A water product seedling quality grading method and system based on seedling vitality image features

Abstract

This invention discloses a method and system for grading the quality of aquatic seedling cultivation based on seedling vitality image features, belonging to the field of image processing technology. This method for grading the quality of aquatic seedling cultivation based on seedling vitality image features acquires a continuous sequence of seedling images and divides local processing areas. It extracts image change information to construct an observed change state vector, determines the foreground processing area based on this vector, constructs a dynamic environment reference set, establishes a local environmental disturbance propagation matrix to separate environmental disturbance changes from non-environmental changes, obtains candidate non-environmental change components, constructs candidate motion evidence chains through association, determines valid motion trajectory events, and calculates confidence weights to construct a seedling vitality feature vector. This invention determines the seedling cultivation quality level by performing quality assessment processing on the seedling vitality feature vector and candidate motion evidence chains, thereby reducing the probability of misjudging environmental fluctuations as seedling activity and making the seedling cultivation quality grading results closer to the actual activity level of the seedlings.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for grading the quality of aquatic seedlings based on the image features of seedling vitality. Background Technology

[0002] In the process of factory-scale aquaculture seedling production, seedlings need to be graded to select high-quality seedlings for subsequent breeding. Seedling ponds usually use aeration to maintain dissolved oxygen in the water. The circulating water flow and the large number of bubbles, water surface ripples, and suspended particles such as uneaten feed and feces generated by aeration keep the water in the pond in a state of continuous dynamic disturbance. Seedlings such as shrimp larvae and fish larvae are very small and their bodies are transparent or semi-transparent. When they gather in high density, it is almost impossible to distinguish individual individuals with the naked eye or conventional imaging methods.

[0003] Currently, most image analysis methods used for evaluating the quality of aquatic seedlings involve acquiring continuous images of seedling populations in seedling ponds, using methods such as inter-frame difference and optical flow calculation to extract regions and intensity of changes in the images, and then evaluating the vitality of the seedling population based on these change characteristics.

[0004] The limitations of existing technologies include at least the following problems: In factory-scale aquaculture seedling production, when acquiring a series of continuous frame images of seedlings in a seedling pond using a fixed image acquisition device, and then using this to grade the quality of seedling population vitality, the continuous operation of aeration and circulating water in the seedling pond results in a large number of air bubbles, refraction changes caused by water surface fluctuations, and suspended particles such as uneaten feed and feces in the water. These environmental factors cause grayscale and texture fluctuations in local image areas between consecutive frames. The natural activities of individual seedlings, such as tail wagging, jumping, and following the current, also manifest as changes in grayscale and texture of local image areas over time. These two types of changes superimpose on each other at the same spatiotemporal location in the image sequence. When extracting image change features, existing technologies struggle to distinguish the source of change from the mixed change information, leading to a deviation between the seedling vitality inferred from image features and the actual activity state of the seedlings. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for grading the quality of aquatic seedlings based on seedling vitality image features. This method solves the problem that the superimposed image changes caused by environmental disturbances and seedling movement make it difficult to accurately extract relevant vitality features.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for grading the quality of aquatic seedling cultivation based on seedling vitality image features, comprising the following steps: acquiring a continuous frame sequence of seedling images in an aquatic seedling pond and dividing it into multiple local processing regions; extracting image change information from each local processing region and constructing a corresponding observation change state vector; performing dynamic analysis on the local processing regions based on the observation change state vector, determining the foreground processing region, and selecting multiple environmental reference regions from a preset range of the foreground processing region to construct a dynamic environmental reference set; establishing a local environmental disturbance propagation matrix based on the observation change state vectors corresponding to each environmental reference region in the dynamic environmental reference set, and separating environmental disturbance changes from non-environmental changes to obtain candidate non-environmental change components; associating the candidate non-environmental change components to construct a candidate motion evidence chain, and determining valid motion trajectory events through motion constraint judgment processing; extracting the motion evidence set of valid motion trajectory events, calculating confidence weights, and constructing a seedling vitality feature vector; and performing quality assessment processing on the seedling vitality feature vector and the candidate motion evidence chain to determine the seedling quality level.

[0007] Furthermore, the specific steps for constructing the observed change state vector are as follows: perform differential processing on adjacent frames of seedling images in a continuous frame seedling image sequence to obtain image change information corresponding to each local processing area; mark the image change information as change feature parameters; perform standardization processing on the change feature parameters to construct the observed change state vector.

[0008] Furthermore, the specific steps for constructing a dynamic environment reference set are as follows: Calculate the corresponding activity level of change based on the observed change state vector of each local processing region; determine the local processing region whose activity level exceeds a preset threshold as the foreground processing region; obtain multiple candidate regions within a preset range of the foreground processing region; calculate the regional association feature set between each candidate region and the foreground processing region, including regional distance, grayscale change difference, and historical observed change state difference; determine multiple environment reference regions from the multiple candidate regions based on the regional association feature set; combine the multiple environment reference regions to form a dynamic environment reference set.

[0009] Furthermore, the specific steps for establishing the local environmental disturbance propagation matrix are as follows: extract the component sequence from the observed change state vector corresponding to each environmental reference area, and calculate the cross-correlation curve to determine the propagation delay of each reference area; perform time correction on the observed change state vector according to the time delay parameter; combine the time-corrected observed change state vectors to obtain the environmental reference state vector; and establish the local environmental disturbance propagation matrix based on the environmental reference state vector and the observed change state vector corresponding to the foreground processing area.

[0010] Furthermore, the specific steps for separating environmental disturbance changes from non-environmental changes are as follows: decompose the observed change state vector corresponding to the foreground processing area into an environmental disturbance change term and a candidate non-environmental change component; separate the candidate non-environmental change component based on the local environmental disturbance propagation matrix in the environmental reference state vector and the environmental disturbance change term; and apply sparsity constraints to the candidate non-environmental change component to obtain the candidate non-environmental change state vector.

[0011] Furthermore, the specific steps for constructing the candidate motion evidence chain are as follows: Obtain the candidate non-environmental change state vectors corresponding to each frame in the continuous frame seedling image sequence; extract the position parameters of the foreground processing region corresponding to each candidate non-environmental change state vector and parse out the motion state parameters; associate the candidate non-environmental change state vectors in adjacent frames based on the position parameters and motion state parameters corresponding to each candidate non-environmental change state vector in adjacent frames; connect the associated candidate non-environmental change state vectors in chronological order to form a candidate non-environmental change state vector sequence; construct the candidate motion evidence chain based on the position parameters and motion state parameters corresponding to each vector in the candidate non-environmental change state vector sequence.

