Camellia oleifera disease and insect pest recognition system based on image recognition technology

Abstract

The invention relates to the technical field of image analysis, in particular to a camellia oleifera disease and insect pest recognition system based on an image recognition technology, which realizes the structured expression of an image sequence by combining the correlation characteristics of an image time sequence and a scab evolution trajectory, forms a time axis index by using shooting time and cleans redundant images. The method effectively improves the continuity and reliability of input data, comprehensively judges the morphological dynamic state of disease spots and accurately depicts the evolutionary states of disease spot expansion, color change and the like through the comparison mode of edge texture change and main color center offset paths, remarkably enhances the staged judgment capability of disease and pest development, improves the image semantic understanding capability, and improves the image quality. Through a multi-dimension matching strategy of a closed region, a color shift direction, area jump and the like, a calibration coding sequence with traceability and staged grading capability is constructed, and the discrimination sensitivity of the system to the early stage, the middle stage and the diffusion stage of diseases is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image analysis, in particular to an oil tea disease and pest identification system based on image recognition technology. BACKGROUND

[0002] The technical field of image analysis is a core sub-direction of computer vision, mainly studying how to extract useful information from digital images and make pattern recognition and intelligent judgment, often relying on machine learning or deep learning algorithms to realize automatic information recognition and processing. Image analysis is widely used in medical diagnosis, industrial detection, security monitoring, agricultural intelligence and other fields. In particular, in the field of agriculture, image analysis can greatly improve the efficiency and accuracy of disease and pest identification.

[0003] The oil tea disease and pest identification system based on image recognition technology aims to solve the problems of low efficiency and high misjudgment in traditional patrol, realize intelligent perception, timely diagnosis and prevention suggestion pushing of crop health status, and build a sustainable operation, high identification accuracy and fast response speed identification tool to serve the early detection and precise intervention of diseases and pests in the process of oil tea planting, so as to reduce the use of pesticides and protect the yield and quality of economic crops.

[0004] The application of existing technology in agricultural disease and pest identification mainly depends on the static features of single frame image, lacks dynamic modeling of disease spot evolution process, and often cannot be effectively identified. Since the time dimension and sequence information are not introduced, the traditional technology is prone to misidentification or missed identification when the image input is numbered disorderly or the disease spot changes slowly. The method lacks attention to medium scale indicators such as texture details and main color shift, resulting in insufficient distinction between adjacent disease spots. The static class labeling or classifier output of common method lacks explainable coding structure, which is not conducive to tracking the disease evolution path and supporting the subsequent development of precise intervention strategy. When facing complex scenes such as high density disease spots and chronic disease evolution, the misjudgment rate is high, the prevention and control opportunity is delayed, and other chain effects are caused, which affects the efficiency of yield and prevention and control resource allocation. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art, and an oil tea disease and pest identification system based on image recognition technology is proposed.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: the oil tea disease and pest identification system based on image recognition technology comprises:

[0007] An image frame sequence structure construction module: based on the evolution track of oil tea disease spot, the image time sequence is extracted, the sequence sorting index is generated through the shooting time information, the non-continuous data is screened out and the time stamp is merged, and the continuous sequence composition list is generated;

[0008] The disease spot evolution sequence screening module comprises the following steps:

[0009] The multi-dimensional label coding construction module comprises the following steps:

[0010] The label legality combination analysis module comprises the following steps:

[0011] The stage level coding generation module comprises the following steps:

[0012] As a further scheme of the present application, the image frame sequence structure construction module comprises:

[0013] The time index generation submodule comprises the following steps:

[0014] The continuous sequence composition submodule comprises the following steps:

[0015] As a further scheme of the present application, the disease spot evolution sequence screening module comprises:

[0016] Edge difference extraction submodule: Based on the continuous sequence composition list, the image grayscale values ​​are separated and the grayscale change segments of the regional boundaries are obtained. The edge texture features are extracted through a convolutional neural network. The pixel gradient is scanned line by line to extract the outer edge position of the lesion contour and mark the coordinate point set. The difference in the coordinates of the corresponding numbered contour points between frames is calculated and the pixel offset coordinates are screened to obtain the edge change segment and mark the pixel change set, thereby generating the edge pixel difference set.

[0017] Area change calculation submodule: Based on the edge pixel difference set, the closed boundary of the adjacent pixel connection area is tracked, the area value of the lesion in the area is calculated, the area change amplitude between consecutive frames is extracted by frame number mapping and comparison, and a change trend list is constructed using the inter-frame difference sequence, the direction of area increase and decrease is marked, and a dynamic area feature sequence is generated;

[0018] Morphological combination construction submodule: Based on the area dynamic feature sequence, the pixel range of the main color area of ​​the lesion in each frame is obtained and the position of the peak number of frequency pixel blocks is extracted, the pixel center coordinates are located, the center point of the main color position is obtained, the Euclidean distance of the coordinate values ​​between frames is calculated and an offset path is generated, the area change value and the center path pair are numbered, and then merged into a sequence of structure segment pairs to generate a morphological evolution change combination group.

