Machine learning based early warning monitoring system for diseases in leopard grouper breeding

Abstract

The application discloses a leopard-bristle-jaw grouper breeding disease early warning monitoring system based on machine learning and concretely relates to the field of aquatic product breeding image monitoring and early warning, and comprises an image modeling module, which is used for acquiring a continuously collected breeding image sequence, performing identification, correction and body surface unfolding on a fish body, segmenting according to a head and tail, and generating a body surface grid map; and an abnormal stringing module, which is used for inputting the body surface grid map into a machine learning identifier trained by a sick fish sample, performing abnormal marking on each grid, and merging abnormal grids into abnormal blocks; the application solves the problem that mechanical damage and continuously evolving pathological abnormalities are difficult to distinguish in the prior art, and the early disease early warning reliability is insufficient, by performing chain tracking and evolution determination on the position continuation, boundary expansion and direction change of the fish body surface abnormalities in the continuously monitored images.

Description

Technical Field

[0001] This invention relates to the field of aquaculture image monitoring and early warning technology, and more specifically, to a disease early warning and monitoring system for leopard-gill spiny perch based on machine learning. Background Technology

[0002] In the early warning and monitoring of diseases in leopard gill spiny perch farming, the existing processing approach mainly focuses on whether suspected diseased individuals can be identified from farming images as early as possible and an early warning result can be formed. The common method is to process the collected fish images and use machine learning to detect, segment, classify or score abnormalities on the body surface, and make a comprehensive judgment by combining the identification results over a continuous period of time. Taking continuous monitoring under high-density recirculating aquaculture conditions as an example, the camera device needs to track abnormal appearances such as scale loss, red spots, superficial damage, and localized discoloration under conditions of frequent fish rubbing against the pond walls or netting, competing for food, overlapping and obstruction of the fish, continuous water disturbance, and constantly changing shooting angles. A persistent limitation in this scenario is that mechanical damage and early pathological abnormalities often exhibit similar static appearances in their initial stages. Furthermore, abnormal areas are characterized by incomplete boundaries, short visibility time, and unstable exposure locations. Under these conditions, relying solely on the abnormal appearance and its short duration is insufficient. Judging based on recurring situations can easily lead to premature inclusion of minor injuries, short-term adhesions, or occasional abnormalities in disease warnings, or situations where results are only output after the abnormality range has continued to expand in order to reduce false alarms. The directly observable phenomenon is that the abnormal indications of some individuals recur in a short period of time and then disappear on their own, while some individuals who subsequently develop persistent lesions have not been identified in time when they showed abnormal signs multiple times in the early stages. The root cause is that the current processing method focuses on whether the abnormality is visible at a certain moment, but lacks the ability to distinguish whether the abnormality continues to evolve along the direction of the disease. Therefore, the technical problem to be solved by this application is: how to distinguish between mechanical damage and continuously evolving pathological abnormalities during the image monitoring of leopard gill sea bass farming, and to achieve reliable disease early warning when the abnormality is in its early stage. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a disease early warning and monitoring system for leopard gill sea bass aquaculture based on machine learning. This system solves the problems mentioned in the background art by performing chain-like tracking and evolution determination of the position continuation, boundary expansion, and directional changes of abnormalities on the fish body surface in continuous monitoring images.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture, comprising: The image modeling module is used to acquire a series of continuously collected aquaculture images, perform recognition, correction and body surface unfolding on the fish, segment by head and tail and row by back and belly, and generate a body surface mesh map; The anomaly concatenation module is used to input the body surface grid map into a machine learning recognizer trained on diseased fish samples, perform anomaly marking on each grid, merge the abnormal grids into anomaly blocks, and concatenate the abnormal blocks with the same or adjacent positions at adjacent time points and the same head-to-tail order into anomaly chains, and output an anomaly chain set. The extended statistics module is used to extract the edge grid, internal grid and external neighbor grid of the abnormal block in the abnormal chain at the previous time step, and then count the number of newly added grids, the number of retained grids and the number of exited grids at the next time step to generate extended records; The evolution determination module is used to compare the number of newly added grids, the number of retained grids, the number of exited grids, and the expansion direction of the same anomalous chain in consecutive time periods. The anomalous chain that has retained grids and newly added grids in the same direction in two consecutive adjacent time periods without direction reversal is identified as an evolutionary anomalous chain, and the evolutionary anomalous chain is output. The early warning output module is used to generate disease early warning results for each evolutionary anomaly in the evolutionary anomaly set according to fish body identification, initial time, duration, coverage grid range and expansion direction, and output the corresponding disease early warning information for each individual leopard gill spiny perch.

[0005] In a preferred embodiment, the image modeling module includes: For each time-phase image in the aquaculture image sequence, perform leopard gill spiny perch body recognition and extract the fish body outline, and output the outline image of each fish body; Based on the outline image, the head position, tail position, dorsal boundary and ventral boundary of each fish are located, and the fish orientation is corrected according to the direction from head to tail, and the corrected image of each fish is output.

[0006] In a preferred embodiment, the image modeling module further includes: Based on the corrected image, the surface area between the dorsal and ventral boundaries of each fish body is unfolded in the head-to-tail direction, and a two-dimensional unfolded map corresponding to the head-to-tail position and the dorsal and ventral position is established, and the unfolded map of the body surface of each fish body is output. The body surface unfolding diagram is segmented along the head-to-tail direction and then divided into rows along the dorsal-ventral direction. The area formed by the intersection of each segment and each row is recorded as the body surface grid. The corresponding body surface grid diagram of each fish at each time point is output.

[0007] In a preferred embodiment, the abnormal connection module includes: Each grid in the body surface grid map is input into a machine learning recognizer to obtain anomaly markers for each grid. The grid marker results are written according to the fish body identification and the collection time, and the anomaly grid map is output. For abnormal grids in the abnormal grid diagram, merge the abnormal grids that are adjacent to the boundary into the same abnormal block, and record the head and tail grid numbers and back and belly grid numbers contained in each abnormal block, and output the abnormal block set.

[0008] In a preferred embodiment, the abnormal connection module further includes: In adjacent time intervals, the head and tail grid numbers and back and belly grid numbers of the abnormal block in the previous time interval are compared with the abnormal block in the next time interval one by one. The abnormal blocks that contain the same grid, head and tail adjacent grids, or back and belly adjacent grids are paired and the concatenated pairs are output. Based on the order of the head and tail grid numbers of the abnormal blocks in the concatenation pair, concatenation pairs with consistent head and tail numbers are retained and concatenated sequentially according to the acquisition time, and the abnormal chain set is output.