[0012] Furthermore, the specific steps of motion constraint determination are as follows: Calculate spatial migration continuity based on the position parameters of adjacent frames in the candidate motion evidence chain; calculate the structural evolution and displacement coordination degree based on the motion state parameters in the candidate motion evidence chain; extract the structural features corresponding to each frame in the candidate motion evidence chain and calculate the structure preservation parameter; count the number of consecutive frames corresponding to the candidate motion evidence chain and determine the duration parameter; perform motion constraint determination on the candidate motion evidence chain based on the spatial migration continuity, structural evolution and displacement coordination degree, structure preservation parameter, and duration parameter; and take the candidate motion evidence chain that passes the motion constraint determination as a valid motion trajectory event.

[0013] Furthermore, the specific steps for constructing the seedling vitality feature vector are as follows: Based on the candidate non-environmental change state vectors of effective movement trajectory events, determine the intensity of movement change; extract the movement evidence set of effective movement trajectory events, and calculate the corresponding confidence weights in combination with the intensity of movement change; use the confidence weights to perform weighted statistics on the movement state parameters corresponding to the effective movement trajectory events; construct the seedling vitality feature vector based on the weighted statistics of the movement state parameters and in combination with the candidate movement evidence chain corresponding to the effective movement trajectory events.

[0014] Further, the specific steps for determining the seedling quality grade are as follows: comprehensively process the seedling vitality feature vector to obtain the seedling comprehensive vitality value; read the motion evidence set in the candidate motion evidence chain corresponding to the effective motion trajectory event and calculate the vitality stability value; determine the credibility of the image evidence, and combine the seedling comprehensive vitality value and vitality stability value to classify the seedling grade and determine the seedling quality grade.

[0015] A system for grading the quality of aquatic seedling cultivation based on seedling vigor image features includes: a data acquisition unit for acquiring a continuous frame sequence of seedling images in an aquatic seedling pond and dividing it into multiple local processing regions; an observation change construction unit for constructing corresponding observation change state vectors; an environmental reference construction unit for constructing a dynamic environmental reference set; a non-environmental change separation unit for establishing a local environmental disturbance propagation matrix and separating environmental disturbance changes from non-environmental changes to obtain candidate non-environmental change components; a motion evidence chain construction unit for constructing candidate motion evidence chains and determining valid motion trajectory events through motion constraint judgment processing; a vigor feature extraction unit for calculating confidence weights and constructing seedling vigor feature vectors; and a seedling quality grading unit for determining the seedling quality level.

[0016] The present invention has the following beneficial effects: (1) The aquatic seedling quality grading method based on seedling vitality image features decomposes the observed change state vector of the foreground processing area by establishing a local environmental disturbance propagation matrix. It separates the environmental disturbance change term caused by environmental factors such as bubbles and water waves from the candidate non-environmental change component caused by seedling activities. The environmental reference area in the seedling pond and the foreground processing area are in similar water environment. There is a corresponding relationship between the transmission of environmental disturbances between the two. By constructing the propagation matrix using the observed change state vectors of multiple environmental reference areas, the amount of change of environmental factors in the foreground processing area can be estimated. After deducting this part of the environmental disturbance change term from the original observed change state vector, the interference of environmental factors in the candidate non-environmental change component is weakened, and the changes caused by seedling activities are highlighted. When calculating the seedling vitality index, the candidate non-environmental change component after separation is used as input, which reduces the probability of misjudging bubble movement or water wave fluctuation as seedling activity, and makes the seedling quality grading result closer to the actual activity level of the seedlings.

[0017] (2) The aquatic seedling quality grading method based on seedling vitality image features screens candidates by constructing a candidate motion evidence chain and judging motion constraints. At the same time, it examines the spatial migration continuity, structural evolution and displacement coordination, structural maintenance parameters and duration parameters. Candidate motion evidence chains are required to meet the threshold in these dimensions before they can be identified as valid motion trajectory events. Thus, it effectively distinguishes between the difficult-to-distinguish environmental pseudo-motion and the real movement of seedlings, and reduces the interference of residual environmental disturbance on vitality evaluation.

[0018] (3) The aquatic seedling quality grading method based on seedling vitality image features uses confidence weights to perform weighted statistics on effective motion trajectory events when constructing the seedling vitality feature vector. The calculation of confidence weights integrates the intensity of motion change and motion evidence. The intensity of motion change reflects the amplitude of the candidate non-environmental change component corresponding to the trajectory event. For trajectory events with smooth spatial migration, intact structure, good coordination of posture and displacement, and long duration, the confidence weight is higher and contributes more to the population vitality statistics. For some trajectory events with low evidence scores, the confidence weight is reduced accordingly. The seedling vitality feature vector obtained by the confidence weight weighted statistics is closer to the real and reliable motion level in the seedling population, so as to improve the credibility of seedling quality grading.

[0019] (4) The aquatic seedling quality grading method system based on seedling vitality image features establishes a local environmental disturbance propagation matrix by using the non-environmental change separation unit, the observation change state vector provided by the observation change construction unit and the dynamic environmental reference set provided by the environmental reference construction unit, and separates the environmental disturbance change items and candidate non-environmental change components in the foreground processing area to reduce the interference of environmental factors such as bubbles and water waves on seedling motion information; the motion evidence chain construction unit performs inter-frame correlation and motion constraint judgment on the candidate non-environmental change components to screen out effective motion trajectory events; the vitality feature extraction unit calculates the confidence weight by comprehensively considering the intensity of motion change and motion evidence, performs weighted statistics on effective motion trajectory events, and constructs a seedling vitality feature vector; the seedling quality grading unit determines the seedling quality level based on the seedling vitality feature vector and the credibility of image evidence, thereby ensuring the overall reliability of seedling population vitality quality grading in complex aquatic environments.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for grading the quality of aquatic seedlings based on seedling vitality image features according to the present invention.