[0019] As a further solution of the present invention, the convolutional neural network is based on the formula:

[0020]

[0021] Where: F′(i,j) represents the edge enhanced convolution response value at position i,j, K(m,n) represents the weight value of the convolution kernel at the mth and nth positions, I(i+m,j+n) represents the pixel grayscale value at position i+m,j+n in the image, G(i+m,j+n) represents the grayscale gradient value at position i+m,j+n in the image, α represents the fusion weight coefficient of the pixel value channel, β represents the fusion weight coefficient of the grayscale gradient channel, i represents the row index of the current output feature map, j represents the column index of the current output feature map, m represents the index position along the row direction in the convolution kernel, and n represents the index position along the column direction in the convolution kernel.

[0022] As a further solution of the present invention, the separated image grayscale value refers to converting the original color image into a single-channel grayscale image, specifically reading the brightness values ​​of the red, green and blue channels of the pixels in each frame of the oil tea leaf image in turn, generating the corresponding grayscale value by weighted superposition, and replacing the original multi-channel pixel information to form a unified grayscale image structure.

[0023] As a further solution of the present invention, the multi-dimensional label encoding construction module includes:

[0024] Area ratio calculation submodule: Based on the morphological evolution change combination group, the pixels of the lesion area in each frame of the image are marked and the lesion pixel positions are extracted. The total number of lesion pixels in a single frame is obtained, and the total number of pixels in the entire frame of the image is counted. After performing time-dependent modeling on the lesion area ratio sequence of consecutive frames through a long short-term memory network, the dynamic change trend is captured. The ratio of the lesion area to the total pixels is calculated and the ratio sequence is recorded to generate a frame area ratio sequence.

[0025] Direction vector extraction submodule: Based on the frame area ratio sequence, extract the pixel blocks of the main color area of ​​each frame image and obtain the main color pixel set, calculate the center of gravity of the main color pixel, obtain the center position of the coordinate point of the main color area within the frame, calculate the coordinates of the main color center of gravity point between each frame and extract the direction angle value, divide the direction angle interval, obtain the offset direction segment number and then associate it with the frame index one-to-one to generate a color offset direction sequence;

[0026] Label segment value assembly submodule: Based on the color offset direction sequence, the area ratio segment number, offset direction segment number and position offset quantity segment number in each frame are obtained, and the three segment numbers are combined into a single-frame label coding field through a fixed number splicing order. The label fields of each sequence are spliced ​​and aggregated into an overall label vector set in frame order to generate a label coding synthetic vector set.

[0027] As a further solution of the present invention, the long short-term memory network is according to the formula:

[0028] H s =Z s tanh(U s ·M s-1 +V s tanh(W z ·[H s-1 ,A s ,D s ,Q s ]+B z ))

[0029] Among them: H s Represents the dynamic state output value of the sth frame, which is used to reflect the degree of evolutionary response in the time series. s Indicates the output gate control factor of the sth frame, which controls whether the state is output at the moment, U s Indicates the retention coefficient of the s-th frame for the memory item, M s-1 Represents the memory state value of the s-1th frame, V s represents the activation control coefficient for the current input update in the sth frame, W z Represents the weight matrix of the state update operation, [H s-1 ,A s ,D s ,Q s] represents the input vector set, H s-1 Indicates the dynamic state output value of the s-1 frame, A s It represents the ratio of the lesion area in the sth frame to the total number of pixels in the entire frame image, D s It represents the absolute jump value of the ratio of the lesion area between the sth frame and the s-1th frame, and is calculated as D s =|A s -A s-1 |, Q s represents the compactness index of the lesion area in the sth frame, B z represents the bias term, s represents the frame index, and tanh represents the hyperbolic tangent function.

[0030] As a further solution of the present invention, the marking of diseased spots refers to locating and pixel-level identifying the areas showing abnormal color and texture features in each frame of the image. Specifically, the image is first converted from RGB to a brightness-chrominance separation color model through color space transformation, and then the color range threshold is set according to the common brown-red and gray-brown tones of oil-tea camellia lesions. The pixel set that meets the conditions is extracted, and the highlight and shadow interference areas close to the tones of healthy leaves are excluded. The closed and semi-closed color abnormal connected areas are extracted as preliminary diseased spot candidate areas. The areas are reviewed based on the lower limit of the area, the degree of edge irregularity and the texture roughness, and the pixel set that meets the typical morphology of the lesion is retained as the diseased spot marking area.

[0031] As a further solution of the present invention, the continuous logical label set group includes:

[0032] Segment value pair screening submodule: Based on the label encoding synthetic vector set, the segment value fields of adjacent frame positions in the label field sequence are extracted. The field sliding window is combined frame by frame to form label segment value pairs. The difference between the numbered segment values ​​is calculated to obtain a record difference sequence. The segment value discontinuous field groups are screened. The cumulative number of the same segment value pair in the entire sequence is counted and then compared with the frequency threshold. The segment value pairs with a cumulative number below the set threshold are deleted to generate a high-frequency legal segment value combination set.

[0033] Logical set construction submodule: Based on the high-frequency legal segment value combination set, the front and back association relationship of each segment value combination field is constructed, the segment value numbers of the front and back items in the combination field are connected and a chain mapping table is established, the chain extension method is used to splice the continuously connected segment values ​​backward to construct a label segment sequence, the segment value structure in each chain is called and whether it is closed, the structure group that meets the chain head and chain tail overlap conditions is screened, the chain group number is used to map the corresponding label frame sequence, a full sequence connection table is established, and a continuous logical label set group is generated.

