshold grayscale space based on standardized image data, selects the largest stable extreme value region whose area change satisfies stability constraints as a candidate region set, and calculates the region stability index for each candidate region.

[0065] The regional structure encoding module takes the candidate region set as input, extracts the hierarchical stability features, area evolution features and boundary geometric distribution features of each candidate region, and fuses them to form an extreme value regional structure description vector. The extreme value regional structure description vector and the corresponding regional stability index are combined to form the regional structure input data.

[0066] The function-level mapping module inputs the regional structure input data into the Kolmogorov-Arnold network, uses the extreme value regional structure description vector as a function variable, embeds the regional stability index into the function mapping process as a modulation parameter, and outputs the function-level feature representation of the candidate region.

[0067] The function-level discrimination module generates discrimination results for disease and pest categories based on function-level feature representation, spatially correlates the discrimination results with the corresponding candidate regions, and outputs agricultural disease and pest identification results.

[0068] In this embodiment, the original image data includes color image data of the farmland scene and acquisition attribute data. The color image data is used to characterize the color distribution and spatial morphology information of crop plants and their surface lesions and insects. The color image data includes pixel-level color value information and spatial location information. The acquisition attribute data establishes a unique association with the corresponding color image data. The acquisition attribute data includes image acquisition time information, acquisition plot identification information, acquisition device identification information, and imaging parameter information. The imaging parameter information includes exposure parameters, focal length parameters, and image resolution parameters. When the same farmland area is acquired multiple times at different times or under different viewing conditions, each acquisition forms an independent image data unit. Each image data unit contains the corresponding color image data and its acquisition attribute data, and they are organized in the order of acquisition time to form an original image data set.

[0069] In this embodiment, the image preprocessing module includes: reading the original width and original height of the color image data, determining the preset target width and preset target height, scaling the color image data proportionally while maintaining the original width-to-height ratio to obtain a scaled image, and performing filling processing on the edge areas of the scaled image when the scaled image does not cover the target width and target height to obtain a uniform scale image, performing color space conversion with the uniform scale image as input to obtain a grayscale image, counting the minimum and maximum grayscale values ​​of the grayscale image, performing linear interval mapping on the grayscale image based on the minimum and maximum grayscale values ​​to obtain a standardized grayscale image whose grayscale value distribution falls within a preset standard interval, and establishing the same index association between the standardized grayscale image and the uniform scale image and combining them to form standardized image data;

[0070] The preset target width and preset target height are uniform spatial size parameters determined before system deployment based on the minimum detection scale of the stable extreme value region and the maximum input resolution of the farmland image. They are used to limit the output resolution of the image after scale unification processing.

[0071] Maintaining the original aspect ratio means that during the scaling process, the color image is resized using the same scaling ratio in both the horizontal and vertical directions, so that the proportional relationship of each pixel in the scaled image on the horizontal and vertical axes remains consistent with the original image, thus keeping the spatial structure of the image unchanged.

[0072] Proportional scaling refers to scaling an image as a whole by applying the same scaling factor to both its width and height during the image resizing process, so that the aspect ratio of the scaled image remains the same as that of the original image.

[0073] Edge region filling refers to the process of adding preset pixel values ​​to the areas around the image that are not covered after the image has been scaled up proportionally, when the image size does not reach the preset target size, so that the overall size of the output image meets the uniform scale requirements.

[0074] Performing color space conversion to obtain a grayscale image means converting multiple color channels of a color image into single-channel brightness data according to a preset channel combination rule, so that each pixel is represented by a grayscale value, thereby forming a grayscale image for grayscale domain processing;

[0075] Linear range mapping refers to adjusting the gray values ​​of each pixel linearly according to the minimum and maximum values ​​of the current gray values, so that the adjusted gray values ​​fall within a preset gray range.

[0076] The preset standard range refers to the range of gray values ​​that the system predetermines during the image preprocessing stage. It is used to unify the gray distribution scale of different images so that the images after gray-scale standardization have a consistent gray-scale representation range.

[0077] In this embodiment, the stable extreme value region generation module includes:

[0078] The multi-threshold grayscale space construction unit is used to generate a grayscale threshold sequence for the standardized grayscale image in the standardized image data according to a preset grayscale step size, and to perform threshold segmentation processing under each grayscale threshold of the grayscale threshold sequence to obtain a binary image. The grayscale threshold sequence and the corresponding binary image together constitute a multi-threshold grayscale space.

[0079] The preset grayscale step size refers to the fixed difference between two adjacent grayscale thresholds when generating a grayscale threshold sequence. It is used to determine the interval at which the grayscale thresholds increase, thereby controlling the fineness of threshold division in the multi-threshold grayscale space.