[0009] In a preferred embodiment, the extended statistics module includes: For the previous time step of the abnormal block in the abnormal chain, check the four-sided adjacency relationship of each grid in the abnormal block one by one. Grids that are connected to non-abnormal grids on at least one side are recorded as edge grids, and the remaining grids are recorded as internal grids. Record the outer direction of each edge grid according to the head direction, tail direction, back direction and ventral direction respectively, and output the edge grid set, internal grid set and edge direction table. Based on the edge direction table, extract one layer of adjacent meshes along the corresponding outer direction for each edge mesh as the outer neighbor mesh, and write direction labels for the head-to-outer neighbor mesh, tail-to-outer neighbor mesh, back-to-outer neighbor mesh, and ventral-to-outer neighbor mesh respectively, and output the outer neighbor mesh set.

[0010] In a preferred embodiment, the extended statistics module further includes: Each grid in the abnormal block at the next time step is compared with the internal grid set, the edge grid set, and the external neighbor grid set one by one. Grids that fall into the internal grid set or the edge grid set are recorded as retained grids. Grids that fall into the external neighbor grid set are recorded as newly added grids in the corresponding direction according to the direction identifier. Output the number of retained grids and the number of newly added grids in each direction. Each grid in the anomalous block of the previous time step is compared with the anomalous block of the next time step one by one. Grids that do not match the anomalous block of the next time step are marked as exited grids. Then, an expansion record containing the main expansion direction and the order of directions is generated according to the number of grids retained, the number of grids added in each direction, and the number of exited grids.

[0011] In a preferred embodiment, the evolution determination module includes: For the same anomaly chain, extract the number of newly added grids, the number of retained grids, the number of exited grids, and the direction of expansion at each time point in the order of collection time. Form a comparison time period with two adjacent time points, and write the number of newly added grids, the number of retained grids, the number of exited grids, and the direction of expansion at the previous and subsequent time points into each comparison time period, and output the time period table; For each comparison period in the time period table, perform retention and direction determination in parallel. The comparison period with a greater than zero number of retained grids in the later time is recorded as the retention period, otherwise it is recorded as the disconnection period. The comparison period with a greater than zero number of newly added grids in the later time and the same expansion direction as the previous time is recorded as the same direction period. The comparison period with a greater than zero number of newly added grids in the later time and the different expansion direction from the previous time is recorded as the turning period. The comparison period with zero number of newly added grids in the later time is recorded as the stop expansion period. Output the time period mark table.

[0012] In a preferred embodiment, the evolution determination module further includes: For two consecutive comparison periods in the same abnormal chain, perform cross-determination. If both the previous and subsequent comparison periods are reserved periods, and both are in the same direction, write a persistence flag. If the previous comparison period is in the same direction and the subsequent comparison period is a turning period, write a bounce flag when the expansion direction of the subsequent comparison period is opposite to that of the previous comparison period; otherwise, write a transition flag. If either the previous or subsequent comparison period is a disconnected period or a stopped expansion period, write an abort flag. For anomalous chains with persistent or transitional markers, a merge determination is performed. If the number of exited grids in the later time step is not greater than the number of exited grids in the previous time step, and the sum of the number of newly added grids in the previous time step and the number of newly added grids in the later time step is greater than the number of exited grids in the later time step, the corresponding anomalous chain is recorded as a candidate evolutionary anomalous chain. Otherwise, the corresponding anomalous chain is written to the check list and the corresponding comparison period is returned to re-extract the expansion direction. Anomalous chains with bounce markers are written to the non-evolutionary list, and anomalous chains with termination markers are written to the check list. For candidate evolutionary anomalies, continue to check subsequent comparison periods in chronological order. If there is a retained period and no bounce marker, the corresponding anomaly chain is identified as an evolutionary anomaly and written into the evolutionary anomaly set before being output. Otherwise, the corresponding anomaly chain is excluded from the evolutionary anomaly set and written into the check table or non-evolutionary table.

[0013] In a preferred embodiment, the early warning output module includes: The evolutionary anomalies in the set of evolutionary anomalies are merged according to the fish body identifier. For each evolutionary anomaly of the same fish body, the initial time, duration, coverage grid range and expansion direction are extracted. The anomalies are sorted and merged according to the order of the initial time, the duration, the size of the coverage grid range and whether the expansion direction is consistent, and the warning records of each fish body are generated. The warning intensity is calculated for each fish in the warning record according to the coverage grid range and duration. The disease warning result for the corresponding fish is generated by combining the expansion direction and the first time. The disease warning information for each individual leopard gill spiny perch is output.

