Water aquaculture precise feeding decision method and system based on visual behavior analysis

Abstract

This invention discloses a method and system for precise feeding decision-making in aquaculture based on visual behavior analysis, belonging to the field of image detection technology. The method and system include the following steps: S1, synchronously observing the activity process of fish schools and the action process of feeding equipment within the aquaculture area; S2, dividing the observation into stages around a single feeding process and analyzing the source of current visual changes; S3, identifying the fish school transfer process and interpreting data from the current aquaculture unit; S4, generating feeding control parameters and correcting the complete control process. This invention effectively improves the accuracy and rationality of precise feeding decisions in aquaculture, solving the problem that the feeding action itself can contaminate visual observation, causing the system to misjudge the disturbance caused by feeding as actual increased feeding activity by fish.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, specifically to a method and system for precise feeding decisions in aquaculture based on visual behavior analysis. Background Technology

[0002] With the development of machine learning, image detection and recognition technologies in the field of aquaculture, visual feeding systems have become an important application direction for intelligent aquaculture. However, existing technologies have not effectively distinguished and isolated the visual disturbances in the feeding process. Most existing systems use the video streams before, during and after feeding as the same statistical caliber to input into the recognition model, which cannot truly reflect the actual feeding needs of the fish.

[0003] For example, invention patent CN115861906B discloses a method, device, system, and feeder for identifying fish feeding intensity. The method includes: acquiring fish feeding information, which includes video information of fish feeding and water quality information for the corresponding video time period; inputting the fish feeding information into a fish feeding intensity identification model to obtain the fish feeding intensity output by the model; the fish feeding intensity identification model is used to fuse video frame features, audio features, and water quality features obtained by feature extraction from the fish feeding information, and to determine the fish feeding intensity based on the fused features; the fish feeding intensity identification model is trained based on samples of fish feeding information and corresponding fish feeding intensity labels. This invention can effectively improve the accuracy and effect of fish feeding intensity identification, achieving high-precision fish feeding intensity identification even in turbid water conditions.

[0004] For example, invention patent CN112215116B discloses a mobile, real-time 3D crab detection method based on 2D images. The method includes: a CMOS camera installed below an automatic feeding vessel takes overhead shots, continuously acquiring underwater 2D RGB images reflecting the natural living state of the crabs as the vessel moves. Annotation tools and augmented reality session data are used to jointly establish a crab dataset with labeled poses and varied shapes. Inspired by the anchorless frame mechanism, a Gaussian central distribution is used, starting from a single RGB image. After feature extraction using an encoder-decoder architecture, a multi-task joint learning approach combining shape, detection, and regression is employed to predict the 2D bounding box of unknown crabs. Then, a refined pose estimation algorithm EPnP is used to extend the 2D prediction to 3D bounding boxes to estimate the crab's pose and physical size, thereby constructing an ultra-lightweight single-order 3D crab detection model. This method can improve the variable feeding efficiency and feeding effect of automatic feeding vessels.

[0005] In existing technologies, most systems assume that continuously captured images from cameras can be directly used to determine feeding status. However, in actual pond, tank, or recirculating aquaculture systems, once the feeding equipment is activated, the feed scattering trajectory, water surface impact, splash diffusion, localized foam, particle settling zone, and the instantaneous sprinting behavior of fish stimulated by feeding will all appear simultaneously in the footage. These changes are visual disturbances directly introduced by the equipment's actions, and they also overlap with actual feeding behavior within a short period. Existing technologies often use the video streams before, during, and after feeding as input to the recognition model with the same statistical caliber. This causes the model to amplify the visual disturbances triggered by the equipment feeding and treat them as indicators of vigorous feeding by the fish, creating a self-excited chain of misjudgments where the more feeding occurs, the more active the fish are judged, and the more active they are, the more feeding continues. This makes it difficult to ensure that the decision results truly reflect the feeding needs of the fish.

[0006] Therefore, in order to address the above problems, there is an urgent need for a precise feeding decision-making method and system for aquaculture based on visual behavior analysis. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and system for precise feeding decisions in aquaculture based on visual behavior analysis. This solves the problem that the feeding action itself can contaminate visual observation, causing the system to misjudge the disturbance caused by feeding as an increase in actual feeding by the fish.