[0080] Threshold segmentation refers to distinguishing pixels in a grayscale image whose grayscale values ​​meet the threshold condition from those that do not, and generating a corresponding binary image based on the distinction result.

[0081] Multi-threshold grayscale space refers to a grayscale analysis space composed of a set of binary images corresponding to multiple different grayscale thresholds, used to characterize the regional changes of the same grayscale image under different grayscale threshold conditions;

[0082] The value region extraction unit is used to extract the connected regions of pixels in each binary image in the multi-threshold grayscale space as value regions, and record the grayscale threshold and region area corresponding to each value region to form a set of value regions with grayscale labels.

[0083] Extracting connected regions of pixels refers to dividing a set of spatially adjacent pixels with the same pixel value into the same connected region in a binary image according to a preset adjacency rule.

[0084] The value region hierarchical structure building unit is used to establish a parent-child relationship between the value region under the low gray threshold and the value region that is completely contained under the high gray threshold according to the spatial inclusion relationship between value regions under different gray thresholds, forming a value region hierarchical structure, and to aggregate the area of ​​each node in the value region hierarchical structure within the continuous gray threshold range to form a region area sequence.

[0085] The spatial inclusion relationship between value regions under different grayscale thresholds refers to the relationship in which the set of pixels in the value region corresponding to a certain threshold is completely covered by the set of pixels in the value region corresponding to another threshold in a grayscale threshold sequence.

[0086] The parent-child relationship refers to the hierarchical structure of value regions, where a value region that contains other value regions is defined as a parent node, and the value regions contained by it are defined as child nodes, thereby establishing a hierarchical correspondence.

[0087] A continuous grayscale threshold range refers to the range of grayscale thresholds in a grayscale threshold sequence, which consists of several adjacent grayscale thresholds. It is used to describe the continuous existence of the same value region under multiple adjacent grayscale threshold conditions.

[0088] The maximum stable extreme value region screening unit is used to calculate the change of region area within a continuous gray threshold interval for the region area sequence corresponding to each node in the value region hierarchy. The node whose region area change meets the stability constraint condition is determined as a stable node. In the same parent-child relationship, the stable node with the smallest region area change and the largest region area is selected, and its corresponding value region is determined as the maximum stable extreme value region, forming a candidate region set.

[0089] Stability constraints refer to the rules for determining the degree of change in the area of ​​a region within a continuous grayscale threshold range. When the change in the area does not exceed a preset range, the region is deemed to meet the stability requirements.

[0090] The stable node with the smallest change in area and the largest area refers to the node with the smallest change in area when comparing multiple nodes that satisfy the stability constraints within the same parent-child relationship, and the node with the largest area when the change in area is the same.

[0091] The regional stability index calculation unit is used to take the set of candidate regions and their corresponding regional area sequences as input, calculate a stability metric value for each candidate region that represents the degree of change in regional area within the corresponding continuous gray-scale threshold range, and determine the stability metric value as the regional stability index, and output it after establishing a correlation with the corresponding candidate region.

[0092] In this embodiment, the region structure encoding module includes:

[0093] The hierarchical index construction unit is used to take the candidate region set, region area sequence and region stability index as input, determine the node identifier, parent-child relationship and continuous grayscale threshold range of each candidate region based on the value region hierarchical structure, and establish a unique association between the node identifier and the region area sequence and the region stability index to generate candidate region hierarchical index data.

[0094] The node identifier is used to uniquely identify a region node in the value region hierarchy. The parent-child relationship is used to represent the hierarchical relationship between region nodes. The continuous grayscale threshold range is used to represent the threshold range in which the value region corresponding to the node exists continuously in the grayscale threshold sequence.

[0095] The hierarchical stable feature extraction unit is used to extract the hierarchical depth and the number of parent and child nodes from the node identifier and parent-child relationship, and extract the threshold interval length and threshold duration from the continuous gray-scale threshold interval to form hierarchical stable features, taking the candidate region hierarchical index data as input.

[0096] Hierarchical depth refers to the hierarchical position of a value region node relative to the root node in the value region hierarchy, and the number of parent-child nodes refers to the number of direct superior and direct subordinate nodes that have established a parent-child relationship with this node.