[0014] The technical effects and advantages of this invention are as follows: 1. By connecting abnormal chains in units of body surface grids and comparing the continuous changes of newly added grids, retained grids and removed grids, it helps to distinguish mechanical damage from continuously evolving pathological abnormalities, thereby relatively improving the reliability of early disease warning. 2. By first unifying the head and tail direction and the dorsal and ventral direction of the fish body, and then unfolding the fish body surface and dividing it into fixed-number body surface grids, the same body surface position of the same fish body at different collection times can be kept consistent, thereby reducing the position mismatch caused by posture changes. 3. By merging the anomalous meshes that are connected at the same time into anomaly blocks, and by connecting the anomalous blocks that are consecutive in position and have the same head and tail order in adjacent time periods into anomaly chains, discrete anomalous ... 4. By dividing the anomaly chain into edge meshes, internal meshes, and external neighbor meshes, and separately counting the retained meshes, newly added meshes, and exiting meshes, the continuation, expansion, and contraction of anomalies can be decomposed and quantified, thereby improving the executability of determining the expansion direction; 5. By performing retention determination, direction determination, crossover determination, and convergence determination on the comparison period, and processing rebound, termination, and pending state separately, short-term fluctuations and directional swings can be suppressed, thereby improving the accuracy of evolution anomaly identification. 6. By sorting and merging evolutionary anomalies according to fish body identifiers and combining the coverage grid range and duration to generate early warning intensity, scattered anomalies can be clustered into early warning records corresponding to the fish bodies, thereby improving the readability of disease early warning information and the targeting of response. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides a machine learning-based early warning and monitoring system for diseases in aquaculture of leopard-gill perch, comprising: The image modeling module is used to acquire a series of continuously collected aquaculture images, perform recognition, correction and body surface unfolding on the fish, segment by head and tail and row by back and belly, and generate a body surface mesh map; In this embodiment, the image modeling module is used to convert continuously acquired aquaculture images into a unified body surface grid map that can be directly called for subsequent anomaly identification. During processing, the fish body object of a single leopard gill spiny perch is first identified in the images at each acquisition time. Then, the head-tail direction and dorsal-ventral direction of the fish body are unified. Subsequently, the surface of the fish body is mapped into a two-dimensional region and divided into a body surface grid according to a fixed numbering rule. After this processing, the same body surface position of the same fish body at different acquisition times can fall into the same grid coordinates. Subsequent anomaly extraction, anomaly block concatenation, and extended statistics are all based on the same position aperture, thereby avoiding position mismatch caused by fish body turning, partial occlusion, and changes in imaging position. The specific steps are as follows: For each time-stamped image in the aquaculture image sequence, fish body regions are extracted frame by frame. Candidate regions with closed contours, complete areas, and not truncated by image boundaries are retained in the current image as valid fish body regions. The outermost closed boundary of each valid fish body region is extracted and recorded as the outer contour of the fish body. The fish body identifier and the acquisition time are written into the contour record along with the outer contour. Among them, the fish body identifiers are preferentially used to retain the existing identifiers that are close in position, have continuous area changes, and move in the same direction in the previous acquisition time. Fish bodies that do not match the existing identifiers generate new identifiers, thereby obtaining the contour images of each fish body. By first establishing the correspondence between the fish body identifiers and the acquisition time, the subsequent calculations at each position can be limited to the same fish body object, avoiding the misuse of contours between different fish bodies. Based on the contour image, the contour center of a single fish is calculated, and the two ends with the largest distance on the outer contour are determined as the endpoints of the major axis. Combining the changes in the width of the fish's shape, the wider end with the front end tapering is marked as the head position, and the narrower end with the tail edge tapering is marked as the tail position. The head and tail positions are connected to form the head-tail axis. The contour boundaries on both sides of the axis are then separated based on the head-tail axis, and the boundary on the upper side is marked as the dorsal boundary, and the boundary on the lower side is marked as the ventral boundary. If the contour is missing at the current moment, the head-tail direction and boundary correspondence of the same fish at the previous acquisition moment are used for correction. After the positioning is completed, the fish is rotated to a unified direction with the head facing one side and the tail facing the other side, generating the corresponding corrected image for each fish. By unifying the head-tail direction and the dorsal-ventral direction, the directional meaning of the subsequent unfolding results can be kept consistent at different acquisition moments. Based on the corrected image, the head-to-tail axis, from the head position to the tail position, is used as the main axis for unfolding. Multiple cross-sectional positions are selected sequentially at equal intervals along the head-to-tail axis, and samples are continuously taken from the dorsal boundary to the ventral boundary at each cross-sectional position. The sampling results are then written into the corresponding positions of the unfolded image in the order from dorsal to ventral. In this way, the front-to-back order in the head-to-tail direction and the up-to-down order in the dorsal-ventral direction remain unchanged in the unfolded image. For blank areas caused by local reflection, occlusion, or missing parts, adjacent effective sampling positions are used to fill in the blank areas, resulting in the unfolded image of the body surface of each fish. After unfolding, the original fish surface is transformed into a two-dimensional region with fixed head-to-tail and dorsal-ventral positions, providing a unified benchmark for subsequent position-by-position comparisons. After the body surface unfolding map is formed, a fixed number of segments are divided along the head-to-tail direction, and a fixed number of rows are divided along the dorsal-ventral direction. The number of segments and rows is kept consistent with the input granularity of the subsequent recognizer during the training phase. Each segment and each row intersect to form a body surface grid, and the grids are numbered using a combination of head-to-tail grid numbers and dorsal-ventral grid numbers. The head-to-tail grid numbers increase sequentially from head to tail, and the dorsal-ventral grid numbers increase sequentially from dorsal to ventral. This generates a body surface grid map corresponding to each fish body at each collection time. By using a unified numbering rule, the same body surface position of the same fish body at different collection times can be directly mapped to the same set of grid numbers, thus providing a stable spatial index for subsequent abnormal grid extraction and abnormal chain concatenation. Through the above processing, the original aquaculture images are uniformly converted into body surface grid maps with fish body identification, collection time, head and tail direction, dorsal and ventral direction and grid coordinates. Subsequent anomaly identification no longer depends on the instantaneous posture of the fish in the original image, but is based on continuous comparison of the same fish body surface area based on a uniform position caliber, thereby reducing the impact of position drift and object confusion on subsequent judgment. In practical applications: The acquisition device continuously acquires images of the aquaculture pond according to a set cycle. The system first extracts the outer contour of a specific leopard gill sea bass, then locates the head, tail, dorsal and ventral boundaries, and generates a body surface unfolding map after orientation correction. The body surface grid map is then formed according to a fixed segmentation and rowing rule. When the fish re-enters the image at a subsequent acquisition time, the system continues to generate a new body surface grid map using the same fish identifier and the same grid numbering rule. This ensures that the same body surface position of the same fish can be directly correlated at consecutive acquisition times, providing a consistent data foundation for subsequent anomaly identification and disease early warning.