[0022] Figure 2This is a flowchart illustrating the specific steps of motion constraint determination in a method for grading aquatic seedling quality based on seedling vitality image features, as described in this invention.

[0023] Figure 3 This is a system block diagram of an aquatic seedling quality grading method based on seedling vitality image features according to the present invention. Detailed Implementation

[0024] Please see Figure 1 This invention provides a technical solution: a method for grading the quality of aquatic seedlings based on seedling vitality image features, comprising the following steps: acquiring a continuous frame sequence of seedling images in an aquatic seedling pond and dividing it into multiple local processing regions; extracting image change information from each local processing region in the continuous frame sequence of seedling images and constructing a corresponding observation change state vector; performing dynamic analysis on the local processing regions based on the observation change state vector, determining the foreground processing region, and selecting multiple environmental reference regions from a preset range of the foreground processing region to construct a dynamic environmental reference set; Based on the observed change state vectors corresponding to each environmental reference area in the dynamic environmental reference set, a local environmental disturbance propagation matrix is ​​established. Based on this matrix, environmental disturbance changes and non-environmental changes are separated from the observed change state vectors corresponding to the foreground processing area, resulting in candidate non-environmental change components and calculating their proportion. Candidate non-environmental change components in consecutive seedling image sequences are correlated to construct candidate motion evidence chains. These chains are then processed by motion constraint judgment to determine valid motion trajectory events. The motion evidence set of valid motion trajectory events is extracted, the confidence weight of each event is calculated, and a seedling vitality feature vector is constructed. Finally, the seedling vitality feature vector and the candidate motion evidence chain are subjected to quality assessment processing to determine the seedling quality level.

[0025] Among them, the observed change state vector is a multi-dimensional standardized quantitative expression of the image changes caused by the movement of seedlings and environmental disturbances such as bubbles, water waves and suspended particles in each local processing area. The dynamic environment reference set is used to provide multiple environmental perturbation reference benchmarks for each foreground processing region, and constructs an environment reference state vector through time delay correction and state combination; The local environmental disturbance propagation matrix depicts the mapping relationship of environmental disturbances propagating from each reference region to the foreground processing region, thus distinguishing between environmental contributions and non-environmental motion changes. The proportion of non-environmental change is used to characterize the degree to which non-environmental factors cause change.

[0026] It should be noted that before dividing the image into multiple local processing regions, the sequence of consecutive frame seedling images (which are grayscale images) needs to be preprocessed, such as performing Gaussian filtering on each frame to suppress random noise from the sensor. The corresponding kernel size can be selected as 3×3, and the standard deviation is preferably 0.8. In addition, a pure environmental sequence without seedlings is also collected.

[0027] The steps for dividing the process into multiple local processing regions are as follows: For the preprocessed sequence of consecutive seedling images, the first frame is taken as the reference frame and divided into a rectangular grid (e.g., 64×64 pixels) as the initial local processing region. Within this reference frame, adjacent initial local processing regions are merged. If the angle between the main gradient directions of adjacent initial local processing regions (obtained by calculating the gradient direction histograms of the two regions in the reference frame image, taking the direction angle corresponding to the peak value of each histogram and calculating the difference) is less than a preset threshold (e.g., 30°). Furthermore, if the difference in structural complexity between the two regions (obtained by calculating the statistical characteristics of their gradient magnitudes in the reference frame image and comparing their differences, for example, by taking their respective gradient magnitude variances and calculating their absolute differences) is less than a preset proportion (e.g., 20%), then the adjacent initial local processing regions are merged into a fixed local processing region. The shape and spatial position of these fixed regions remain unchanged across all frames of the entire image sequence. Let the merged region be the first... Each local processing area is .

[0028] Specifically, the steps for constructing the observed change state vector are as follows: The seedling image sequence of consecutive frames is processed by differential processing of adjacent frames to obtain the image change information corresponding to each local processing area, including grayscale change, texture change, gradient structure change, and motion vector, specifically: Gray scale change The mean absolute difference of pixels in the local processing area is used to calculate the texture change. The gradient structure change is obtained by calculating the chi-square distance of the histogram after extracting local binary pattern features from the region. Pick Subtract the normalized cross-correlation similarity between the two frames' gradient magnitude maps; motion vector The region motion vector, including the motion amplitude, can be obtained using optical flow algorithms. and direction The information, among which, Indicates the frame number; Image change information is marked as change feature parameters of each local processing region. The change feature parameters include gray level change, texture change, structural change and motion vector. The above four parameters describe the image changes that occur in the local processing region between adjacent frames from four dimensions: gray level, texture, edge structure and motion. The changing characteristic parameters are standardized to construct the observed changing state vector, i.e.: Direction of movement Use its sine component Sum and cosine components This indicates that, after standardization, a local processing area is defined. In the Observational change state vector of a frame The vector is composed of six components arranged in sequence: the standardized grayscale change, the texture change, the structural change, the motion amplitude, the sinusoidal component of the motion direction, and the cosine component of the motion direction.