[0034] As a further solution of the present invention, the stage level code generation module includes:

[0035] The label segment group determination submodule extracts the closed edge segment values ​​in the label field and identifies the closed number field based on the continuous logical label set group. The numbered mark area with closed characteristics is extracted through coordinate connectivity screening. The fields in the set are matched. The offset direction segment value is extracted and located in the same frame label field. The area jump segment value and position offset segment value corresponding to the above position are matched frame by frame. The segment value field is combined to establish a three-segment group comparison structure. The field configuration template in the combination rule set is matched. Non-compliant combinations are screened out to generate a label combination segment group list.

[0036] Stage coding generation submodule: Based on the label combination segment group list, split the segment group number field structure and extract the index value in each combination position, map the segment value combination field with the preset coding rule table one by one, locate the mapping coding number, merge the number sequence according to the original frame order, use the sequential sequence structure to establish the mapping index between the frame and the stage number, integrate the mapping relationships and output the continuous segment number results to generate the stage calibration coding sequence.

[0037] Compared with the prior art, the advantages and positive effects of the present invention are:

[0038] 1. This method combines the correlation characteristics of image time sequence and lesion evolution trajectory to achieve structured expression of image sequences. It uses shooting time to form a time axis index and clean up redundant images, effectively improving the continuity and reliability of input data.

[0039] 2. This invention uses a method that compares edge texture changes and the main color center offset path to comprehensively judge the morphological dynamics of the lesions, accurately depicting the evolution of lesions such as expansion and color change, significantly enhancing the ability to judge the stage of pest development and improve image semantic understanding capabilities;

[0040] 3. In this invention, through multi-dimensional matching strategies such as closed areas, color shift direction and area jump, a calibration coding sequence with traceability and stage-by-stage grading capabilities is constructed, which greatly improves the system's sensitivity to the early, middle and spread stages of the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a system flow chart of the present invention;

[0042] Figure 2 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0045] Example

[0046] See also Figure 1 The present invention provides a technical solution: a tea plant disease and insect pest identification system based on image recognition technology includes:

[0047] Image frame sequence structure construction module: Based on the evolution trajectory of oil-tea camellia lesions, the image time sequence is extracted, the sequence sorting index is generated through the shooting time information, the discontinuous data is filtered out and the timestamps are merged to generate a continuous sequence composition list;

[0048] Lesion evolution sequence screening module: Based on a continuous sequence composition list, a convolutional neural network is used to extract the edge texture features of the lesion. The area values ​​of the contour lines of the previous and next frames are calculated to obtain the area change. The distance between the main color center points is compared to obtain the offset path, and a morphological evolution change combination group is generated.

[0049] Multidimensional label coding construction module: Based on the morphological evolution change combination group, the inter-frame evolution characteristics are modeled through the long short-term memory network, the ratio of the pixel area of ​​the lesion area to the total pixels in each frame is calculated, and the segment value numbers corresponding to each parameter are sequentially aggregated and assembled into label codes to generate a label code synthetic vector set;

[0050] Label legitimacy combination analysis module: Based on the label encoding synthetic vector set, the label segment value fields are combined in pairs, and pairs that do not meet the continuous segment value arrangement structure are filtered out. Label combination groups with label encoding combination frequencies above the threshold are retained and logically reconstructed to generate a continuous logical label set group.

[0051] Stage level coding generation module: Based on the continuous logical label set group, it judges the closed edge segment and the color offset direction label segment value, matches the area jump value and the position offset segment value, queries and queries the combination rule set for the segment group, obtains the number of the aggregated segment group code, and generates the stage calibration coding sequence.

[0052] The image frame sequence structure building blocks include:

[0053] The time index generation submodule extracts image capture time information based on the evolution trajectory of oil-tea camellia lesions, reads the capture time field and establishes a correspondence between image and time, uses the time field to arrange the image frame sequence, compares the image number field and detects the number increment pattern, filters out frames with missing numbers, removes duplicate recorded frame data, establishes a unified frame sequence index sequence by matching the number and time field, merges frames with adjacent time field intervals less than the set valley value, synchronizes the timestamps, and generates a time index sequence set;

[0054] Continuous sequence composition submodule: Based on the time index sequence set, the difference in the timestamp field between frames is calculated and the interval interval value is recorded. The sequence gap position is filtered by the time interval threshold, the data frame records above the interval peak are deleted and the images corresponding to the discontinuous frames are removed. The missing frame data is supplemented and the frame number segments are merged according to the time sequence. The continuous image frame sequence structure is constructed, and the combined frame set is renumbered to generate a continuous sequence composition list.

[0055] Time index generation submodule: Based on the evolution trajectory of oil-tea camellia lesions, the image EXIF ​​information parsing method is used to extract the image shooting time field row, which is converted into the time field format. The correspondence between image and time is established through the time image mapping structure. The image frame sequence is arranged in ascending order using the time field. The image number fields are compared in sequence and the number sequence increment detection is performed. The number jump frame records are marked and filtered out. The hash summary algorithm is called to compare the image file fingerprint and identify image data redundancy items. Duplicate frame data with completely consistent hash values ​​are cleared. By comparing the consistency of the image number and time field sequence, a unified frame sequence index sequence is constructed. Image frames with time field differences between adjacent frames less than the set valley value are merged. The frame group time field is synchronously adjusted to generate a time index sequence set.