[0097] Threshold interval length refers to the size of the grayscale threshold range covered by the continuous grayscale threshold interval, and threshold duration refers to the number of consecutive thresholds in which the candidate region appears and remains connected in the grayscale threshold sequence;

[0098] The area evolution feature extraction unit is used to extract the starting and ending areas, the positions of the area's positive and negative values, the direction of area change, and the statistical measures of the area change amplitude under adjacent thresholds from the area sequence in the candidate region hierarchical index data, thereby forming area evolution features.

[0099] The starting area and ending area are the area values ​​corresponding to the beginning and end of the area sequence. The maximum and minimum values ​​of the area are the maximum or minimum values ​​of the area sequence and their corresponding threshold numbers. The direction of area change is the overall trend of area increase or decrease as the threshold increases. The statistical measure of the area change amplitude under adjacent thresholds is the statistical value of the difference between adjacent areas.

[0100] The boundary geometric distribution feature extraction unit is used to extract the boundary pixel set and generate the boundary point sequence by taking the boundary pixel set of each candidate region in the candidate region set as input. Based on the boundary point sequence, the boundary perimeter, the size of the circumscribed rectangle and the aspect ratio are calculated, and the direction change of the boundary point sequence is statistically analyzed to generate the boundary direction distribution feature, thus forming the boundary geometric distribution feature.

[0101] The direction change of the boundary point sequence refers to calculating the direction angle of the line connecting adjacent boundary points sequentially along the candidate region boundary point sequence, and statistically analyzing the changes in the direction angle in the sequence, which is used to characterize the turning point and curvature distribution characteristics of the boundary morphology;

[0102] The structural description vector construction unit is used to concatenate hierarchical stability features, area evolution features, and boundary geometric distribution features according to a preset feature arrangement order to generate an extreme value region structural description vector corresponding to the node identifier.

[0103] The preset feature arrangement order refers to the rule for arranging hierarchical stable features, area evolution features, and boundary geometric distribution features in a predetermined fixed positional order when constructing the structural description vector of the extreme value region;

[0104] The regional structure input data construction unit is used to call the regional stability index corresponding to the extreme value regional structure description vector in the candidate regional level index data with the extreme value regional structure description vector as the primary key index, and combine the extreme value regional structure description vector and the corresponding regional stability index to form the regional structure input data.

[0105] In this embodiment, the stability coupling function level mapping module includes:

[0106] The index alignment unit is used to take the regional structure input data as input, read the extreme value regional structure description vector and regional stability index corresponding to the same candidate region in each regional structure input data, and align them according to the candidate region index to form an aligned input pair;

[0107] The function variable construction unit is used to take the extreme region structure description vector in the aligned input pair as input, write the features of each dimension of the extreme region structure description vector into the function variable position according to the preset feature arrangement order, and generate the function variable input corresponding to the candidate region index;

[0108] Writing each feature into the function variable position means that, according to the preset feature arrangement order, each feature value in the extreme value region structure description vector is sequentially mapped to a fixed position of the input variable of the Kolmogorov-Arnold network, maintaining a one-to-one correspondence between the feature and the variable position;

[0109] The modulation parameter construction unit is used to generate modulation parameters corresponding to candidate region indices by taking the region stability index in the aligned input pair as input, and to establish the same index association between the modulation parameters and the function variable input.

[0110] The modulation embedding unit is used to take the function variable input and modulation parameters as inputs, and perform modulation processing on the function variable input after the function variable input is formed to obtain the modulated function variable input. The modulation processing is achieved by proportional modulation or offset modulation of the features of each dimension of the function variable input, so that the modulated function variable input is jointly determined by the function variable input and the modulation parameters.

[0111] The network input / output unit is used to take the modulated function variable as input to the Kolmogorov-Arnold network to obtain the network output, and establish a correspondence between the network output and the candidate region index to obtain the function-level feature representation of the candidate region.

[0112] In this embodiment, the Kolmogorov-Arnold network includes:

[0113] The input receiving unit is used to take the modulated function variable input as input, read the corresponding modulated function variable input according to the candidate region index, and generate the input variable group corresponding to the candidate region index;

[0114] The one-dimensional variable expansion unit is used to split the input variable group into a one-dimensional variable sequence according to the preset feature arrangement order, while maintaining the correspondence between the one-dimensional variable sequence and the candidate region index.

[0115] The interaction discovery unit is used to calculate the interaction metric value of each pair of one-dimensional variables as input, and generate a variable interaction set based on the interaction metric value and the preset interaction filtering rules. The variable interaction set includes index information of several one-dimensional variable pairs, which is used to indicate the one-dimensional variable pairs that enter the binary join function mapping.