[0029] The specific steps for constructing a dynamic environment reference set are as follows: Based on the observed change state vectors of each local processed region in the continuous frame seedling image sequence, the corresponding change activity level is calculated, specifically as follows: For each local processing area Take the sequence of motion amplitude components in the observed state vector within a preset time window (e.g., 2 seconds), and calculate the variance of the sequence. ; Simultaneously, wavelet packet decomposition (e.g., using db4 wavelet, with 3 to 4 decomposition levels) is performed on the motion amplitude component sequence, decomposing the signal into multiple frequency bands and extracting the energy of the activity-varying frequency bands respectively. and environmental low-frequency energy The frequency band of activity changes can be preset according to the seedling type. For example, the jumping and tail wagging frequency of shrimp larvae is usually 3 to 8 Hz, and the tail wagging frequency of fish larvae is 2 to 6 Hz. The low frequency band of the environment is usually the water wave and overall light fluctuation frequency below 1.5 Hz. The degree of change : ; In the formula, It is a small positive number used to prevent the denominator from being zero; the closer its value is to zero, the better. This indicates that the area is more likely to contain active seedlings. Finally, the calculation results for all locally processed areas in the entire map are subjected to min-max normalization to map the degree of change to interval; Local treatment regions whose activity levels exceed a preset threshold (which can be taken as the 95th percentile of the activity level distribution of all regions in a seedless pure environment sequence) are identified as foreground treatment regions. When there is no foreground processing area, the analysis time window can be expanded to recalculate the degree of change activity in order to avoid missing resting seedlings; Obtain multiple candidate regions within a preset range of the foreground processing area, specifically as follows: Using the center of each foreground processing region as the center, search all local processing regions marked as background processing regions as candidate regions within a preset search radius (such as 3 to 5 times the side length of the local processing region); Calculate the regional association feature set between each candidate region and the foreground processing region, including regional distance, difference in grayscale change, and difference in historical observation change status, specifically as follows: The region distance is the Euclidean distance between the center of the candidate region and the center of the foreground processing region. The difference in grayscale change is taken as the absolute difference between the grayscale change (original value before standardization) of the two regions in the current frame; The difference in historical observed state of change is taken as the mean of the Euclidean distance between the observed state of change vectors of the two regions over the past several frames (e.g., 20 frames). Based on the regional association feature set, multiple environmental reference regions are determined from multiple candidate regions, specifically as follows: From the candidate regions, select the candidate regions that simultaneously meet all the following conditions: the distance between regions does not exceed the search radius, the difference in grayscale change is less than the preset grayscale change difference threshold (preferably set to 10 to 20, for example, 15), and the difference in historical observation change status is less than the preset status difference threshold (preferably set to 0.5 to 1.5, for example, 0.8). If the number of candidate regions that meet all the conditions is greater than the preset number (e.g., 3 to 5), then the preset number is selected from the regions closest to furthest. If the number is less than the preset number, then the search radius is increased step by step (e.g., 0.5 to 1 times the side length of the local processing region), and the region is re-selected after each increase until the number of regions that meet the conditions reaches the preset number or the search radius reaches the preset upper limit (e.g., 1.5 times the search radius). If the number is still insufficient, then the actual number is selected. Multiple environmental reference regions are combined to form a dynamic environmental reference set. .

[0030] Specifically, the steps for establishing the local environmental disturbance propagation matrix are as follows: Extract the component sequences from the observed state vectors corresponding to each environmental reference region in the dynamic environmental reference set, and calculate the cross-correlation curves between these components and the corresponding component sequences in the foreground processing region to determine the propagation delay of each reference region, i.e.: For each environmental reference region in the dynamic environmental reference set, extract the observed change state vector sequence (after standardization) within the time window; extract the first component (i.e., the standardized grayscale change) in the observed change state vector of the reference region to form a sequence, and perform cross-correlation analysis with the corresponding component sequence of the foreground processing region to calculate the cross-correlation curve, i.e., the curve of cross-correlation coefficient changing with the number of lag frames; The number of lag frames corresponding to the peak value of the cross-correlation curve is taken as the propagation delay of the j-th environmental reference region relative to the foreground processing region. If the absolute values ​​of all cross-correlation coefficients in the cross-correlation curves do not exceed a preset peak threshold, the 95th percentile of the maximum cross-correlation curve value of all background regions in the pure environment sequence can be taken as this threshold. In this case, it is determined that there is no significant peak value. ; Based on the time delay parameter, time correction is performed on the observed change state vector corresponding to each environmental reference area, specifically as follows: For the environmental reference area at time The observed change state vector, using time delay Perform correction and take its value at time [time]. The observed change state vector is used as the time of the foreground processing region. Alignment reference state vector; The observed change state vectors corresponding to multiple time-corrected environmental reference areas are combined to obtain the environmental reference state vector, i.e.: The observed state vectors of all environmental reference regions, after time delay correction, are concatenated end-to-end to form the environmental reference state vector. It is used to characterize the comprehensive environmental change status of multiple environmental reference areas after time correction; Based on the environmental reference state vector and the observed change state vector corresponding to the foreground processing region, a local environmental disturbance propagation matrix is ​​established, namely: In the historical time window ( Preferred Within each frame, collect the environment reference state vector. and the corresponding foreground processing region observation change state vector ,in, For historical time window The frame number within; Each frame Arrange the columns to form the input matrix, and then... The columns are arranged to form the output matrix, and the local environmental disturbance propagation matrix is ​​obtained by solving the least squares method. To satisfy .

[0031] The specific steps for separating environmental disturbances from non-environmental changes are as follows: The observed change state vector corresponding to the foreground processing region is decomposed into an environmental disturbance change term, a candidate non-environmental change component, and a random noise term, specifically as follows: Time Observation change state vector of the foreground processing region The model is established as the sum of three terms, that is... ; in, This is the local environmental disturbance propagation matrix. The environmental disturbance change term represents the environmental change component calculated jointly by the local environmental disturbance propagation matrix and the reference state vector; These are candidate non-environmental change components, used to represent change components that are difficult to explain by environmental disturbances, including changes caused by seedling activity, bubble transients, and other unmodeled disturbances. This is a random noise term, representing the small random fluctuations introduced during the data acquisition and calculation process; Based on the separation of the environmental reference state vector and the local environmental disturbance propagation matrix in the environmental disturbance change term, candidate non-environmental change components are obtained, specifically as follows: Using the local environmental disturbance propagation matrix and the reference state vector of the current frame Directly calculate the predicted environmental change state value of the current frame. ; Subtracting the predicted value from the observed state change vector yields the candidate non-environmental change components for the current frame. ; Furthermore, sparsity constraints are applied to the candidate non-environmental change components to obtain the candidate non-environmental change state vector, i.e.: By historical time window Solve the following optimization problem internally, refining the candidate non-environmental variation components: ; In the formula, For sparse regularization strength, the preferred range is: to This optimization will reduce small random noise. Incorporating candidate non-environmental change components In the middle, by Norm constraints encourage it to be as sparse as possible; when screening based on the motion evidence chain in the subsequent process, weak and discontinuous changes caused by noise will be naturally eliminated, and the optimization problem can be solved iteratively using the alternating direction multiplier method (ADMM). The candidate non-environmental change components obtained from the final solution As a candidate non-environmental change state vector, denoted as ; The proportion of non-environmental changes is calculated based on this vector. : ; In the formula, Let be a pre-defined minimal positive number, when ( Given a preset minimum positive number, the preferred value is... When ), directly order The range of values ​​for the proportion of non-environmental changes is: The value is used to characterize the degree of change caused by non-environmental factors. The higher the value, the less the change follows the prediction law of environmental disturbance change term. However, a high proportion of non-environmental change may also be caused by model error or random noise, so it is only used as a candidate screening indicator in the future.