[0056] Continuous sequence composition submodule: Based on the time index sequence set, the frame time field difference is used to calculate the execution time interval, record and extract the time interval interval value between each image frame, set the time interval peak as the filtering condition, delete the frame record above the peak and remove the corresponding image data, mark the break position between frames and organize the numbered segments accordingly, call the image interpolation processing method to fill in the missing numbered segments, and construct a timestamp field for the supplemented frame, number and integrate the corrected image frames, and construct a structured frame sequence set through the combination of the number and time double fields to generate a continuous image frame sequence structure, output the correspondence between the renumbered frame path, time field and number field, and generate a continuous sequence composition list.

[0057] The lesion evolution sequence screening module includes:

[0058] Edge Difference Extraction Submodule: Based on the continuous sequence composition list, the image grayscale values ​​are separated and the grayscale change segments of the regional boundaries are obtained. The edge texture features are extracted through the convolutional neural network. The pixel gradient is scanned line by line to extract the outer edge position of the lesion contour and mark the coordinate point set. The difference between the coordinates of the corresponding numbered contour points between frames is calculated and the pixel offset coordinates are filtered to obtain the edge change segment and mark the pixel change set, thereby generating the edge pixel difference set.

[0059] Area change calculation submodule: Based on the edge pixel difference set, it tracks the closed boundary of the adjacent pixel connection area, calculates the area value of the lesion in the area, extracts the area change amplitude between consecutive frames by using frame number mapping comparison, constructs a change trend list using the inter-frame difference sequence, marks the direction of area increase and decrease, and generates an area dynamic feature sequence;

[0060] Morphological combination construction submodule: Based on the area dynamic feature sequence, the pixel range of the main color area of ​​the lesion in each frame is obtained and the location of the peak number of frequency pixel blocks is extracted. The pixel center of gravity coordinates are located, the center point of the main color position is obtained, the Euclidean distance of the coordinate values ​​between frames is calculated and an offset path is generated. The area change value and the center of gravity path pair are numbered and then merged into a sequence of structure segment pairs to generate a morphological evolution change combination group;

[0061] Edge difference extraction submodule: Based on the continuous sequence composition list, the image grayscale channel separation method is used to perform grayscale conversion on the image pixel matrix, extract the single channel matrix and calculate the grayscale value, use the Sobel operator to calculate the gradient in the horizontal and vertical directions respectively to generate the edge response map, extract the non-maximum grayscale change segment, use the convolutional neural network to extract the edge texture features of the image, define the network structure as a combination of three convolution layers and two pooling layers, the convolution kernel size is set to 3×3, the stride is 1, and the activation function uses ReLU. The output result is flattened and then connected to the fully connected layer to obtain the edge feature map, extract the gradient response mutation position and locate the outer edge boundary coordinates of the contour to form a coordinate point set, use the frame number matching method to calculate the coordinate difference of the points with the same number in adjacent frames and filter out the zero difference point pairs, record the non-zero difference points and mark the pixel positions, output the image dot matrix set where the edge pixels have changed, and generate the edge pixel difference set;

[0062] Area change calculation submodule: Based on the edge pixel difference set, the contour closure tracking method is used to detect the pixel difference area. The contour search function is called to locate the boundaries of continuous pixel blocks in the binary image and identify the closed contour area. The pixel values ​​within the closed area are counted to obtain the total area pixel of the lesion area. The area of ​​the same area in the previous and next frames is calculated by number mapping. The increase or decrease in area between frames is obtained and recorded as the change value. The area difference between consecutive frames is sorted by number to construct a sequence list. The increase or decrease direction is determined based on the positive or negative value of the area difference to generate a dynamic area feature sequence.

[0063] Morphology combination construction sub-module: based on the area dynamic feature sequence, the image main color extraction method is used to convert the image to execute the color space, the RGB image is converted to the HSV color space, the main color distribution frequency in each pixel block is counted, the position index of the frequency maximum pixel block area is obtained to obtain the region mask graph, the pixel barycenter coordinates are calculated in the mask graph, the center point is obtained as the main color position, the center point coordinates of each frame main color are paired with the previous frame, the Euclidean distance between the corresponding points is calculated and recorded as the path displacement value, the area change value and the path displacement value of each frame are paired and assigned a number label, the numbered segments are combined into a structured segment pair sequence, and the morphology evolution change combination group is output and arranged.

[0064] The convolutional neural network is according to the formula:

[0065]

[0066] Wherein: F'(i,j) represents the edge enhancement convolution response value at position i,j, K(m,n) represents the weight value of the convolution kernel at the m,n position, I(i+m,j+n) represents the pixel gray value of position i+m,j+n in the image, G(i+m,j+n) represents the gray gradient value of position i+m,j+n in the image, alpha represents the fusion weight coefficient of the pixel value channel, beta represents the fusion weight coefficient of the gray gradient channel, i represents the row index of the current output feature map, j represents the column index of the current output feature map, m represents the index position of the convolution kernel along the row direction, and n represents the index position of the convolution kernel along the column direction.

[0067] The execution process is as follows: first, the gray value I(i+m,j+n) of the current pixel point at the index position (i+m,j+n) in the input image is extracted, the gray difference graph is constructed combined with the adjacent pixels, the gray gradient value G(i+m,j+n) of the corresponding position is obtained, the fusion weight coefficients alpha and beta are applied to the image pixel channel and the gradient channel respectively, the weighted combination term alpha·I(i+m,j+n)+beta·G(i+m,j+n) is formed, the weighted value is multiplied by the weight value K(m,n) of the position (m,n) in the convolution kernel, and the weighted sum is performed by traversing the entire convolution kernel region, and finally the edge enhancement convolution response value F'(i,j) at position (i,j) is obtained.