[0116] Calculating the interaction metric of pairwise combinations of one-dimensional variables refers to statistically analyzing the joint change relationship of any two one-dimensional variables in the same training batch to obtain a value that reflects the degree of correlation or the strength of the cooperative change between the two variables, which is used to determine whether to establish a binary connection function mapping.

[0117] Preset interaction filtering rules refer to the judgment rules that retain one-dimensional variable pairs that meet the filtering conditions by comparing interaction metrics with preset thresholds and combining the frequency or stability requirements of variable occurrence.

[0118] The one-dimensional connection function mapping unit is used to take a one-dimensional variable sequence as input, apply a one-dimensional connection function mapping to each one-dimensional variable in the one-dimensional variable sequence, and generate a one-dimensional connection function output sequence corresponding to the candidate region index.

[0119] Applying a one-dimensional connection function mapping refers to inputting a single one-dimensional variable into the corresponding one-dimensional connection function, and obtaining the nonlinear transformation result of the variable through function operation;

[0120] The binary link function mapping unit is used to apply binary link function mapping to the one-dimensional variable pairs indicated by the variable interaction set and the one-dimensional variable sequence as input, generate the binary link function output sequence corresponding to the candidate region index, and align and merge the binary link function output sequence with the one-dimensional link function output sequence to form the extended function output set.

[0121] Binary link function mapping refers to taking a pair of one-dimensional variables as joint input, performing function operations through the corresponding binary link function, and obtaining a function output that represents the joint change relationship between the two variables.

[0122] Alignment merging refers to arranging the output sequences of one-dimensional join functions and binary join functions according to their corresponding relationships based on the candidate region indices, and combining them to form a unified set of extended function outputs;

[0123] The function aggregation unit is used to take the output set of the extended function as input, perform summation and function composition operations on the output set of the extended function according to the preset aggregation topology, generate intermediate aggregation representations, and maintain the correspondence between the intermediate aggregation representations and the candidate region indices;

[0124] Preset aggregation topology refers to the pre-determined combination relationship and addition order between one-dimensional and binary connection function outputs during function aggregation, which is used to limit the structural form of function outputs participating in aggregation operations;

[0125] Performing summation and function composition operations refers to summing the outputs of multiple connection functions according to a preset aggregation topology, and then inputting the summation result into the outer function for further function operations to form an intermediate representation.

[0126] The outer aggregation function unit is used to take the intermediate aggregation representation as input, apply the outer aggregation function transformation to the intermediate aggregation representation, and generate the network output corresponding to the candidate region index;

[0127] The outer aggregation function transformation refers to applying a pre-defined nonlinear function operation to the intermediate aggregation representation, converting the combination result into the output value of the network at the current stage;

[0128] The contribution evaluation unit is used to calculate the contribution index of each connection function channel in the one-dimensional connection function mapping unit and the binary connection function mapping unit to the loss, based on the candidate region index, and to generate channel contribution data, with the network output as input.

[0129] The contribution index of each connection function channel to the loss refers to the degree of influence of the gradient of the corresponding parameter or output of each one-dimensional or binary connection function channel on the overall loss change during backpropagation, and obtain the channel contribution value accordingly.

[0130] The structure self-evolution unit is used to perform a structure update operation on the connection function channels with channel contribution data as input, and generate an updated set of connection function channels. The structure update operation includes deleting connection function channels with channel contribution data lower than a preset threshold, adding connection function channels to connection function channels with contribution data that meet preset conditions, or performing function refinement replacement on connection function channels. The updated set of connection function channels is then written back to the one-dimensional connection function mapping unit and the binary connection function mapping unit.

[0131] The output alignment unit is used to establish a correspondence between the network output and the candidate region index, generating a function-level feature representation of the candidate region.

[0132] In this embodiment, the function-level discrimination module includes:

[0133] The index alignment receiving unit is used to take the function-level feature representation of the candidate region as input, read the corresponding function-level feature representation according to the candidate region index, and establish the same candidate region index alignment relationship for the candidate region set.

[0134] The discriminative input building unit is used to take the function-level feature representation as input, perform dimension-consistency processing on the function-level feature representation to obtain the discriminative feature vector, and establish a binding relationship between the discriminative feature vector and the candidate region index.

[0135] Performing dimension consistency processing refers to adjusting the dimensions of function-level feature representations to ensure that the feature representations corresponding to different candidate regions are consistent in terms of feature dimension and arrangement order.