[0032] Specifically, the steps for constructing the candidate motion evidence chain are as follows: The candidate non-environmental change state vectors corresponding to each frame in a continuous sequence of seedling images are obtained as follows: For each foreground processing region in each frame of a continuous sequence of seedling images, extract the candidate non-environmental change state vectors obtained through the aforementioned steps. Only the non-environmental change proportion is greater than zero and the motion amplitude component in the candidate non-environmental change state vector exceeds a preset minimum threshold for subsequent association. The preset minimum threshold is set to 0.5 to 2.0, preferably 1.0. Candidate non-environmental change components at different spatial locations in each frame are treated as independent candidate events for subsequent association. Extract candidate non-environmental change state vectors from each frame The corresponding position parameters of the foreground processing region are obtained, and the motion state parameters used for position prediction and association are parsed from this vector, namely: The position parameter is taken as the center coordinate of the foreground processing area. The motion components in the standardized state vector are restored to the actual motion vector: the motion amplitude component is destandardized, that is, multiplied by the historical standard deviation and added to the mean to obtain the actual motion amplitude. The actual motion direction angle is recovered by arctangent operation of the sine and cosine components of the motion direction. This actual motion vector will be used to predict the position of the candidate event in the next frame. The predicted position is the current actual position plus the actual motion vector. Based on the position parameters and motion parameters corresponding to each candidate non-environmentally changed state vector in adjacent frames, the candidate non-environmentally changed state vectors in adjacent frames are associated, specifically as follows: For the Each candidate non-environmental change state vector in the frame, in the th The successor is searched independently within the frame; the search range is all candidate non-environmental change state vectors within a preset radius centered on the predicted position. The candidate with the lowest matching cost is selected as the successor. The matching cost is obtained by weighted summation of three normalized indices: The first term is the Euclidean distance between the predicted location and the center coordinates of the foreground processing region corresponding to the candidate, divided by the search radius; the second term is the absolute value of the angle between the two motion directions, divided by... The third term is the chi-square distance between the local binary pattern histograms of the two image patches and their theoretical maximum value. The ratio (if the histogram is not normalized, it needs to be normalized first so that the distance is within) (Scope); the three weights are preferably equal, and all are taken. ; If the minimum cost still exceeds the preset matching cost threshold, the candidate non-environmental change state vector breaks in this frame and is not included in any evidence chain. The preset matching cost threshold is automatically calibrated by the pure environment sequence: the same candidate change state extraction and inter-frame association operation as above is performed on the pure environment sequence without seedlings. At this time, all the associations obtained are pseudo motion associations such as bubble movement and water wave propagation caused by environmental disturbances. Collect the matching costs generated during the association process for all frames in the pure environment sequence (i.e., the cost between each candidate and the best matching successor in each frame), and then calculate the first value of these costs. The percentile is used as the preset matching cost threshold, which reflects the minimum cost level that pseudo-motion caused only by environmental disturbance can achieve when associated. If the matching cost exceeds this threshold, it indicates that the matching degree of this association is even worse than... The environmental pseudo-motion, this candidate change is likely noise that is difficult to sustain correlation; Since each frame may have multiple candidate events, and each candidate event may match multiple successors, multiple candidate association paths will be formed simultaneously across the entire graph. By connecting the candidate non-environmentally changed state vectors in chronological order, multiple sequences of candidate non-environmentally changed state vectors are formed, specifically as follows: From the Starting from each candidate non-environmental change state vector in the frame, each vector is independently and recursively associated along the time direction until continuous Until a frame cannot find a successor that satisfies the minimum threshold of matching cost, a candidate vector sequence is formed. Sequences starting from different starting positions in each frame, as well as sequences starting from different candidate events in the same frame, are constructed in parallel as independent candidate sequences. Based on the position parameters and motion state parameters corresponding to each vector in the candidate non-environmental change state vector sequence, a candidate motion evidence chain is constructed, specifically as follows: Each successfully associated candidate non-environmental change state vector sequence is denoted as a candidate motion evidence chain composed of multiple candidate events. Each event contains the candidate non-environmental change state vector, position parameters, and motion state parameters of the corresponding frame. During the analysis period, multiple candidate motion evidence chains will be formed in parallel. Each evidence chain will organize the candidate non-environmental changes scattered in each frame and position into a candidate motion sequence with temporal continuity.