[0018] The anomaly concatenation module is used to input the body surface grid map into a machine learning recognizer trained on diseased fish samples, perform anomaly marking on each grid, merge the abnormal grids into anomaly blocks, and concatenate the abnormal blocks with the same or adjacent positions at adjacent time points and the same head-to-tail order into anomaly chains, and output an anomaly chain set. In this embodiment, the anomaly chaining module is used to form a continuously trackable anomaly chain based on the body surface mesh map. During processing, firstly, anomaly marking is performed on the body surface mesh of each fish at each collection time to obtain an anomaly mesh map with fish identification, collection time, and mesh coordinates. Then, anomaly meshes with adjacent boundaries at the same collection time are merged into anomaly blocks. Subsequently, anomaly blocks are paired one by one between adjacent collection times to determine the correspondence between the same body surface anomaly in the preceding and following times. Finally, the pairing results that meet the requirements of continuous position and consistent head and tail order are sequentially chained into an anomaly chain according to the collection time order. After this processing, the local anomalies that were originally scattered in each collection time can be organized into anomaly objects with continuous temporal relationship, which can be directly called for subsequent extended statistics and evolution determination. The implementation process includes the following steps: First, each grid in the body surface grid map is input into a machine learning recognizer. The recognizer outputs an abnormal category score and a normal category score for each grid. Grids with an abnormal category score greater than the normal category score are marked as abnormal grids, and grids with a normal category score not less than the abnormal category score are marked as normal grids. Then, the grid marking results are written according to the fish body identification, collection time, head and tail grid number, and dorsal and ventral grid number to form an abnormal grid map corresponding to each fish body and each collection time. After this processing, each abnormal grid has a clear fish body affiliation, time affiliation, and location coordinates. Subsequent abnormal block merging and abnormal block concatenation are all performed under the same fish body identification, so as not to be confused with the grids of other fish bodies. Furthermore, for abnormal grids in the abnormal grid map, the adjacent relationships are checked according to the shared boundary method. Abnormal grids that are directly connected in the head-tail direction or the dorsal-ventral direction are merged into the same abnormal block, while abnormal grids that are obliquely contacted but do not share a boundary are not merged into the same abnormal block. After the merging is completed, all head-tail grid numbers and dorsal-ventral grid numbers contained in each abnormal block are extracted and written into the abnormal block record in ascending order of grid number. The same abnormal block corresponds to one abnormal block number, and one abnormal block number belongs to only one fish body identifier and one collection time. Thus, the abnormal block set is output. By fixing the merging caliber of using shared boundaries, the excessive merging of abnormal blocks due to oblique contact can be avoided, and a unified block-level object is also provided for subsequent block-by-block comparison. Subsequently, during adjacent data collection times, the abnormal blocks from the previous time step and the abnormal blocks from the next time step under the same fish body identifier are compared one by one. The comparison includes whether the two contain grids with the same head and tail grid numbers and dorsal and ventral grid numbers, or whether they contain no identical grids but have an adjacent relationship where the head and tail grid numbers differ by one and the dorsal and ventral grid numbers are the same, or whether they have an adjacent relationship where the dorsal and ventral grid numbers differ by one and the head and tail grid numbers are the same. If any relationship is satisfied, the abnormal blocks from the previous time step are recorded as a pair. If the same abnormal block from the previous time step corresponds to multiple abnormal blocks from the next time step, the abnormal block from the next time step with the most overlapping grids is retained first. If the number of overlapping grids is the same, the one with the smaller sum of the differences between the head and tail grid numbers is retained. If they are still the same, the one with the smaller sum of the differences between the dorsal and ventral grid numbers is retained. If the same abnormal block from the next time step corresponds to multiple abnormal blocks from the previous time step, the unique pair is retained in the same order, and the other pairs are discarded. By adding this adjudication process, it is possible to avoid a single abnormal block being linked into multiple chains at the same time, ensuring that the subsequent abnormal chain objects are unique. Finally, a consistency check is performed based on the order of the head and tail grid numbers of the abnormal blocks in the concatenation pair. If the grid arrangement of the abnormal block in the head-to-tail direction is consistent with that of the abnormal block in the previous time step, the concatenation pair is retained. If the position of the abnormal block in the head-to-tail direction of the subsequent time step crosses the position of the abnormal block in the previous time step, the concatenation pair is discarded. For the retained concatenation pairs, they are connected sequentially according to the acquisition time order: if the abnormal block in the subsequent time step of the previous concatenation pair is the same abnormal block as the abnormal block in the previous time step of the subsequent concatenation pair, the two concatenation pairs are grouped into the same abnormal chain. If an abnormal block fails to form a valid concatenation pair with any abnormal block in the next acquisition time step, the abnormal block is used as the termination block of the corresponding abnormal chain. If an abnormal block in the subsequent time step does not match an abnormal block in the previous time step, the abnormal block is used as the starting block of a new abnormal chain. Thus, the abnormal chain set corresponding to each fish body is output. By writing the start, continuation, and termination relationships together, the generation process of the abnormal chain in continuous acquisition time can be kept closed. Through the above processing, the anomaly chaining module organizes local anomalies grid by grid into anomaly chains defined by fish body identification, acquisition time, and location continuity. This allows subsequent extended statistics to be calculated not separately for each independent anomaly block, but for the change process of the same anomaly object in consecutive acquisition times. This processing maintains the continuity of anomaly locations and suppresses the problems of false merging, false chaining, and duplicate assignment of multiple chains by merging shared boundaries, pairing blocks by blocks, and conflict resolution, thereby improving the stability of subsequent evolution determination. In practical applications: When a leopard-gill spiny perch exhibits localized anomalies in its body surface unfolding plot across multiple consecutive sampling times, the system first marks the abnormal grids in the corresponding body surface grid plots at each sampling time. Then, it merges the abnormal grids with adjacent boundaries at the same sampling time into anomaly blocks. Subsequently, it compares whether the anomaly blocks between adjacent sampling times contain the same or adjacent grids, and determines the unique pairing relationship based on the number of overlapping grids and the difference in grid numbers. For pairing results that satisfy the head-to-tail order, they are sequentially linked into anomaly chains according to the order of sampling times, so that persistent and gradually changing body surface anomalies on the same fish are identified as the same continuous object, providing a stable input for subsequent expansion direction statistics and disease early warning.