[0136] The disease category discrimination unit is used to take the discrimination feature vector as input, calculate the category of the discrimination feature vector according to the preset disease category mapping relationship, generate the disease category discrimination result corresponding to the candidate region, and establish a correlation between the disease category discrimination result and the candidate region index;

[0137] Pre-defined disease category mapping relationship refers to the correspondence rules or mapping structure between features and disease categories established in advance during the category discrimination stage, which is used to map the discrimination feature vector to a specific disease category identifier;

[0138] Category calculation refers to inputting the discriminative feature vector into the corresponding category mapping relationship to obtain the score or confidence level of each category, and determining the output category label based on the score or confidence level;

[0139] The pest category discrimination unit is used to take the discrimination feature vector as input, calculate the category of the discrimination feature vector according to the preset pest category mapping relationship, generate the pest category discrimination result corresponding to the candidate region, and establish a correlation between the pest category discrimination result and the candidate region index;

[0140] Pre-defined pest category mapping relationship refers to the corresponding rules or mapping structure of features to pest categories established in advance during the category discrimination stage, which is used to map the discrimination feature vector to a specific pest category identifier;

[0141] The spatial information reading unit is used to extract the pixel set and bounding rectangle position information of the corresponding candidate region according to the candidate region index, taking the candidate region set as input, and obtain the spatial information of the candidate region corresponding to the candidate region index.

[0142] The pixel set refers to the set of all pixel coordinate points that constitute the candidate region, and the bounding rectangle position information refers to the position information of the smallest rectangle that can cover the pixel set, which is used to represent the spatial range of the candidate region in the image.

[0143] The spatial association output unit is used to combine the disease category discrimination results and pest category discrimination results with the corresponding candidate region spatial information, using the candidate region index as the primary key, to generate regional-level identification results, and to aggregate all regional-level identification results to form agricultural disease and pest identification results.

[0144] Example 1:

[0145] To verify the feasibility of this invention in practice, it was applied to large-scale farmland inspection. Crop diseases and pests are typically characterized by dispersed locations, diverse morphological changes, and weak early-stage features. Taking leafy crops as an example, early lesion diameters are usually between 2 and 6 millimeters, and the body length of insect larvae is mostly concentrated around 3 millimeters, occupying only a tiny proportion of pixels in high-resolution farmland images. Traditional manual inspection methods have limited coverage of farmland area per inspector per day, and the identification results are highly dependent on personal experience, easily leading to missed detections and misjudgments. Existing automatic identification methods based on convolutional features of the entire image have shown some effectiveness in experimental environments, but in real farmland scenarios, where image resolutions generally exceed 4000×3000 pixels, the model needs to perform convolution calculations on more than ten million pixels, resulting in a long average processing time per image. Furthermore, the stability of identifying tiny lesions and insects is insufficient under complex background conditions.

[0146] In this embodiment, the resolution of a single original image is approximately 4200×3200 pixels. The image acquisition module records imaging parameters, including exposure time, focal length, and image resolution, while simultaneously acquiring the image. The image preprocessing module performs scale-uniform processing on the original image, scaling it to a preset target size range while maintaining the original aspect ratio. After scaling, the longer side of the image is approximately 1600 pixels, and the shorter side is automatically determined based on the aspect ratio. After processing, the total number of pixels in a single image is reduced to approximately 2.5 million pixels, a reduction of about 75% compared to the original image, thus reducing the computational load for subsequent processing.

[0147] After scaling, the system performs grayscale conversion and grayscale standardization on the image. Through linear interval mapping, the grayscale value distribution is uniformly mapped to a preset standard interval, so that the grayscale mean fluctuation range of the image under different acquisition conditions is controlled within 5%, which significantly reduces the impact of illumination changes on subsequent region detection.

[0148] In the stable extremum region generation stage, the system processes the standardized grayscale image in a multi-threshold grayscale space. The grayscale threshold sequence consists of multiple adjacent grayscale thresholds, and the threshold step size is set to a fixed interval. In actual operation, a single image usually generates more than one hundred binary images corresponding to grayscale thresholds. The system extracts pixel connected regions in each binary image and calculates the area of ​​the corresponding region. In the process of constructing the value region hierarchy, the system associates regions that exist continuously at the same spatial location under different grayscale thresholds to form a complete hierarchical path. Statistical analysis shows that in farmland images, the number of initially extracted value regions can reach thousands, but after performing stability constraint screening, the number of maximum stable extremum regions that meet the region area change conditions is controlled on average between 15 and 30, accounting for less than 1% of the original number of value regions.

[0149] Through the aforementioned filtering mechanism, the system compresses the objects for subsequent depth processing from millions of pixels in the entire image to a limited number of candidate regions, with the number of pixels in a single candidate region typically ranging from 200 to 3000. This processing method significantly reduces background noise areas while ensuring that lesions and insect bodies are preserved.