[0068] Separating the image gray value means converting the original color image into a single-channel gray image, specifically, the brightness values of the red, green and blue channels of the pixels in each frame of oil tea leaf image are read in turn, the corresponding gray value is generated by weighted superposition, and the original multi-channel pixel information is replaced to form a unified gray image structure.

[0069] The multi-dimensional label encoding construction module comprises:

[0070] Area ratio calculation submodule: Based on the morphological evolution change combination group, the pixels in the lesion area in each frame are marked and the lesion pixel locations are extracted. The total number of lesion pixels in a single frame is obtained, and the total number of pixels in the entire frame image is counted. After performing time-dependent modeling on the lesion area ratio sequence of consecutive frames through a long short-term memory network, the dynamic change trend is captured. The ratio of the lesion area to the total pixels is calculated and recorded in the ratio sequence to generate a frame area ratio sequence.

[0071] Direction vector extraction submodule: Based on the frame area ratio sequence, it extracts the pixel blocks in the main color area of ​​each frame image and obtains the main color pixel set. It calculates the center of gravity of the main color pixel and obtains the center position of the coordinate point of the main color area within the frame. It calculates the coordinates of the main color center of gravity between frames and extracts the direction angle value. It divides the direction angle into intervals, obtains the offset direction segment number and associates it with the frame index one-to-one to generate the color offset direction sequence.

[0072] Label segment value assembly submodule: Based on the color offset direction sequence, it obtains the area ratio segment number, offset direction segment number, and position offset number segment number in each frame. It combines the three segment numbers into a single-frame label encoding field through a fixed number splicing sequence. It then splices the label fields of each sequence and aggregates them into an overall label vector set in frame order to generate a label encoding composite vector set.

[0073] Area ratio calculation submodule: Based on the morphological evolution change combination group, a long short-term memory network model is used. The input frame sequence length is set to 10 frames, and the input of each frame is the normalized lesion area ratio. The hidden layer dimension is set to 64, the number of stacked network layers is set to 2, and the bidirectional structure is enabled. The number of pixels in the lesion area of ​​each frame image is extracted. The total number of image pixels is counted as the divisor and the ratio is normalized. The input sequence is constructed in frame order, and a time step state transfer path is established to enter the network. An update gate is set to control the state output, and the state output tensor of each frame is cached. The predicted values ​​are spliced ​​by frame index to generate a frame area ratio sequence.

[0074] Direction vector extraction submodule: Based on the frame area ratio sequence, the main color area position is extracted. The number of clusters is set to 3, and the brightness-independent color channel is selected and clustered. The cluster center adopts the optimal initialization strategy. The number of iterations is limited to 300, and the error threshold is 1%. The largest category pixel area is selected, the pixel coordinate set of the largest category pixel area is extracted, and the center of gravity position is counted. The angular offset between the main color centers of adjacent frames is calculated, and the complete angle interval is divided into 8 segments. The mapping number corresponds one-to-one with the frame index to generate a color offset direction sequence.

[0075] Label segment value assembly submodule: Based on the color offset direction sequence, the segment number splicing method is used to integrate the area ratio segment number, direction segment number and position offset segment number respectively. The area ratio is divided into ten segments from zero to one, the direction number adopts the aforementioned mapping, and the position offset is divided into five segments according to the displacement distance. The three numbers are spliced ​​in a unified order to form a single-frame label field. Each frame is arranged in sequence according to the label to form a label vector set, generating a label encoding synthetic vector set.

[0076] Long short-term memory network follows the formula:

[0077] H s =Z s tanh(U s ·M s-1 +V s tanh(W z ·[H s-1 ,A s ,D s ,Q s ]+B z ))

[0078] Among them: H s Represents the dynamic state output value of the sth frame, which is used to reflect the degree of evolutionary response in the time series. s Indicates the output gate control factor of the sth frame, which controls whether the state is output at the moment, U s Indicates the retention coefficient of the s-th frame for the memory item, M s-1 Represents the memory state value of the s-1th frame, V s represents the activation control coefficient for the current input update in the sth frame, W z Represents the weight matrix of the state update operation, [H s-1 ,A s ,D s ,Q s ] represents the input vector set, H s-1 Indicates the dynamic state output value of the s-1 frame, A s It represents the ratio of the lesion area in the sth frame to the total number of pixels in the entire frame image, D s It represents the absolute jump value of the ratio of the lesion area between the sth frame and the s-1th frame, and is calculated as D s =|A s -A s-1 |, Q s represents the compactness index of the lesion area in the sth frame, B z represents the bias term, s represents the frame index, and tanh represents the hyperbolic tangent function;

[0079] Execution process: first, the morphological evolution change combination group is processed, the total number of pixel points in the lesion area and the total number of pixels in the whole frame image are extracted, the lesion area ratio A is calculated s , and the difference D s between the ratio and the previous frame is combined s = |A s-1 | as a structural index representing area mutation, the compactness coefficient Q s of the lesion area is extracted as a geometric constraint parameter reflecting the convergence degree of the area shape, the three indexes and the previous frame dynamic state output H s-1 are spliced into the input sequence vector [H s-1 , A s , D s , Q s ], the vector is linearly combined with the weight matrix W z and the bias term B z , then input to the hyperbolic tangent activation unit, output the current frame activation response, and then combine the previous frame memory state M s-1 with the gate coefficients U s , V s weighted and superimposed to update the current frame state, and then control whether to output the current state response H s through the output gate Z s , and construct a continuous frame area ratio dynamic sequence.