[0019] The extended statistics module is used to extract the edge grid, internal grid and external neighbor grid of the abnormal block in the abnormal chain at the previous time step, and then count the number of newly added grids, the number of retained grids and the number of exited grids at the next time step to generate extended records; In this embodiment, the extended statistics module is used to extract the boundary changes and outward expansion direction of the same anomaly between adjacent acquisition times, based on the established anomaly chain. During processing, instead of directly comparing the overall area of ​​the anomaly blocks at two times, the module first divides the anomaly block at the previous time into edge grids and internal grids, and then establishes an outer neighbor grid outside the edge grids, thus forming the inner region, boundary region, and outward expansion region corresponding to the anomaly block at the previous time. Subsequently, each grid in the anomaly block at the next time is mapped to the above three regions one by one, distinguishing between retained grids and newly added grids. Finally, the anomaly block at the previous time checks the anomaly block at the next time to determine the grid exit and generate an extension record accordingly. After this processing, the continuation, outward expansion, and contraction of the anomaly block between adjacent acquisition times can be quantified separately under a unified grid coordinate system, providing a directly callable directional and quantitative basis for subsequent evolution determination. The implementation process includes the following steps: First, for the previous anomalous block in the anomalous chain, read all the grids contained in the anomalous block one by one, and check the adjacent grids in the head-to-head, tail-to-head, back-to-head, and ventral directions of each grid. If a grid is connected to a non-abnormal grid in at least one of the four directions, it is marked as an edge grid; if a grid is connected to a grid within the same anomalous block in all four directions, it is marked as an internal grid. The non-abnormal grids here include grids that are not marked as anomalous in the current fish body surface grid diagram, as well as grids located outside the boundary of the body surface grid diagram that do not have valid grids, and therefore located at the outer edge of the anomalous block and close to the boundary of the fish body surface. Mesh can also be identified as edge mesh; after completing the edge mesh and internal mesh division, the outer direction of each edge mesh is recorded; if the head direction of an edge mesh is connected to a non-abnormal mesh, the head-outward mark of that mesh is written into the edge direction table; the tail direction, back direction, and ventral direction are written in the same way; if the same edge mesh is connected to a non-abnormal mesh in two or more directions, the corresponding direction mark is written respectively; thus forming the edge mesh set, internal mesh set, and edge direction table; by separating the edge mesh and internal mesh and recording the outer direction of the edge mesh, a clear starting point and directional basis can be provided for the subsequent extraction of neighboring meshes; After obtaining the edge direction table, the outer direction of each edge grid is read item by item, and the adjacent grids sharing the boundary with the edge grid along the outer direction are extracted as the outer neighbor grids. Specifically, for edge grids marked with a head-outward mark, an adjacent grid is extracted in its head direction and recorded as the head-outward outer neighbor grid; for edge grids marked with a tail-outward mark, an adjacent grid is extracted in its tail direction and recorded as the tail-outward outer neighbor grid; the extraction methods for the back and front directions are the same; if the same outer neighbor grid is extracted from multiple edge grids, only one grid record is retained in the outer neighbor grid set, but all direction sources are retained in this grid record; if the same outer neighbor grid is written to two or more direction sources at the same time, the grid participates in the counting of the corresponding directions in the subsequent direction statistics, but is only recorded as one outer neighbor grid in the grid existence judgment; for positions that exceed the boundary of the body surface grid map, no outer neighbor grid record is generated; thus, the outer neighbor grid set is obtained; the purpose of this processing is to limit the position on the outside of the abnormal block that is most likely to expand in the previous moment to the layer of grids directly connected to it, thereby limiting the judgment of subsequent new grids to a clear and verifiable spatial range; Subsequently, each mesh in the abnormal block at the next moment is compared with the internal mesh set, the edge mesh set, and the external neighbor mesh set. If a mesh at the next moment has the same coordinates as any mesh in the internal mesh set or any mesh in the edge mesh set, it is first recorded as a reserved mesh. Only if the mesh is not recorded as a reserved mesh is it checked again to see if it has the same coordinates as any mesh in the external neighbor mesh set. If they do, it is recorded as a new mesh in the corresponding direction according to the direction of origin of the external neighbor mesh. For external neighbor meshes with multiple direction sources, if a mesh falls into the external neighbor mesh at the next moment, it is recorded once in the new mesh count for the corresponding direction, but only one record of the mesh is retained in the new mesh details table. After all comparisons are completed, the number of reserved meshes is counted, and the number of new meshes in the head direction, tail direction, back direction, and ventral direction are counted separately. By determining the order of retention first and then addition, the same mesh can be avoided from being counted as both a reserved mesh and a new mesh at the same time. By counting the number of new meshes in each direction separately, a direct basis can be provided for the generation of the subsequent main expansion direction. Finally, the coordinates of each grid in the anomaly block from the previous time step are compared with those of each grid in the anomaly block from the next time step. If a grid from the previous time step does not have a corresponding grid with the same coordinates in the anomaly block from the next time step, that grid is marked as an exited grid. After completing all comparisons, the number of exited grids is obtained. Based on this, an extended record is generated, which at least includes the fish body identifier, anomaly chain identifier, previous collection time, next collection time, number of retained grids, number of newly added grids in the head direction, number of newly added grids in the tail direction, number of newly added grids in the back direction, number of newly added grids in the ventral direction, and number of exited grids. The main expansion direction is determined by the largest number of newly added grids in the four directions. If two or more directions have the same number of newly added grids, the main expansion direction recorded in the previous comparison period is used first. If there is no main expansion direction in the previous comparison period, the directions are selected in a fixed order: head, tail, back, and ventral. The direction order is generated by arranging the number of newly added grids in the four directions from largest to smallest. If the number of newly added grids is the same, they are arranged in the above fixed order. This results in an expansion record containing the main expansion direction and the direction order. By simultaneously writing the number of retained grids, the number of newly added grids in each direction, and the number of grids exited, the expansion record can not only reflect whether the anomaly continues to exist, but also where the anomaly expands and whether local contraction occurs. Through the above processing, the extended statistics module decomposes the changes of anomalous blocks at adjacent collection times in the anomaly chain into three categories: inner continuation, boundary expansion, and exit of the original region. It further provides the main expansion direction and direction order, so that subsequent evolution judgment no longer depends on a general comparison of the appearance of the anomalous blocks, but is based on the specific changes under a unified grid coordinate system. This not only improves the feasibility of expansion direction judgment, but also reduces misjudgments caused by slight fluctuations in the shape of the anomalous blocks. In practical applications: When a leopard-gill sea bass corresponds to an anomaly block at the previous time and the next time on the same anomaly chain at two consecutive acquisition times, the system first divides the anomaly block at the previous time into edge grids and internal grids, and records the outer direction of the edge grids according to the head, tail, back, and ventral directions, and then extracts an outer neighbor grid layer accordingly; then, each grid in the anomaly block at the next time is compared with the internal grid, edge grid, and outer neighbor grid in turn, and the grids that still fall within the original anomaly block range are recorded as retained grids, and the grids that fall within the outer neighbor grid range are recorded as newly added grids according to their direction; then, it checks in reverse which grids in the anomaly block at the previous time did not continue to appear at the next time and records them as exited grids; finally, it generates an expansion record based on the number of retained grids, the number of newly added grids in each direction, and the number of exited grids, which is used to determine whether the anomaly is continuously expanding, locally shrinking, or just remaining in its original state.