[0150] During the region structure encoding stage, the system extracts hierarchical stability features, area evolution features, and boundary geometric distribution features for each candidate region. Taking hierarchical stability features as an example, the length of the continuous gray-level threshold interval in the value region hierarchy of a candidate region is typically between 8 and 20 thresholds, significantly higher than the 2 to 4 thresholds for background noise regions. In the area evolution features, the area change trend of lesion regions shows a slow change, with an average area change amplitude of less than 10%, while the change amplitude of non-target regions often exceeds 30%. In the boundary geometric distribution features, the boundary perimeter and aspect ratio of the circumscribed rectangle of lesion regions show a stable distribution, while stray noise regions exhibit a highly irregular boundary direction distribution.

[0151] The aforementioned features are concatenated to form extreme value region structure description vectors. The dimension of a single vector is controlled to be on the order of tens of dimensions, which significantly reduces feature redundancy compared to the feature representations in traditional convolutional networks that often have thousands of dimensions. Each extreme value region structure description vector, together with the corresponding region stability index, constitutes the region structure input data, which serves as the input for function-level modeling.

[0152] In the stability coupling function-level mapping stage, the system maps the extreme value region structure description vector to the function variable input, and embeds the region stability index as a modulation parameter into the function mapping process, so that the region stability directly affects the function response strength. The improved Kolmogorov-Arnold network performs variable interaction discovery processing during training. Typically, in the initial training stage, it identifies high-interaction variable combinations that account for about 20% of the total variable pairs and applies binary connection function mapping to them. As the training iteration progresses, the structure self-evolution mechanism gradually deletes connection function channels with low contribution. Finally, the number of effective channels retained is reduced by about 40% compared with the initial state, and the network structure is more compact.

[0153] In the function-level discrimination stage, the system generates disease and pest classification results based on function-level feature representations, and spatially correlates these results with the pixel set or bounding rectangle location information of the candidate region. In continuous test images, the system's average processing time for a single image is significantly lower than traditional whole-image convolutional recognition methods, while its recognition success rate for early lesions and insect samples is significantly improved. In statistical comparisons, the system detects a significantly higher number of minute lesions than the control method, while maintaining a low number of false positives.

[0154] This embodiment clearly demonstrates that the present invention significantly reduces the involvement of irrelevant regions through the maximum stable extreme value region screening mechanism, improves the ability to express fine-grained structural differences through region structural feature encoding and function-level modeling, and achieves efficient and stable pest and disease identification through the improved Kolmogorov-Arnold network, thus solving the key problems of existing technologies in terms of efficiency in processing large-scale farmland images and early pest and disease identification capabilities.

[0155] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent identification system for agricultural pests and diseases, characterized in that, Includes the following modules: The image acquisition module acquires images of farmland scenes and generates raw image data; The image preprocessing module performs scale unification and grayscale standardization on the original image data to obtain standardized image data; The stable extreme value region generation module constructs an extreme value region hierarchy in a multi-threshold grayscale space based on standardized image data, selects the largest stable extreme value region whose area change satisfies stability constraints as a candidate region set, and calculates the region stability index for each candidate region. The regional structure encoding module takes the candidate region set as input, extracts the hierarchical stability features, area evolution features and boundary geometric distribution features of each candidate region, and fuses them to form an extreme value regional structure description vector. The extreme value regional structure description vector and the corresponding regional stability index are combined to form the regional structure input data. The function-level mapping module inputs the regional structure input data into the Kolmogorov-Arnold network, uses the extreme value regional structure description vector as a function variable, embeds the regional stability index into the function mapping process as a modulation parameter, and outputs the function-level feature representation of the candidate region. The function-level discrimination module generates discrimination results for disease and pest categories based on function-level feature representation, spatially correlates the discrimination results with the corresponding candidate regions, and outputs agricultural disease and pest identification results.

2. The intelligent agricultural pest and disease identification system based on deep learning according to claim 1, characterized in that, The raw image data includes color image data of the farmland scene and acquisition attribute data. The color image data is used to characterize the color distribution and spatial morphology information of crop plants and their surface lesions and insects. The color image data includes pixel-level color value information and spatial location information. The acquisition attribute data establishes a unique association with the corresponding color image data. The acquisition attribute data includes image acquisition time information, acquisition plot identification information, acquisition device identification information, and imaging parameter information. The imaging parameter information includes exposure parameters, focal length parameters, and image resolution parameters.