[0080] Labeling the lesion area means locating and identifying the pixel level of the area showing abnormal color and texture features in each frame image. Specifically, first, the image is converted from RGB to a color model with brightness-chroma separation through color space transformation, then the color range threshold is set according to the common brownish red and grayish brown tones of tea plant disease spots, the pixel set that meets the conditions is extracted, the high light and shadow interference area with similar color tone to healthy leaves is excluded, the closed and semi-closed color abnormal connected region is extracted as the preliminary disease spot candidate area, and the area lower limit, edge irregularity and texture roughness are reviewed to retain the pixel set that meets the typical morphology of disease spots as the disease spot labeling area.

[0081] Continuous logical tag set group includes:

[0082] Segment value pair screening submodule: based on the label encoding synthesis vector set, the segment value field of adjacent frame position in the label field sequence is extracted, the field sliding window is combined frame by frame, and the label segment value pair is formed, the difference value of each numbered segment value is calculated to obtain the record difference value sequence, the segment value discontinuous field group is screened, the cumulative number of the same segment value pair in the whole sequence is counted, and then compared with the frequency threshold, the segment value group pair with cumulative number less than the set threshold is deleted, and the high-frequency legal segment value combination set is generated;

[0083] Logical set construction submodule: based on the high-frequency legal segment value combination set, the forward and reverse mapping dictionary table and index structure are established, the chain splicing is carried out according to the frame order, the splicing termination condition is set as reaching the maximum chain length, the maximum chain length is set as 20 field units, the chain structure record unit is established, the chain starting point and ending point segment numbers are marked, and whether the chain is closed is judged according to the condition that the chain starting number is equal to the tail number, the chain group structure meeting the closed judgment rule is extracted, the frame segment value in the chain group is mapped with the original label index position, the many-to-one structure index table between the chain group number and the frame number is established, and the continuous logical label set group is generated.

[0084] Segment value pair screening submodule: based on the label coding synthesis vector set, the window step is set as 1 and the window width is set as 2, the adjacent two label fields in the label vector sequence are extracted, the numbers in the three segment values in each frame field are obtained, the current frame and the next frame number segment are combined into a pair group by calling the offset comparison structure, the difference of each segment number in the group is recorded, the difference value is calculated by using the fixed-length integer mapping method, the integer value difference after the number conversion is calculated, the segment value difference tolerance is set as 1, the segment value combination with the difference value exceeding the tolerance is marked as a non-continuous field, the counting accumulation structure is called to record each segment value combination, the frequency threshold is set as 5, the screening table structure is constructed to traverse each combination and delete the combinations below the threshold, the remaining field group set is output, and the high-frequency legal segment value combination set is generated.

[0085] Logical set construction submodule: based on the high-frequency legal segment value combination set, the segment value field pairs in the combination set are extracted, the former segment number is taken as the starting point and the latter segment number is taken as the connection target, the forward mapping dictionary table and the reverse mapping index structure are established, the chain splicing is carried out according to the frame order, the splicing termination condition is set as reaching the maximum chain length, the maximum chain length is set as 20 field units, the chain structure record unit is established, the chain starting point and ending point segment numbers are marked, whether the chain is closed is judged according to the condition that the chain starting number is equal to the tail number, the chain group structure meeting the closed judgment rule is extracted, the frame segment value in the chain group is mapped with the original label index position, the many-to-one structure index table between the chain group number and the frame number is established, and the continuous logical label set group is generated.

[0086] The stage level coding generation module comprises:

[0087] Label segment group determination submodule: based on the continuous logical label set group, the closed edge segment value in the label field is extracted and the closed number field is identified, the number marker area with the closed feature is extracted through the coordinate connectivity screening, the fields in the set are matched, the offset direction segment value is extracted and positioned to the same frame label field, the area jump segment value and the position offset segment value corresponding to the above position are matched frame by frame, the three segment group comparison structure is established by combining the segment value fields, the field configuration template in the combination rule set is matched, the irregular combination is screened out, and the label combination segment group list is generated.

[0088] Phase code generation submodule: Based on the label combination segment group list, the segment group number field structure is split and the index value in each combination position is extracted. The segment value combination field is mapped one by one with the preset coding rule table, the mapping code number is located, the number sequence is merged according to the original frame order, and the mapping index between the frame and the phase number is established using the sequential sequence structure. The mapping relationships are integrated and the continuous segment number results are output to generate the phase calibration code sequence.

[0089] Label segment group determination submodule: Based on the continuous logical label set group, the segment value screening condition is set as the field number is included in the closed segment value index set, the closed edge segment value in each frame label field is extracted and the number index is marked, the number adjacency judgment condition is used to determine the inter-frame connectivity of the closed number segment, the segment difference threshold is set to 1, the frame number set that meets the connectivity relationship is extracted, the offset direction segment value is extracted according to the number index and located at the same row number field position, and the area jump segment value and position offset segment value of each frame are extracted at the same time, and then spliced ​​in the order of area number, direction number, and position number to form a three-segment group structure, the field template matching table is set as a fixed number set, each group number field is compared with the configuration item in the table, the unmatched structure is eliminated, and a label combination segment group list is generated;

[0090] Stage code generation submodule: Based on the label combination segment group list, each combination word is extracted in the set order to form a three-field number group with a fixed structure. The mapping rule table is set as a static number comparison table. The corresponding code number is searched according to the field combination item. The mapped code numbers are arranged and merged in the original frame order. After constructing the mapping array, a dictionary structure is established with the frame index. Each frame number is matched with the sequence of code numbers to generate a stage calibration code sequence.