[0033] like Figure 2 As shown, the specific steps for motion constraint determination and processing are as follows: Based on the position parameters corresponding to the candidate non-environmentally changed state vectors of adjacent frames in the candidate motion evidence chain, the spatial migration continuity is calculated as follows: For each candidate event corresponding to a candidate non-environmental change state vector in the evidence chain, take the center coordinates of its foreground processing area, calculate the Euclidean distance between its center coordinates and those of adjacent events, and use it as the inter-frame displacement. Repeat this operation for all adjacent frame pairs in the evidence chain to obtain an inter-frame displacement sequence. Based on this displacement sequence, calculate the absolute value of the difference between the displacements of two adjacent frames, which is used as the acceleration amplitude sequence. Take the standard deviation of this acceleration amplitude sequence. ; Spatial migration continuity Calculate using the following formula: ; In the formula, The preset scaling factor has a preferred range of [values]. to The smaller the acceleration fluctuation, The closer This indicates that the smoother the spatial migration, the more consistent it is with the characteristics of continuous seedling movement; Based on the motion state parameters corresponding to the candidate non-environmental change state vectors in adjacent frames of the candidate motion evidence chain, the structural evolution and displacement coordination degree is calculated. Specifically, it is as follows: For each candidate event corresponding to a candidate non-environmental change state vector in the evidence chain, the displacement direction from that candidate event to the next candidate event is calculated using its corresponding actual motion vector, denoted as... ; Simultaneously, the local moving target structure direction of the local processing region where the candidate event is located is extracted: the image gradient of each pixel in this region is used to construct a structure tensor, where each element is the sum of the gradient products of all pixels in the corresponding direction; the structure tensor is subjected to feature decomposition to obtain two eigenvalues, where the direction of the eigenvector corresponding to the larger eigenvalue is the local moving target structure direction of the region, denoted as . ; Based on two eigenvalues ​​of the structure tensor, the degree of anisotropy is calculated, which is defined as the ratio of the difference between the two eigenvalues ​​to the sum of the two eigenvalues. Only when it is greater than a preset threshold (preferably 0.4) is the orientation estimation of the frame considered reliable and it is regarded as a valid frame. Calculate the change in displacement direction between adjacent frames respectively. and structural direction change And normalize the angle differences before calculation; Structural evolution and displacement synergy Calculate using the following formula: ; In the formula, This is the set of frame indices that are valid for the structural orientation of the locally moving target in the evidence chain. This represents the number of valid frames in the set. If the number of valid frames is less than half the length of the evidence chain, then the coordination constraint is directly determined to be unsatisfactory. Pick ; Extract the structural features corresponding to each frame in the candidate motion evidence chain and calculate the structure preservation parameters, specifically: For each candidate event in the candidate evidence chain, a square image patch with a side length equal to the side length of the smallest bounding square of the region is taken as the center of the geometric center of the foreground processing region where the event is located, and the gradient direction histogram descriptor of the image patch is extracted; the structure preservation parameter is taken as the mean of the cosine similarity of the descriptors of adjacent frames, which measures the degree of preservation of the structure of the local region during the motion process. The number of consecutive frames corresponding to candidate motion evidence chains is counted, and the duration parameter is determined. Specifically, the duration parameter is taken as the sum of the actual length of the evidence chain and the minimum effective chain length (preferred). to The ratio of frames) and The smaller of the two values; Based on spatial migration continuity, structural evolution and displacement synergy, structural retention parameters, and duration parameters, motion constraint determination is performed on each candidate motion evidence chain, specifically as follows: Determine the appropriate thresholds for spatial migration continuity, structural evolution and displacement synergy, structural retention parameters, and duration parameters, and optimize each threshold. to The threshold for spatial migration continuity is set to 0.6, the threshold for structural evolution and displacement synergy is set to 0.55, the threshold for structural retention parameter is set to 0.65, and the threshold for duration parameter is set to 0.5. Candidate motion evidence chains that simultaneously satisfy all four constraints are determined by motion constraints. The candidate motion evidence chain determined by motion constraints is taken as a valid motion trajectory event, specifically as follows: Each candidate motion evidence chain that passes the motion constraint determination constitutes a valid motion trajectory event; otherwise, it is discarded. All candidate non-environmental change state vectors contained in the evidence chain that passed the judgment are included in the set of valid motion trajectory events, and their corresponding spatial migration continuity, structural evolution and displacement coordination, structural retention parameters and duration parameters are included in the motion evidence set of the valid motion trajectory event.

[0034] Specifically, the steps for constructing the seedling vitality feature vector are as follows: Based on each valid motion trajectory event The candidate non-environmental change state vector is used to determine the intensity of motion change, that is, the candidate non-environmental change state vector. norm As the intensity of its motion change; Extract the set of motion evidence for valid motion trajectory events, and calculate the corresponding confidence weights based on the intensity of motion changes. Extract the inter-transition continuity, structural evolution and displacement co-occurrence, structural preservation parameters, and duration parameters from the motion evidence set of this event, and assign their confidence weights. Calculate using the following formula: ; in, ( These correspond to spatial migration continuity, structural preservation parameters, structural evolution and displacement coordination, and duration parameters, respectively. Assigning weights to each score, and selecting four equally weighted factors, i.e., each factor is... ; It is the minimum value among the four parameters; Confidence weights are used to weight and statistically analyze the motion state parameters corresponding to valid motion trajectory events. Specifically, for all valid motion trajectory events, confidence weights are applied... As a weighting coefficient, the motion state parameters are weighted and accumulated; Based on the weighted statistical motion state parameters and combined with the candidate motion evidence chain corresponding to the effective motion trajectory events, a seedling vitality feature vector is constructed, which is as follows: The seedling vitality feature vector includes the following dimensions: Exercise intensity is calculated as the average of confidence-weighted exercise amplitudes. Activity coverage is the time average of the proportion of high-confidence-weighted regions to the total region within the analysis period. The duration of the motion is the median length of the evidence chain corresponding to all valid motion trajectory events in the candidate motion evidence chain. Spatial activity balance is determined by the normalized entropy of the distribution of high-confidence weighted events in the pool space grid, which uses the same rectangular grid as the local processing area. Directional change characteristics are used to determine the proportion of turning events in the evidence chain. This is done when the absolute value of the angle between the motion directions of adjacent frames in the evidence chain exceeds a preset angle threshold (e.g., ...). The event is marked as a turning event, and the proportion of turning events is the ratio of the number of turning events to the total number of adjacent frame intervals in the evidence chain; Furthermore, for each dimension, only valid motion trajectory events with confidence weights exceeding a preset activity threshold are counted. The activity threshold is automatically determined by the pure environment sequence. Specifically, the pure environment sequence is processed using the steps described above, the confidence weights of all obtained pseudo-motion events in the environment are calculated, and their distribution is taken. The quantile is used as the preset activity threshold.