[0008] Technical solution

[0009] To achieve the above objectives, the present invention provides the following technical solution: a precision feeding decision-making method for aquaculture based on visual behavior analysis, comprising: S1, synchronously observing the activity process of fish schools and the operation process of feeding equipment within the aquaculture area to form visual observation data, and performing uniform scale processing to establish a baseline observation window; S2, dividing the observation process into stages around a single feeding process, analyzing the source of current visual changes, identifying the switching of current observation segments, and constructing a set of purification behavior observation segments; S3, based on the set of purification behavior observation segments, identifying the fish school transfer process, interpreting the data of the current aquaculture unit, and obtaining a feeding interpretation report; S4, generating feeding control parameters based on the data interpretation conclusions, correcting the current round of control process, forming a closed-loop sample, and correcting the overall control process.

[0010] Furthermore, the process of synchronously observing the fish activity process and the feeding equipment operation process within the aquaculture area to form visual observation data, and performing uniform scale processing to establish a baseline observation window, is as follows: Observation is conducted within the aquaculture unit; feeding equipment operation information is synchronously accessed; image sequences, sampling time information, aquaculture unit identification, camera device identification, feeding trigger information, feeding equipment operation information, and post-feeding feedback images are collected and written into the visual observation data; the image sequences and equipment operation information are aligned according to timestamps; missing frames are marked for sampling times with missing frames; the brightness distribution, contrast distribution, and background area in the image sequences are uniformized; based on the image sequences, a fish target extraction method combining background modeling and inter-frame difference is used to separate the fish movement area; the separated fish movement area is subjected to connected component filtering and morphological correction to obtain the fish target area; based on the centroid position, circumscribed region position, and region overlap relationship of the fish target area at adjacent sampling times, cross-frame correlation tracking is performed to construct fish movement trajectory segments. Fish target detection, motion trajectory correlation, and continuous behavior analysis methods were employed to characterize the changes in regional aggregation, swimming variation, directional migration, and following response of fish during the observation period. Behavioral features corresponding to the fish aggregation range, swimming variation rate, directional vector discrete value, and following migration rate were extracted. Each behavioral feature was linearly normalized to a uniform scale. Before the feeding action was initiated, a static stability screening was performed on the image sequence. Only if the current observation segment met the static stability condition was the current observation segment included in the baseline observation window; if the static stability condition was not met, the current observation segment was discarded and the screening was repeated backward to generate the baseline observation report for the current aquaculture unit.

[0011] Furthermore, the specific process of analyzing the sources of current visual changes by dividing the observation stages around a single feeding process is as follows: Based on the feeding equipment's action information, the visual observation data is divided into the pre-feeding observation stage, the feeding process stage, and the post-feeding feedback stage; within each stage, changes in fish aggregation, local water surface disturbance, bait diffusion, particle sinking, and group movement transfer are extracted to form staged observation content; based on the image sequence and behavioral feature quantities within the feeding process stage, the sources of current visual changes are analyzed, the disturbance occurrence time window is determined based on the feeding equipment's action information, and the significance of the above disturbance components is detected within the time window to form a purified segment, in order to distinguish the disturbance directly caused by the equipment's action from the actual feeding behavior of the fish, and thus identify the disturbance components directly caused by the equipment's action; the disturbance components include the bait throwing trajectory recognition results, the water surface texture change area recognition results, the foam area recognition results, the particle strip area recognition results, and the fish sprint trajectory recognition results.

[0012] Furthermore, the specific process for switching the current observation segment is as follows: The disturbance score of the current observation segment is obtained through a fusion calculation method. When the disturbance score is greater than or equal to the disturbance threshold, the current observation segment is switched. If the number of newly appearing bait falling trajectories is greater than the trajectory number threshold, and at least one of the areas of the water surface texture change region, the foam region, and the particle strip region is greater than the corresponding area threshold, and the number of fish sprint trajectories is greater than the sprint trajectory threshold, the current observation segment is determined to be in a state dominated by disturbance induced by the feeding action. If the number of newly appearing bait falling trajectories is less than or equal to the trajectory number threshold... When the areas of water surface texture change, foam, and particle stripe are all less than or equal to the corresponding area thresholds, and the proportion of fish moving towards the feeding area is greater than the direction proportion threshold, and the proportion of fish maintaining the migration relationship towards the feeding area is greater than the following migration rate threshold, the current observation segment is determined to be a feeding-determinable state. State switching is only performed when the determination conditions for the disturbance-dominant state induced by the feeding action or the determination conditions for the feeding-determinable state are all met within the corresponding consecutive N sampling periods. If any statistical item fails to meet the corresponding determination condition within the consecutive N sampling periods, the current state remains unchanged.