[0020] The evolution determination module is used to compare the number of newly added grids, the number of retained grids, the number of exited grids, and the expansion direction of the same anomalous chain in consecutive time periods. The anomalous chain that has retained grids and newly added grids in the same direction in two consecutive adjacent time periods without direction reversal is identified as an evolutionary anomalous chain, and the evolutionary anomalous chain is output. In this embodiment, the evolution determination module is used to determine whether the same anomalous chain exhibits continuous expansion pathological changes between consecutive acquisition times, based on the already formed extended records. This process does not make a one-time judgment on a single comparison period, but first organizes the comparison periods according to the order of acquisition time, then completes the retention determination and direction determination separately, and then uses the consecutive comparison periods as the cross-determination objects to distinguish the continuation, transition, rebound and termination of the anomalous chain. Finally, it performs a convergence determination based on the exit and addition cases, and continues to verify in subsequent comparison periods, thereby writing the anomalous chains that truly have continuous expansion characteristics into the evolutionary anomalous set. Through this processing method, short-term fluctuations, directional swings and local interruptions can be distinguished from continuous evolution, so that the subsequent early warning results are based on continuous temporal evidence. The implementation process includes the following steps: First, for the same abnormal chain, the corresponding extended records for each time point are read sequentially according to the acquisition time. The number of newly added grids, the number of retained grids, the number of exited grids, and the expansion direction are extracted from each extended record, and two adjacent acquisition times are used as a comparison time period and written into the time period table. Each comparison time period records the number of newly added grids, the number of retained grids, the number of exited grids, and the expansion direction of the previous and subsequent times, and writes the fish body identifier, the abnormal chain identifier, the previous acquisition time, and the subsequent acquisition time. If the same abnormal chain covers multiple consecutive acquisition times, multiple comparison time periods are formed sequentially according to the acquisition time. The previous acquisition time of the later comparison time period is consistent with the subsequent acquisition time of the previous comparison time period, and no cross-time period jump ratio is performed. After this processing, the change process of the same abnormal chain at consecutive acquisition times is organized into a time period table with a clear order, providing a unified input for subsequent marking and cross-judgment. Subsequently, retention and direction determination are performed in parallel for each comparison period in the time period table. Retention determination is based on the number of retained grids at the time following the comparison period; periods with a greater than zero retained grid count at the time following the comparison period are designated as retention periods, while periods with zero retained grid count at the time following the comparison period are designated as disconnection periods. Direction determination is based on the number of new grids at the time following the comparison period and the expansion directions of the two time periods before and after the comparison period; periods with a greater than zero new grid count at the time following the comparison period and the same expansion direction at the time following the comparison period are designated as unidirectional periods; periods with a greater than zero new grid count at the time following the comparison period and a different expansion direction at the time following the comparison period are designated as turning periods; and periods with zero new grid count at the time following the comparison period are designated as stop-expansion periods. Here, retention periods and disconnection periods form one set of mutually exclusive flags, while unidirectional periods, turning periods, and stop-expansion periods form another set of mutually exclusive flags. Each comparison period has both a retention flag and a direction flag. After marking all comparison periods, a time period flag table is output. This processing ensures that each comparison period reflects both whether the anomaly continues to exist within its original range and whether the anomaly continues to expand outwards, as well as whether the expansion direction has changed. After the time period marking table is formed, cross-judgment is performed on two consecutive comparison time periods in the same anomaly chain. During cross-judgment, only two comparison time periods sharing the same intermediate acquisition time are processed, and non-consecutive comparison time periods are not directly compared. When both the previous and next comparison time periods are reserved time periods, the direction-type marker is checked first. If both the previous and next comparison time periods are in the same direction, a persistence marker is written, indicating that the anomaly continues to expand in the same direction within two consecutive comparison time periods. If the previous comparison time period is in the same direction and the next comparison time period is a turning time period, the expansion direction of the next comparison time period is further checked to see if it is the same as the expansion direction of the previous comparison time period. Regarding the directional relationship, if the previous comparison period is head-oriented and the next comparison period is tail-oriented, or the previous comparison period is tail-oriented and the next comparison period is head-oriented, or the previous comparison period is back-oriented and the next comparison period is ventral, or the previous comparison period is ventral and the next comparison period is back-oriented, it is marked as a bounce mark. Other directional changes are marked as transition marks. If at least one of the previous or next comparison periods is a disconnection period, or at least one of the expansion stops, then a stop mark is directly written. After this processing, the changes of the abnormal chain in two consecutive comparison periods can be divided into four states: stable continuation, directional transition, directional bounce, and timing stoppage. Furthermore, a convergence determination is performed on the abnormal chains that have been written with persistent or transitional markers. During the convergence determination, the number of exited grids and the number of newly added grids corresponding to the previous and subsequent comparison periods are read. If the number of exited grids in the subsequent comparison period is not greater than the number of exited grids in the previous comparison period, and the sum of the number of newly added grids in the previous and subsequent comparison periods is greater than the number of exited grids in the subsequent comparison period, the abnormal chain is recorded as a candidate evolutionary abnormality. If the above conditions are not met, the abnormal chain is written to the check table, and the corresponding comparison period is returned to re-extract the expansion direction. During re-extraction, the corresponding abnormal chain in the expansion statistics module is called. The records are expanded, and the main expansion direction is re-determined according to the direction with the most newly added grids. If the number of newly added grids is the same, the expansion direction recorded in the previous comparison period is used first. If it is still impossible to distinguish, the direction is selected in a fixed order of head, tail, back, and ventral. Anomaly chains marked with a bounce mark are directly written into the non-evolutionary table. Anomaly chains marked with a stop mark are directly written into the check table. By introducing a joint judgment of exit and new cases after the continuous marking or transition marking, it is possible to avoid misrecording obviously contracted anomalies as candidate evolutionary anomalies simply because of the continuity of the direction. At the same time, it also preserves a verification path for anomalies with local instability in the direction. Finally, the candidate evolutionary anomalies are checked in subsequent comparison periods according to the acquisition time sequence. During the check, the comparison period markers corresponding to each candidate evolutionary anomaly are read one by one. If a retention period exists continuously in the subsequent comparison period and no bounce marker appears, the anomaly chain is identified as an evolutionary anomaly and written into the evolutionary anomaly set. If a bounce marker appears in the subsequent comparison period, the anomaly chain is excluded from the evolutionary anomaly set and written into the non-evolutionary table. If a disconnection period or a stoppage period appears in the subsequent comparison period, the anomaly chain is excluded from the evolutionary anomaly set and written into the check list. For anomaly chains written into the check list, if the retention and directional continuity conditions are met again in a new comparison period, they can re-enter the convergence determination. If the aborted state continues to occur, they are retained in the check list and not written into the evolutionary anomaly set. Through this process, a unique write relationship is maintained between the evolutionary anomaly set, the check list, and the non-evolutionary table. The same anomaly chain retains only one state in the same round of determination, thus making the evolution determination process completely closed. Through the above processing, the evolution judgment module decomposes the expansion process of the same abnormal chain in multiple consecutive acquisition times into five consecutive steps: time period organization, parallel marking, cross-judgment, convergence judgment, and subsequent verification. This makes the judgment of evolution anomalies no longer dependent on a single comparison result, but based on the joint basis of continuous temporal evidence, directional change relationship, and exit contraction situation. In this way, it can not only identify anomalies that are truly continuously expanding, but also exclude anomalies such as directional rebound, local cessation of expansion, and short-term disconnection, thereby improving the stability and reliability of subsequent disease warning results. In practical applications: When an abnormal chain corresponding to a certain leopard-gill sea bass covers multiple consecutive acquisition times, the system first organizes the extended records between two adjacent acquisition times into multiple comparison periods. Then, it determines whether each comparison period belongs to a retention period and whether it belongs to a same-direction period, a turning period, or a period of cessation of expansion. Subsequently, it performs cross-determination on two consecutive comparison periods, marking abnormal chains that expand continuously in the same direction as continuous, abnormal chains that change in different directions as transitions, abnormal chains that change in opposite directions as bounces, and abnormal chains that break or stop expanding as terminated. For continuous or transitional abnormal chains, it combines the number of newly added grids and the number of grids that exited in the preceding and following periods to determine whether they still maintain the overall expansion trend, and records those that meet the conditions as candidate evolutionary anomalies. Finally, it continues to check whether the chain remains retained and does not bounce in subsequent comparison periods. If it meets the conditions, the abnormal chain is identified as an evolutionary anomaly and written into the evolutionary anomaly set; otherwise, it is written into the check table or the non-evolutionary table, thereby providing stable abnormal evolution results for subsequent disease warning output.