[0091] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The oil-tea camellia pest and disease identification system based on image recognition technology is characterized by: The system comprises: Image frame sequence structure construction module: Based on the evolution trajectory of oil-tea camellia lesions, the image time sequence is extracted, the sequence sorting index is generated through the shooting time information, the discontinuous data is filtered out and the timestamps are merged to generate a continuous sequence composition list; Lesion evolution sequence screening module: Based on the continuous sequence composition list, the convolutional neural network is used to extract the edge texture features of the lesion, the contour line area values ​​of the previous and next frames are calculated to obtain the area change, the distance between the main color center points is compared and the offset path is obtained, and the morphological evolution change combination group is generated; Multidimensional label coding construction module: Based on the morphological evolution change combination group, the inter-frame evolution characteristics are modeled through a long short-term memory network, the ratio of the pixel area of ​​the lesion area to the total pixels in each frame is calculated, the segment value numbers corresponding to each parameter are aggregated in sequence and the label codes are assembled to generate a label code synthetic vector set; Label legitimacy combination analysis module: Based on the label encoding synthetic vector set, the label segment value fields are combined in pairs, and pairs that do not meet the continuous segment value arrangement structure are filtered out. Label combination groups with label encoding combination frequencies above the threshold are retained and logically reconstructed to generate a continuous logical label set group; Stage level coding generation module: Based on the continuous logical label set group, determine the closed edge segment and color offset direction label segment value, match the area jump value and position offset segment value, query and query the combination rule set for segment group, obtain the number of the aggregated segment group code, and generate the stage calibration code sequence.

2. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 1, characterized in that: The image frame sequence structure building module includes: The time index generation submodule extracts image capture time information based on the evolution trajectory of oil-tea camellia lesions, reads the capture time field and establishes a correspondence between image and time, uses the time field to arrange the image frame sequence, compares the image number field and detects the number increment pattern, filters out frames with missing numbers, removes duplicate recorded frame data, establishes a unified frame sequence index sequence by matching the number and time field, merges frames with adjacent time field intervals less than the set valley value, synchronizes the timestamps, and generates a time index sequence set; Continuous sequence composition submodule: Based on the time index sequence set, calculate the difference in the timestamp field between frames and record the interval value, filter the sequence discontinuity position through the time interval threshold, delete the data frame records above the interval peak and remove the corresponding images of the discontinuous frames, supplement the missing frame data and merge the frame number segments according to the time sequence, construct a continuous image frame sequence structure, renumber the combined frame set, and generate a continuous sequence composition list.

3. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 1 is characterized in that: The lesion evolution sequence screening module includes: Edge difference extraction submodule: Based on the continuous sequence composition list, the image grayscale values ​​are separated and the grayscale change segments of the regional boundaries are obtained. The edge texture features are extracted through a convolutional neural network. The pixel gradient is scanned line by line to extract the outer edge position of the lesion contour and mark the coordinate point set. The difference in the coordinates of the corresponding numbered contour points between frames is calculated and the pixel offset coordinates are screened to obtain the edge change segment and mark the pixel change set, thereby generating the edge pixel difference set. Area change calculation submodule: Based on the edge pixel difference set, the closed boundary of the adjacent pixel connection area is tracked, the area value of the lesion in the area is calculated, the area change amplitude between consecutive frames is extracted by frame number mapping and comparison, and a change trend list is constructed using the inter-frame difference sequence, the direction of area increase and decrease is marked, and a dynamic area feature sequence is generated; Morphological combination construction submodule: Based on the area dynamic feature sequence, the pixel range of the main color area of ​​the lesion in each frame is obtained and the position of the peak number of frequency pixel blocks is extracted, the pixel center coordinates are located, the center point of the main color position is obtained, the Euclidean distance of the coordinate values ​​between frames is calculated and an offset path is generated, the area change value and the center path pair are numbered, and then merged into a sequence of structure segment pairs to generate a morphological evolution change combination group.

4. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 3 is characterized in that: The convolutional neural network is based on the formula: Where: F′(i,j) represents the edge enhanced convolution response value at position i,j, K(m,n) represents the weight value of the convolution kernel at the mth and nth positions, I(i+m,j+n) represents the pixel grayscale value at position i+m,j+n in the image, G(i+m,j+n) represents the grayscale gradient value at position i+m,j+n in the image, α represents the fusion weight coefficient of the pixel value channel, β represents the fusion weight coefficient of the grayscale gradient channel, i represents the row index of the current output feature map, j represents the column index of the current output feature map, m represents the index position along the row direction in the convolution kernel, and n represents the index position along the column direction in the convolution kernel.

5. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 3 is characterized in that: The separated image grayscale value refers to converting the original color image into a single-channel grayscale image, specifically reading the brightness values ​​of the red, green, and blue channels of the pixels in each frame of the oil tea leaf image in turn, generating the corresponding grayscale value through weighted superposition, and replacing the original multi-channel pixel information to form a unified grayscale image structure.

6. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 1 is characterized in that: The multi-dimensional label encoding building block includes: Area ratio calculation submodule: Based on the morphological evolution change combination group, the pixels of the lesion area in each frame of the image are marked and the lesion pixel positions are extracted. The total number of lesion pixels in a single frame is obtained, and the total number of pixels in the entire frame of the image is counted. After performing time-dependent modeling on the lesion area ratio sequence of consecutive frames through a long short-term memory network, the dynamic change trend is captured. The ratio of the lesion area to the total pixels is calculated and the ratio sequence is recorded to generate a frame area ratio sequence. Direction vector extraction submodule: Based on the frame area ratio sequence, extract the pixel blocks of the main color area of ​​each frame image and obtain the main color pixel set, calculate the center of gravity of the main color pixel, obtain the center position of the coordinate point of the main color area within the frame, calculate the coordinates of the main color center of gravity point between each frame and extract the direction angle value, divide the direction angle interval, obtain the offset direction segment number and then associate it with the frame index one-to-one to generate a color offset direction sequence; Label segment value assembly submodule: Based on the color offset direction sequence, the area ratio segment number, offset direction segment number and position offset quantity segment number in each frame are obtained, and the three segment numbers are combined into a single-frame label coding field through a fixed number splicing order. The label fields of each sequence are spliced ​​and aggregated into an overall label vector set in frame order to generate a label coding synthetic vector set.

7. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 6 is characterized in that: The long short-term memory network is based on the formula: H s =Z s ·tanh(U s ·M s-1 +V s ·tanh(W z ·[H s-1 ,A s ,D s ,Q s ]+B z )) Among them: H s Represents the dynamic state output value of the sth frame, which is used to reflect the degree of evolutionary response in the time series. s Indicates the output gate control factor of the sth frame, which controls whether the state is output at the moment, U s Indicates the retention coefficient of the s-th frame for the memory item, M s-1 Represents the memory state value of the s-1th frame, V s represents the activation control coefficient for the current input update in the sth frame, W z Represents the weight matrix of the state update operation, [H s-1 ,A s ,D s ,Q s ] represents the input vector set, H s-1 Indicates the dynamic state output value of the s-1 frame, A s It represents the ratio of the lesion area in the sth frame to the total number of pixels in the entire frame image, D s It represents the absolute jump value of the ratio of the lesion area between the sth frame and the s-1th frame, and is calculated as D s =|A s -A s-1 |, Q s represents the compactness index of the lesion area in the sth frame, B z represents the bias term, s represents the frame index, and tanh represents the hyperbolic tangent function.

8. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 6, characterized in that: The marking of diseased spots refers to locating and pixel-level identifying areas showing abnormal color and texture features in each frame of the image. Specifically, the image is first converted from RGB to a brightness-chromaticity separation color model through color space transformation, and then a color range threshold is set based on the common brown-red and gray-brown tones of oil-tea camellia lesions. A set of pixels that meet the conditions is extracted, and highlight and shadow interference areas close to the tones of healthy leaves are excluded. Closed and semi-closed color abnormal connected areas are extracted as preliminary diseased spot candidate areas. The areas are reviewed based on the lower limit of the area, the degree of edge irregularity and the roughness of the texture, and the pixel sets that meet the typical morphology of the lesions are retained as the diseased spot marking areas.

9. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 1, characterized in that: The continuous logical label set group includes: Segment value pair screening submodule: Based on the label encoding synthetic vector set, the segment value fields of adjacent frame positions in the label field sequence are extracted. The field sliding window is combined frame by frame to form label segment value pairs. The difference between the numbered segment values ​​is calculated to obtain a record difference sequence. The segment value discontinuous field groups are screened. The cumulative number of the same segment value pair in the entire sequence is counted and then compared with the frequency threshold. The segment value pairs with a cumulative number below the set threshold are deleted to generate a high-frequency legal segment value combination set. Logical set construction submodule: Based on the high-frequency legal segment value combination set, the front and back association relationship of each segment value combination field is constructed, the segment value numbers of the front and back items in the combination field are connected and a chain mapping table is established, the chain extension method is used to splice the continuously connected segment values ​​backward to construct a label segment sequence, the segment value structure in each chain is called and whether it is closed, the structure group that meets the chain head and chain tail overlap conditions is screened, the chain group number is used to map the corresponding label frame sequence, a full sequence connection table is established, and a continuous logical label set group is generated.

10. The oil-tea camellia pest and disease identification system based on image recognition technology according to claim 1, characterized in that: The stage level code generation module includes: The label segment group determination submodule extracts the closed edge segment values ​​in the label field and identifies the closed number field based on the continuous logical label set group. The numbered mark area with closed characteristics is extracted through coordinate connectivity screening. The fields in the set are matched. The offset direction segment value is extracted and located in the same frame label field. The area jump segment value and position offset segment value corresponding to the above position are matched frame by frame. The segment value field is combined to establish a three-segment group comparison structure. The field configuration template in the combination rule set is matched. Non-compliant combinations are screened out to generate a label combination segment group list. Stage coding generation submodule: Based on the label combination segment group list, split the segment group number field structure and extract the index value in each combination position, map the segment value combination field with the preset coding rule table one by one, locate the mapping coding number, merge the number sequence according to the original frame order, use the sequential sequence structure to establish the mapping index between the frame and the stage number, integrate the mapping relationships and output the continuous segment number results to generate the stage calibration coding sequence.

Images