3. The intelligent agricultural pest and disease identification system based on deep learning according to claim 1, characterized in that, The image preprocessing module includes: reading the original width and height of the color image data, determining the preset target width and height, scaling the color image data proportionally while maintaining the original aspect ratio to obtain a scaled image, and performing filling processing on the edge areas of the scaled image when it does not cover the target width and height to obtain a uniform-scale image. Using the uniform-scale image as input, a color space conversion is performed to obtain a grayscale image. The minimum and maximum grayscale values ​​of the grayscale image are counted. Based on the minimum and maximum grayscale values, a linear interval mapping is performed on the grayscale image to obtain a standardized grayscale image whose grayscale value distribution falls within a preset standard interval. The standardized grayscale image and the uniform-scale image are associated with the same index and combined to form standardized image data.

4. The intelligent agricultural pest and disease identification system based on deep learning according to claim 1, characterized in that, The stable extreme value region generation module includes: The multi-threshold grayscale space construction unit is used to generate a grayscale threshold sequence for the standardized grayscale image in the standardized image data according to a preset grayscale step size, and to perform threshold segmentation processing under each grayscale threshold of the grayscale threshold sequence to obtain a binary image. The grayscale threshold sequence and the corresponding binary image together constitute a multi-threshold grayscale space. The value region extraction unit is used to extract the connected regions of pixels in each binary image in the multi-threshold grayscale space as value regions, and record the grayscale threshold and region area corresponding to each value region to form a set of value regions with grayscale labels. The value region hierarchical structure building unit is used to establish a parent-child relationship between the value region under the low gray threshold and the value region that is completely contained under the high gray threshold according to the spatial inclusion relationship between value regions under different gray thresholds, forming a value region hierarchical structure, and to aggregate the area of ​​each node in the value region hierarchical structure within the continuous gray threshold range to form a region area sequence. The maximum stable extreme value region screening unit is used to calculate the change of region area within a continuous gray threshold interval for the region area sequence corresponding to each node in the value region hierarchy. The node whose region area change meets the stability constraint condition is determined as a stable node. In the same parent-child relationship, the stable node with the smallest region area change and the largest region area is selected, and its corresponding value region is determined as the maximum stable extreme value region, forming a candidate region set. The regional stability index calculation unit is used to take a set of candidate regions and their corresponding regional area sequences as input, calculate a stability metric for each candidate region, determine the stability metric as the regional stability index, and output it after establishing a correlation with the corresponding candidate region.

5. The intelligent agricultural pest and disease identification system based on deep learning according to claim 1, characterized in that, The region structure encoding module includes: The hierarchical index construction unit is used to take the candidate region set, region area sequence and region stability index as input, determine the node identifier, parent-child relationship and continuous grayscale threshold range of each candidate region based on the value region hierarchical structure, and establish a unique association between the node identifier and the region area sequence and the region stability index to generate candidate region hierarchical index data. The hierarchical stable feature extraction unit is used to extract the hierarchical depth and the number of parent and child nodes from the node identifier and parent-child relationship, and extract the threshold interval length and threshold duration from the continuous gray-scale threshold interval to form hierarchical stable features, taking the candidate region hierarchical index data as input. The area evolution feature extraction unit is used to extract the starting and ending areas, the positions of the area's positive and negative values, the direction of area change, and the statistical measures of the area change amplitude under adjacent thresholds from the area sequence in the candidate region hierarchical index data, thereby forming area evolution features. The boundary geometric distribution feature extraction unit is used to extract the boundary pixel set and generate the boundary point sequence by taking the boundary pixel set of each candidate region in the candidate region set as input. Based on the boundary point sequence, the boundary perimeter, the size of the circumscribed rectangle and the aspect ratio are calculated, and the direction change of the boundary point sequence is statistically analyzed to generate the boundary direction distribution feature, thus forming the boundary geometric distribution feature. The structural description vector construction unit is used to concatenate hierarchical stability features, area evolution features, and boundary geometric distribution features according to a preset feature arrangement order to generate an extreme value region structural description vector corresponding to the node identifier. The regional structure input data construction unit is used to call the regional stability index corresponding to the extreme value regional structure description vector in the candidate regional level index data with the extreme value regional structure description vector as the primary key index, and combine the extreme value regional structure description vector and the corresponding regional stability index to form the regional structure input data.