[0035] The specific steps for determining the quality grade of seedlings are as follows: The seedling vitality feature vector is comprehensively processed to obtain the seedling comprehensive vitality value, which is as follows: Normalize each dimension of the seedling vitality feature vector. When normalizing, the global maximum and minimum values ​​based on historical batch data statistics can be used, or the extreme value of the current batch can be used as a temporary range for the first use. The overall vitality value is the weighted sum of the features of each normalized dimension, and the sum of the weights is... The corresponding weights are set as follows: exercise intensity weight 0.30, activity coverage weight 0.20, typical duration of exercise weight 0.25, spatial activity balance weight 0.15, and directional change characteristics weight 0.10. The higher the comprehensive vitality value, the stronger the overall activity level of the seedling population. Based on the set of motion evidence in the candidate motion evidence chain corresponding to valid motion trajectory events (with confidence weights exceeding a preset activity threshold), the activity stability value is calculated. Specifically, the entire analysis cycle is divided into preset durations (e.g., each...). The time interval is divided into multiple equal-length sub-windows, and the local comprehensive vitality value in each sub-window is calculated separately. For each sub-window, based only on the valid motion trajectory events that appear within that window (with confidence weights exceeding the preset activity threshold), the local seedling vitality feature vector (with consistent dimensions and definitions) is extracted using the same steps as the overall model. Each dimension is then normalized (the normalization reference range can be the global extreme value of the entire sequence), and a weighted sum is obtained to obtain the local comprehensive vitality value. Based on the mean and standard deviation of the local comprehensive vitality values ​​of each sub-window, the vitality stability value is determined as follows: when the mean of the local comprehensive vitality value is When the standard deviation is also The vitality stability value is Otherwise ; When the mean is greater than At that time, calculate Subtract the ratio of the standard deviation to the mean, and compare the result with... The larger value between these two ranges is used as the vitality stability value, ensuring that the value range is within a certain range. ; The closer the vitality stability value is to This indicates that the seedling vigor is more evenly distributed over the analysis period, and there are no drastic fluctuations. The credibility of the image evidence is determined, and the seedling quality grade is determined by classifying the seedlings according to their overall vitality value and vitality stability value. Specifically: The credibility of image evidence is a positive indicator that integrates two aspects: the quality of environmental separation and the quality of the motion evidence chain. The effective reference area ratio refers to the proportion of environmental reference areas that exist within a preset range of the foreground processing area. This indicator reflects the reliability of anomaly separation. Fit stability refers to the propagation matrix estimated for each foreground processing region in each frame (or every few frames) within the analysis period. The degree of fit of the environmental model is determined based on the deviation between the predicted results and the actual observed changes, and the median of all effective degree of fit values ​​is taken as the index. The percentage of abnormal environmental events is determined by the percentage of non-environmental changes exceeding a preset threshold during the analysis period (e.g., ...). The proportion of events to all events, where all events refer to the observed changes in the state vectors of all foreground processing regions in each frame within the analysis period. The norm is greater than a certain minimum threshold (e.g.) The total number of events; The total number of valid motion events refers to the total number of valid motion trajectory events determined by motion constraints within the analysis period, whose confidence weight exceeds a preset activity threshold. If the total number of valid motion events is... If so, the credibility of the image evidence is directly determined to be insufficient; The proportion of high-confidence changes, i.e., in the set of valid motion trajectory events, the confidence weight exceeds a preset threshold (such as...). The proportion of events, which reflects the overall quality of the chain of evidence for the movement; After normalizing the above five indicators, the credibility of the image evidence is calculated by weighting. The weights of each indicator can be set according to the importance of each indicator to the grading results in the actual application scenario, and the equal weighting method is preferred. If the proportion of the normalized effective reference area is lower than the first absolute lower limit (preferably set to 0.05 to 0.10, for example 0.05), or the total number of normalized effective motion events is lower than the second absolute lower limit (preferably set to 0.01 to 0.05, for example 0.02), then the credibility of the image evidence is directly determined to be unsatisfactory, and no further weighted calculation is required.

[0036] Hierarchical decision-making based on the credibility of image evidence: If the credibility of the image evidence is lower than the preset credibility threshold (preferred) This indicates that the current image sequence is severely disturbed by the environment or lacks effective motion information, making it difficult to draw reliable grading conclusions. In this case, the seedlings of this batch are directly classified as "awaiting re-inspection" and no grading is performed to avoid misjudging poor image quality as poor seedling quality. If the credibility of the image evidence is not lower than the preset credibility threshold, it indicates that the image evidence is sufficient, and the level will continue to be classified according to the comprehensive vitality value and vitality stability. The division rules are as follows: When both the overall vitality value and vitality stability are higher than the corresponding high threshold, it is judged as first-level seedling quality, indicating that the seedling population has strong and continuous vitality. When both the overall vitality value and vitality stability are higher than the corresponding low threshold but not higher than the corresponding high threshold, the seedling quality is judged as Level II, indicating that the seedling population has moderate vitality. The remaining cases that meet the image evidence credibility threshold of not less than the preset credibility threshold are judged as Level 3 seedling quality, indicating that the seedling population has weak or unstable vitality. The preferred range for the high threshold of the comprehensive vitality value is 0.60 to 0.80, for example, 0.70. The preferred range for the low threshold value of the comprehensive vitality value is 0.35 to 0.55, for example, set to 0.45; The preferred range for the high threshold of vitality stability is 0.60 to 0.80, for example, set to 0.70; The preferred range for the low threshold value of vitality stability is 0.35 to 0.55, for example, set to 0.45.

[0037] Please see Figure 3This invention provides a technical solution: a system for grading the quality of aquatic seedlings based on seedling vitality image features, comprising: a data acquisition unit for acquiring a continuous frame sequence of seedling images in an aquatic seedling pond and dividing it into multiple local processing regions; an observation change construction unit for extracting image change information from each local processing region and constructing a corresponding observation change state vector; an environmental reference construction unit for dynamically analyzing the local processing regions based on the observation change state vector, determining the foreground processing region, and selecting multiple environmental reference regions from a preset range of the foreground processing region to construct a dynamic environmental reference set; and a non-environmental change separation unit for establishing a local environmental disturbance propagation matrix based on the observation change state vectors corresponding to each environmental reference region in the dynamic environmental reference set, and separating environmental disturbance changes from non-environmental changes to obtain candidate non-environmental change components. The motion evidence chain construction unit is used to associate candidate non-environmental change components, construct candidate motion evidence chains, and determine valid motion trajectory events through motion constraint judgment processing; the vitality feature extraction unit is used to extract the motion evidence set of valid motion trajectory events, calculate confidence weights, and construct seedling vitality feature vectors; the seedling quality grading unit is used to perform quality assessment processing on the seedling vitality feature vectors and candidate motion evidence chains to determine the seedling quality level.

[0038] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0039] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for grading the quality of aquatic seedlings based on seedling vitality image features, characterized in that, Includes the following steps: Acquire a continuous frame sequence of seedling images from an aquatic seedling nursery and divide it into multiple local processing regions; Image change information is extracted from each local processing area to construct the corresponding observed change state vector; Dynamic analysis of the local processing area is performed based on the observed change state vector to determine the foreground processing area. Multiple environmental reference areas are selected from the preset range of the foreground processing area to construct a dynamic environmental reference set. Based on the observed change state vectors corresponding to each environmental reference region in the dynamic environmental reference set, a local environmental disturbance propagation matrix is ​​established, and environmental disturbance changes and non-environmental changes are separated to obtain candidate non-environmental change components. The candidate non-environmental change components are correlated to construct a candidate motion evidence chain, and after motion constraint judgment processing, the valid motion trajectory events are determined. Extract the motion evidence set of valid motion trajectory events, calculate the confidence weights, and construct the seedling vitality feature vector; The seedling vitality feature vector and candidate motion evidence chain are subjected to quality assessment processing to determine the seedling quality level.