[0021] The early warning output module is used to generate disease early warning results for each evolutionary anomaly in the evolutionary anomaly set according to fish body identification, first time, duration, coverage grid range and expansion direction, and output the corresponding disease early warning information for each individual leopard gill spiny perch. In this embodiment, the early warning output module is used to organize the judgment results of the evolutionary anomaly set into disease early warning information that can be directly output. During processing, the evolutionary anomalies are first merged at the fish body level to form early warning records corresponding to the same fish body. Then, disease early warning results are generated based on the coverage grid range, duration, expansion direction, and first time in the early warning records. The purpose of this processing is to unify the multiple evolutionary anomaly results obtained in the previous steps to the fish body object level, avoid the same fish body generating repeated early warnings in similar locations and time periods, and at the same time make the output results have three types of information: location range, duration, and expansion trend. The implementation process includes the following steps: First, the evolutionary anomalies in the set are merged according to fish body identifiers. Under the same fish body identifier, the initial time, duration, coverage grid range, and expansion direction of the corresponding evolutionary anomaly are read one by one, and these fields are written into the anomaly merging table of that fish body. The coverage grid range is represented by the minimum head-tail grid number, the maximum head-tail grid number, the minimum dorsal grid number, and the maximum dorsal grid number. The duration is obtained by subtracting the initial time from the last acquisition time when the evolutionary condition of the anomaly was met, and then adding one acquisition cycle. After the field extraction is completed, multiple evolutionary anomalies under the same fish body are sorted by the initial time from earliest to latest. If the initial times are the same, they are sorted by the duration from longest to shortest. If the duration is the same, they are sorted by the total number of grids corresponding to the coverage grid range from largest to smallest. After sorting, Next, two adjacent evolutionary anomalies are checked sequentially: if their coverage grid ranges overlap, or if the maximum head and tail grid number of one anomaly differs from the minimum head and tail grid number of the other anomaly by one and their dorsal and ventral grid ranges overlap, and their expansion directions are the same, then the two anomalies are merged into the same warning record. When merging, the earlier of the two anomalies is taken as the initial time, the duration is taken as the overall time span covered, the coverage grid range is taken as the minimum outer grid range after the two anomalies are merged, and the expansion direction remains the original common direction. If the above conditions are not met, they are retained as different warning records. This generates warning records for each fish body. By sorting and then merging, it can be ensured that spatiotemporally continuous and directional evolutionary anomalies on the same fish body are merged and output, while evolutionary anomalies that are separate or have different directions are kept as independent records. Subsequently, the warning intensity was calculated for each fish in the warning record, and disease warning results were generated accordingly. The calculation first involved determining the total number of covered grids based on the coverage area, specifically by subtracting the minimum head and tail grid number from the maximum head and tail grid number and adding one, and then multiplying this result by subtracting the minimum dorsal and ventral grid number from the maximum dorsal and ventral grid number and adding one. The total number of covered grids was then multiplied by the duration to obtain the warning intensity corresponding to that warning record. For fish with multiple warning records, each fish's warning intensity and corresponding location range were retained without further merging. After calculating the warning intensity, the fish identifier, initial time, duration, coverage area, expansion direction, and warning intensity were written into the disease warning results. The initial time indicates the earliest time the fish entered the warning state, the expansion direction indicates the main expansion trend of the abnormality, and the coverage area indicates the location of the abnormality on the body surface. Finally, the disease warning results were compiled into disease warning information output for each individual leopard gill spiny perch. This processing preserves both the warning location and time information, as well as the abnormality expansion trend and duration information, facilitating subsequent display, alerting, or manual review. Through the above processing, the early warning output module unifies the judgment results of the evolutionary anomaly concentration to the fish body level, and completes sorting, merging and intensity calculation within the same fish body. This makes the final output disease early warning information no longer a scattered evolutionary anomaly result, but an early warning record that corresponds one-to-one with the fish body object. This reduces repeated alarms for the same fish body and enables the early warning result to reflect the abnormal location, duration and expansion trend at the same time, improving the readability and executability of the output result. In practical applications: When a leopard-gill spiny perch develops multiple evolutionary anomalies during continuous monitoring, the system first groups these anomalies into the same fish based on the fish's body identifier. Then, it extracts the initial time, duration, coverage grid range, and expansion direction of each anomaly and sorts them sequentially by initial time, duration, and coverage range. For evolutionary anomalies with contiguous or overlapping coverage areas and consistent expansion directions, the system merges them into a single warning record. Subsequently, the warning intensity is calculated based on the total number of coverage grids and duration of the warning record, and the disease warning result for the fish is generated by combining the expansion direction and initial time. Finally, the system outputs the corresponding disease warning information for the fish, allowing aquaculture personnel to locate abnormal areas and handle them promptly.

[0022] Working Principle: This scheme first processes continuously collected fish images from the fish farm to identify individual leopard-gill spiny perch. The fish bodies are then uniformly aligned to a fixed head-to-tail and dorsal-ventral direction. The fish surface is then unfolded into a two-dimensional region and divided into numbered surface grids. Based on this, machine learning is used to mark the presence of anomalies in each grid. Anomaly grids with adjacent boundaries at the same time are merged into anomaly blocks. Anomaly blocks with consecutive positions and consistent head-to-tail order at adjacent time points are then chained together into anomaly chains. Subsequently, the retained, newly added, and removed grids, as well as the expansion direction, are compared around the anomaly chain to determine whether the anomaly is a short-term fluctuation, a local interruption, or a continuous expansion. Only anomalies meeting the continuous expansion condition are identified as evolutionary anomalies. Finally, evolutionary anomalies on the same fish are grouped by time, range, and direction to generate disease warning results for the corresponding fish. In other words, this scheme does not only look at whether there are anomalies in a single frame, but connects the abnormal changes of the same fish over consecutive time points to distinguish between incidental damage and truly continuously developing pathological anomalies. For example, in a high-density recirculating aquaculture system, cameras continuously record fish schools. A leopard-gill sea bass initially shows only a few abnormal grids on its side. Subsequently, these abnormal grids continue to appear in adjacent grids and gradually expand in the same direction. The system first uses machine learning to identify these local anomalies, then links them into a chain based on their location. It then compares which grids remain, which are newly added, and which have disappeared, and determines whether the expansion direction is continuous. If the anomaly persists for multiple consecutive moments and expands in the same direction, it is recorded as an evolutionary anomaly. The system then generates disease warning information for the fish based on the earliest appearance time, duration, coverage area, and expansion direction of the anomaly. In this way, fish farmers see not only that the fish is abnormal, but also which fish, which area of ​​its body is affected, how long it has been present, and whether it is still expanding, allowing for earlier intervention.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture, characterized in that, include: The image modeling module is used to acquire a series of continuously collected aquaculture images, perform recognition, correction and body surface unfolding on the fish, segment by head and tail and row by back and belly, and generate a body surface mesh map; The anomaly concatenation module is used to input the body surface grid map into a machine learning recognizer trained on diseased fish samples, perform anomaly marking on each grid, merge the abnormal grids into anomaly blocks, and concatenate the abnormal blocks with the same or adjacent positions at adjacent time points and the same head-to-tail order into anomaly chains, and output an anomaly chain set. The extended statistics module is used to extract the edge grid, internal grid and external neighbor grid of the abnormal block in the abnormal chain at the previous time step, and then count the number of newly added grids, the number of retained grids and the number of exited grids at the next time step to generate extended records; The evolution determination module is used to compare the number of newly added grids, the number of retained grids, the number of exited grids, and the expansion direction of the same anomalous chain in consecutive time periods. The anomalous chain that has retained grids and newly added grids in the same direction in two consecutive adjacent time periods without direction reversal is identified as an evolutionary anomalous chain, and the evolutionary anomalous chain is output. The early warning output module is used to generate disease early warning results for each evolutionary anomaly in the evolutionary anomaly set according to fish body identification, initial time, duration, coverage grid range and expansion direction, and output the corresponding disease early warning information for each individual leopard gill spiny perch.

2. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture as described in claim 1, characterized in that: The image modeling module includes: For each time-phase image in the aquaculture image sequence, perform leopard gill spiny perch body recognition and extract the fish body outline, and output the outline image of each fish body; Based on the outline image, the head position, tail position, dorsal boundary and ventral boundary of each fish are located, and the fish orientation is corrected according to the direction from head to tail, and the corrected image of each fish is output.

3. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture as described in claim 2, characterized in that: The image modeling module also includes: Based on the corrected image, the surface area between the dorsal and ventral boundaries of each fish body is unfolded in the head-to-tail direction, and a two-dimensional unfolded map corresponding to the head-to-tail position and the dorsal and ventral position is established, and the unfolded map of the body surface of each fish body is output. The body surface unfolding diagram is segmented along the head-to-tail direction and then divided into rows along the dorsal-ventral direction. The area formed by the intersection of each segment and each row is recorded as the body surface grid. The corresponding body surface grid diagram of each fish at each time point is output.

4. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture as described in claim 3, characterized in that: The abnormal connection module includes: Each grid in the body surface grid map is input into a machine learning recognizer to obtain anomaly markers for each grid. The grid marker results are written according to the fish body identification and the collection time, and the anomaly grid map is output. For abnormal grids in the abnormal grid diagram, merge the abnormal grids that are adjacent to the boundary into the same abnormal block, and record the head and tail grid numbers and back and belly grid numbers contained in each abnormal block, and output the abnormal block set.

5. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture according to claim 4, characterized in that: The abnormal connection module also includes: In adjacent time intervals, the head and tail grid numbers and back and belly grid numbers of the abnormal block in the previous time interval are compared with the abnormal block in the next time interval one by one. The abnormal blocks that contain the same grid, head and tail adjacent grids, or back and belly adjacent grids are paired and the concatenated pairs are output. Based on the order of the head and tail grid numbers of the abnormal blocks in the concatenation pair, concatenation pairs with consistent head and tail numbers are retained and concatenated sequentially according to the acquisition time, and the abnormal chain set is output.

6. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture according to claim 5, characterized in that: The extended statistics module includes: For the previous time step of the abnormal block in the abnormal chain, check the four-sided adjacency relationship of each grid in the abnormal block one by one. Grids that are connected to non-abnormal grids on at least one side are recorded as edge grids, and the remaining grids are recorded as internal grids. Record the outer direction of each edge grid according to the head direction, tail direction, back direction and ventral direction respectively, and output the edge grid set, internal grid set and edge direction table. Based on the edge direction table, extract one layer of adjacent meshes along the corresponding outer direction for each edge mesh as the outer neighbor mesh, and write direction labels for the head-to-outer neighbor mesh, tail-to-outer neighbor mesh, back-to-outer neighbor mesh, and ventral-to-outer neighbor mesh respectively, and output the outer neighbor mesh set.

7. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture according to claim 6, characterized in that: The extended statistics module also includes: Each grid in the abnormal block at the next time step is compared with the internal grid set, the edge grid set, and the external neighbor grid set one by one. Grids that fall into the internal grid set or the edge grid set are recorded as retained grids. Grids that fall into the external neighbor grid set are recorded as newly added grids in the corresponding direction according to the direction identifier. Output the number of retained grids and the number of newly added grids in each direction. Each grid in the anomalous block of the previous time step is compared with the anomalous block of the next time step one by one. Grids that do not match the anomalous block of the next time step are marked as exited grids. Then, an expansion record containing the main expansion direction and the order of directions is generated according to the number of grids retained, the number of grids added in each direction, and the number of exited grids.

8. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture according to claim 7, characterized in that: The evolution determination module includes: For the same anomaly chain, extract the number of newly added grids, the number of retained grids, the number of exited grids, and the direction of expansion at each time point in the order of collection time. Form a comparison time period with two adjacent time points, and write the number of newly added grids, the number of retained grids, the number of exited grids, and the direction of expansion at the previous and subsequent time points into each comparison time period, and output the time period table; For each comparison period in the time period table, perform retention and direction determination in parallel. The comparison period with a greater than zero number of retained grids in the later time is recorded as the retention period, otherwise it is recorded as the disconnection period. The comparison period with a greater than zero number of newly added grids in the later time and the same expansion direction as the previous time is recorded as the same direction period. The comparison period with a greater than zero number of newly added grids in the later time and the different expansion direction from the previous time is recorded as the turning period. The comparison period with zero number of newly added grids in the later time is recorded as the stop expansion period. Output the time period mark table.

9. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture according to claim 8, characterized in that: The evolution determination module also includes: For two consecutive comparison periods in the same abnormal chain, perform cross-determination. If both the previous and subsequent comparison periods are reserved periods, and both are in the same direction, write a persistence flag. If the previous comparison period is in the same direction and the subsequent comparison period is a turning period, write a bounce flag when the expansion direction of the subsequent comparison period is opposite to that of the previous comparison period; otherwise, write a transition flag. If either the previous or subsequent comparison period is a disconnected period or a stopped expansion period, write an abort flag. For anomalous chains with persistent or transitional markers, a merge determination is performed. If the number of exited grids in the later time step is not greater than the number of exited grids in the previous time step, and the sum of the number of newly added grids in the previous time step and the number of newly added grids in the later time step is greater than the number of exited grids in the later time step, the corresponding anomalous chain is recorded as a candidate evolutionary anomalous chain. Otherwise, the corresponding anomalous chain is written to the check list and the corresponding comparison period is returned to re-extract the expansion direction. Anomalous chains with bounce markers are written to the non-evolutionary list, and anomalous chains with termination markers are written to the check list. For candidate evolutionary anomalies, continue to check subsequent comparison periods in chronological order. If there is a retained period and no bounce marker, the corresponding anomaly chain is identified as an evolutionary anomaly and written into the evolutionary anomaly set before being output. Otherwise, the corresponding anomaly chain is excluded from the evolutionary anomaly set and written into the check table or non-evolutionary table.

10. The machine learning-based disease early warning and monitoring system for leopard-gill spiny perch aquaculture according to claim 9, characterized in that: The early warning output module includes: The evolutionary anomalies in the set of evolutionary anomalies are merged according to the fish body identifier. For each evolutionary anomaly of the same fish body, the initial time, duration, coverage grid range and expansion direction are extracted. The anomalies are sorted and merged according to the order of the initial time, the duration, the size of the coverage grid range and whether the expansion direction is consistent, and the warning records of each fish body are generated. The warning intensity is calculated for each fish in the warning record according to the coverage grid range and duration. The disease warning result for the corresponding fish is generated by combining the expansion direction and the first time. The disease warning information for each individual leopard gill spiny perch is output.

Images