[0013] Furthermore, the specific process of constructing the purification behavior observation segment set is as follows: continuous observation segments in a state where feeding can be determined are retained, and the corresponding behavioral feature quantities of the segments are retained simultaneously to form a purification behavior observation segment set; segments in a state dominated by disturbances induced by feeding actions are not directly included in the feeding interpretation, but are used as equipment action interference segments to update the disturbance template, update the discrimination threshold, update the classifier parameters, or update the mask region library to participate in feedback correction; continuous segments in the purification behavior observation segment set are structurally bound according to the breeding unit number, feeding stage information, sampling time information, and equipment number information, and written into the purification observation segment record.

[0014] Furthermore, based on the set of observation segments of purification behavior, the specific process of identifying the fish school transfer process is as follows: Based on the image sequences and behavioral features in the set of observation segments of purification behavior, a method of continuous comparison of sliding analysis windows and fusion of stage transfer indicators is used to identify the transfer process of the fish school from the waiting state to the feeding enhancement stage, from the feeding enhancement stage to the continuous feeding stage, and from the continuous feeding stage to the feeding decline stage; within each sliding analysis window, the aggregation-contraction ratio, directional consistency coefficient, feeding zone companion rate, and following-maintain similarity are extracted and jointly judged according to the sampling time order: when each indicator changes from a missing state to a formed state, it is determined that the fish school has changed from the waiting state to the feeding enhancement stage; when each indicator changes from a missing state to a formed state, it is determined that the fish school has changed from the waiting state to the feeding enhancement stage; when each indicator changes from a missing state to a formed state, it is determined that the fish school has changed from the waiting state to the feeding enhancement stage; when each indicator changes from a missing state to a formed state, it is determined that the fish school has changed from the feeding ... decline stage. If all three parameters remain valid within N consecutive sliding analysis windows, the fish population is determined to have transitioned from an enhanced feeding phase to a sustained feeding phase. When the aggregation-contraction ratio, directional consistency coefficient, and follow-and-maintain similarity successively decrease, and the main activity area of ​​the fish population moves away from the feeding area or particle distribution area, the fish population is determined to have transitioned from a sustained feeding phase to a feeding decline phase. Regional statistics, trajectory correlation, directional distribution analysis, migration connection discrimination, and phase comparison analysis methods are used to extract changes in the fish population aggregation range, follow-and-maintain status, swimming direction concentration changes, group migration continuity, feeding decline changes, and response status after disturbance withdrawal. These are then structured and arranged according to the sampling time sequence to generate a feeding behavior interpretation report corresponding to the current observation period.

[0015] Further, the specific process of interpreting data from the current aquaculture unit to obtain a feeding interpretation report is as follows: The movement status of the fish population in the current aquaculture unit over N consecutive sampling periods is determined. The state switching record corresponding to the current observation segment is read. When the current observation segment is in a feeding-determinable state, and the number of bait falling trajectories, the area of ​​water surface texture changes, the area of ​​foam areas, the area of ​​particle strip areas, and the number of fish sprint trajectories do not meet the conditions corresponding to the dominant state of disturbance induced by feeding actions, the disturbance exit is determined to be established. Simultaneously, if the feeding interpretation report indicates that the fish population is in a continuous feeding phase or an enhanced feeding phase, the current aquaculture unit is deemed to have feeding control conditions, and a feeding interpretation report is obtained. Within the observation segment of the same aquaculture unit, if the feeding interpretation report remains consistent and the feeding control conditions do not change in the opposite direction, the current data interpretation conclusion is maintained. When the interpretation direction repeatedly switches between consecutive observation segments, the baseline observation report and the stage state switching record are re-called for verification.

[0016] Furthermore, based on the data interpretation conclusions, feeding control parameters are generated, and the specific process of correcting this round of control is as follows: When the current aquaculture unit is determined to meet the feeding control conditions, a feeding start command, a single feeding duration parameter, and a single feeding amount parameter are output to the feeding equipment; when the current aquaculture unit is determined not to meet the conditions for continued feeding control, a stop command or a waiting command is output to the feeding equipment, and the start time parameter for the next feeding interpretation is generated; when the proportion of the fish group moving towards the feeding area or particle distribution area in N consecutive sampling periods is greater than the direction proportion threshold, the movement is identified as following migration behavior; when the displacement velocity of the fish group in a single sampling period is greater than the intrusion velocity threshold, the movement is identified as sprinting intrusion behavior; if the number of sampling periods identified as following migration behavior is greater than the number of sampling periods identified as sprinting intrusion behavior in N consecutive sampling periods, the feedback stage is considered to be a following migration interpretation result; otherwise, the feedback stage is considered to be a sprinting intrusion interpretation result.