2. The aquatic seedling quality grading method based on seedling vitality image features according to claim 1, characterized in that, The specific steps for constructing the observed state change vector are as follows: Differential processing is performed on adjacent frames of seedling images in a continuous sequence of seedling images to obtain image change information corresponding to each local processing area; Image change information is labeled as change feature parameters; The changing characteristic parameters are standardized to construct the observed changing state vector.

3. The aquatic seedling quality grading method based on seedling vitality image features according to claim 2, characterized in that, The specific steps for constructing a dynamic environment reference set are as follows: Calculate the corresponding activity level of change based on the observed change state vector of each local processing area; Local processing areas where the activity level of change exceeds a preset threshold are identified as foreground processing areas; Obtain multiple candidate regions within a preset range of the foreground processing area; Calculate the regional association feature set between each candidate region and the foreground processing region, including regional distance, difference in grayscale change, and difference in historical observation change status; Based on the regional association feature set, multiple environmental reference regions are determined from multiple candidate regions; Multiple environmental reference regions are combined to form a dynamic environmental reference set.

4. The aquatic seedling quality grading method based on seedling vitality image features according to claim 1, characterized in that, The specific steps for establishing the local environmental disturbance propagation matrix are as follows: Extract the component sequence from the observed state change vector corresponding to each environmental reference area, and calculate the cross-correlation curve to determine the propagation delay of each reference area; Based on the time delay parameter, time correction is performed on the observed change state vector; The time-corrected observed state vectors are combined to obtain the environmental reference state vector; A local environmental disturbance propagation matrix is ​​established based on the environmental reference state vector and the observed change state vector corresponding to the foreground processing area.

5. The aquatic seedling quality grading method based on seedling vitality image features according to claim 4, characterized in that, The specific steps for separating environmental disturbances from non-environmental changes are as follows: The observed change state vector corresponding to the foreground processing area is decomposed into environmental disturbance change term and candidate non-environmental change component. Based on the separation of the environmental reference state vector and the local environmental disturbance propagation matrix in the environmental disturbance change term, candidate non-environmental change components are obtained. Furthermore, sparsity constraints are imposed on the candidate non-environmental change components to obtain the candidate non-environmental change state vector.

6. The aquatic seedling quality grading method based on seedling vitality image features according to claim 1, characterized in that, The specific steps for constructing a candidate motion evidence chain are as follows: Obtain the candidate non-environmental change state vectors corresponding to each frame in a continuous frame seedling image sequence; Extract the position parameters of the foreground processing region corresponding to the candidate non-environmental change state vectors in each frame, and parse out the motion state parameters; Based on the position parameters and motion state parameters corresponding to each candidate non-environmental change state vector in adjacent frames, the candidate non-environmental change state vectors in adjacent frames are associated. Connect the candidate non-environmental change state vectors after association in chronological order to form a sequence of candidate non-environmental change state vectors. Based on the position parameters and motion state parameters corresponding to each vector in the candidate non-environmental change state vector sequence, a candidate motion evidence chain is constructed.

7. The aquatic seedling quality grading method based on seedling vitality image features according to claim 6, characterized in that, The specific steps for motion constraint determination and processing are as follows: Calculate the spatial migration continuity based on the position parameters of adjacent frames in the candidate motion evidence chain; Calculate the structural evolution and displacement coherence degree based on the motion state parameters in the candidate motion evidence chain; Extract the structural features corresponding to each frame in the candidate motion evidence chain and calculate the structure preservation parameters; The number of consecutive frames corresponding to candidate motion evidence chains is counted, and the duration parameter is determined. Based on spatial migration continuity, structural evolution and displacement synergy, structural retention parameters, and duration parameters, motion constraint determination is performed on candidate motion evidence chains. Candidate motion evidence chains determined by motion constraints are considered as valid motion trajectory events.

8. The aquatic seedling quality grading method based on seedling vitality image features according to claim 7, characterized in that, The specific steps for constructing the seedling vitality feature vector are as follows: The intensity of motion change is determined based on the candidate non-environmental change state vectors of effective motion trajectory events; Extract the set of motion evidence for valid motion trajectory events, and calculate the corresponding confidence weights based on the intensity of motion changes; Confidence weights are used to perform weighted statistics on the motion state parameters corresponding to valid motion trajectory events. Based on the weighted statistical motion state parameters and combined with the candidate motion evidence chain corresponding to the effective motion trajectory events, a seedling vitality feature vector is constructed.

9. The aquatic seedling quality grading method based on seedling vitality image features according to claim 8, characterized in that, The specific steps for determining the quality grade of seedlings are as follows: The seedling vitality feature vector is comprehensively processed to obtain the seedling comprehensive vitality value; Read the set of motion evidence from the candidate motion evidence chain corresponding to the valid motion trajectory event, and calculate the vitality stability value; Determine the credibility of image evidence, and classify seedling grades based on seedling comprehensive vitality value and vitality stability value to determine seedling quality grade.

10. A system for grading the quality of aquatic seedling cultivation based on seedling vigor image features, employing the aquatic seedling cultivation quality grading method based on seedling vigor image features as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire a continuous frame sequence of seedling images in the aquatic seedling pond and divide it into multiple local processing areas; The observation change construction unit is used to construct the corresponding observation change state vector; Environment reference building blocks are used to build dynamic environment reference sets; The non-environmental change separation unit is used to establish the local environmental disturbance propagation matrix and separate environmental disturbance changes from non-environmental changes to obtain candidate non-environmental change components. The motion evidence chain construction unit is used to construct candidate motion evidence chains and, after motion constraint judgment processing, determine the valid motion trajectory events. The vitality feature extraction unit is used to calculate confidence weights and construct seedling vitality feature vectors. The seedling quality grading unit is used to determine the seedling quality grade.

Images