6. The intelligent agricultural pest and disease identification system based on deep learning according to claim 1, characterized in that, The stability coupling function-level mapping module includes: The index alignment unit is used to take the regional structure input data as input, read the extreme value regional structure description vector and regional stability index corresponding to the same candidate region in each regional structure input data, and align them according to the candidate region index to form an aligned input pair; The function variable construction unit is used to take the extreme region structure description vector in the aligned input pair as input, write the features of each dimension of the extreme region structure description vector into the function variable position according to the preset feature arrangement order, and generate the function variable input corresponding to the candidate region index; The modulation parameter construction unit is used to generate modulation parameters corresponding to candidate region indices by taking the region stability index in the aligned input pair as input, and to establish the same index association between the modulation parameters and the function variable input. The modulation embedding unit is used to perform modulation processing on the function variable input after the function variable input is formed, taking the function variable input and modulation parameters as input, to obtain the modulated function variable input. The network input / output unit is used to take the modulated function variable as input to the Kolmogorov-Arnold network to obtain the network output, and establish a correspondence between the network output and the candidate region index to obtain the function-level feature representation of the candidate region.

7. The intelligent agricultural pest and disease identification system based on deep learning according to claim 6, characterized in that, The Kolmogorov-Arnold network includes: The input receiving unit is used to take the modulated function variable input as input, read the corresponding modulated function variable input according to the candidate region index, and generate the input variable group corresponding to the candidate region index; A one-dimensional variable expansion unit is used to split an input variable set into a one-dimensional variable sequence according to a preset feature arrangement order, taking the input variable set as input. The interaction discovery unit is used to take a sequence of one-dimensional variables as input, calculate the interaction metric value of each pair of one-dimensional variables, and generate a variable interaction set based on the interaction metric value and a preset interaction filtering rule. The variable interaction set includes index information of several one-dimensional variable pairs. The one-dimensional connection function mapping unit is used to take a one-dimensional variable sequence as input, apply a one-dimensional connection function mapping to each one-dimensional variable in the one-dimensional variable sequence, and generate a one-dimensional connection function output sequence corresponding to the candidate region index. The binary link function mapping unit is used to apply binary link function mapping to the one-dimensional variable pairs indicated by the variable interaction set and the one-dimensional variable sequence as input, generate the binary link function output sequence corresponding to the candidate region index, and align and merge the binary link function output sequence with the one-dimensional link function output sequence to form the extended function output set. The function aggregation unit is used to take the output set of the extended function as input, perform summation and function composition operations on the output set of the extended function according to a preset aggregation topology, and generate an intermediate aggregation representation. The outer aggregation function unit is used to take the intermediate aggregation representation as input, apply the outer aggregation function transformation to the intermediate aggregation representation, and generate the network output corresponding to the candidate region index; The contribution evaluation unit is used to calculate the contribution index of each connection function channel in the one-dimensional connection function mapping unit and the binary connection function mapping unit to the loss, based on the candidate region index, and to generate channel contribution data, with the network output as input. The self-evolutionary structural unit is used to perform structural update operations on the connection function channels with channel contribution data as input, generate an updated set of connection function channels, and write back the updated set of connection function channels to the one-dimensional connection function mapping unit and the binary connection function mapping unit. The output alignment unit is used to establish a correspondence between the network output and the candidate region index, generating a function-level feature representation of the candidate region.

8. The intelligent agricultural pest and disease identification system based on deep learning according to claim 1, characterized in that, The function-level discrimination module includes: The index alignment receiving unit is used to take the function-level feature representation of the candidate region as input, read the corresponding function-level feature representation according to the candidate region index, and establish the same candidate region index alignment relationship for the candidate region set. The discriminative input building unit is used to take the function-level feature representation as input, perform dimension-consistency processing on the function-level feature representation to obtain the discriminative feature vector, and establish a binding relationship between the discriminative feature vector and the candidate region index. The disease category discrimination unit is used to take the discrimination feature vector as input, calculate the category of the discrimination feature vector according to the preset disease category mapping relationship, generate the disease category discrimination result corresponding to the candidate region, and establish a correlation between the disease category discrimination result and the candidate region index; The pest category discrimination unit is used to take the discrimination feature vector as input, calculate the category of the discrimination feature vector according to the preset pest category mapping relationship, generate the pest category discrimination result corresponding to the candidate region, and establish a correlation between the pest category discrimination result and the candidate region index; The spatial information reading unit is used to extract the pixel set and bounding rectangle position information of the corresponding candidate region according to the candidate region index, taking the candidate region set as input, and obtain the spatial information of the candidate region corresponding to the candidate region index. The spatial association output unit is used to combine the disease category discrimination results and pest category discrimination results with the corresponding candidate region spatial information, using the candidate region index as the primary key, to generate regional-level identification results, and to aggregate all regional-level identification results to form agricultural disease and pest identification results.

Images