[0017] Furthermore, the specific process of forming closed-loop samples and correcting the overall control process is as follows: When the feeding interpretation report, the equipment execution process, and the following migration behavior in the feedback phase are consistent, the current feeding process is recorded as a closed-loop consistent sample; when the feeding interpretation report does not match the equipment execution process, or when the number of sampling cycles for sprinting behavior in the feedback phase is greater than the number of sampling cycles for another type of behavior, or when the following migration behavior is not formed, the current feeding process is recorded as a closed-loop deviation sample; the closed-loop consistent samples and closed-loop deviation samples are classified and archived, and the closed-loop consistent samples and closed-loop deviation samples formed in each cycle are summarized to generate a feeding closed-loop report for the corresponding aquaculture unit, and the overall control process is corrected: the feeding interpretation threshold, disturbance template, stage transition criterion weight, or feeding parameter mapping function are updated using the closed-loop consistent samples and closed-loop deviation samples, respectively, and the stage transition criterion weight is updated according to the temporal characteristics of the following migration behavior in the closed-loop consistent samples, so that the stage transition judgment is more in line with the actual feeding response pattern.

[0018] Furthermore, the second aspect of this invention provides a precision feeding decision-making system for aquaculture based on visual behavior analysis, applied to a precision feeding decision-making method for aquaculture based on visual behavior analysis, comprising: an observation acquisition and baseline calibration module, used to simultaneously observe the activity process of fish schools and the action process of feeding equipment within the aquaculture area, form visual observation data, perform uniform scale processing, and establish a baseline observation window; a feeding action-induced disturbance identification module, used to divide the observation stage around a single feeding process, analyze the source of current visual changes, distinguish the switching of current observation segments, and construct a set of purification behavior observation segments; a feeding demand interpretation module, used to identify the fish school transfer process based on the set of purification behavior observation segments, perform data interpretation on the current aquaculture unit, and obtain a feeding interpretation report; and an execution control and feedback correction module, used to generate feeding control parameters based on the data interpretation conclusions, correct the current round of control process, form a closed-loop sample, and correct the overall control process.

[0019] Beneficial effects

[0020] The present invention has the following beneficial effects: (1) This invention effectively avoids the problem of reverse contamination of visual observation by feeding action. By establishing a closed-loop visual decoupling mechanism for feeding disturbance isolation and behavior purification, the feeding device action period is marked as an independent event segment and segmented modeling is performed to eliminate the disturbance caused by feeding and allow the system to make judgments based on real feeding behavior information.

[0021] (2) This invention can accurately identify the source of visual changes during the feeding process and complete the state determination. It divides the feeding observation stage and analyzes the visual change disturbance components through the disturbance identification module induced by the feeding action. It also sets strict state switching judgment conditions to accurately distinguish between the disturbance-dominated state and the feeding judgment state. By retaining the observation segments in the feeding judgment state to form a purification behavior observation segment set, and at the same time establishing templates for disturbance segments and performing reduction processing, the fish feeding behavior interpretation is freed from equipment disturbance interference.

[0022] (3) This invention realizes the refined stage identification of the feeding state of fish groups. By using the feeding demand judgment module and the method of integrating the sliding analysis window and the stage transition index, it accurately identifies the stage transition process of fish groups waiting for food, increased feeding, continuous feeding, and decreased feeding, which is consistent with the actual feeding behavior change pattern of fish groups.

[0023] (4) This invention achieves precise output and dynamic correction of feeding control. By executing the control and feedback correction module, personalized feeding parameters are output based on the interpretation conclusion. It can also accurately distinguish between fish following migration and sprinting behavior, and adjust the feeding instructions based on the behavior recognition results. It can generate a closed-loop feeding report for the breeding unit, realize the continuous optimization of the feeding process of the same breeding unit, and continuously improve the precise feeding decision-making ability with the breeding cycle.

[0024] 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

[0025] Figure 1 This is a flowchart of the precision feeding decision-making method for aquaculture based on visual behavior analysis, as described in this invention. Figure 2 This is a diagram illustrating the architecture of the aquaculture precision feeding decision-making system based on visual behavior analysis, as described in this invention. Figure 3 This is a statistical diagram of the disturbance during the feeding process of this invention; Figure 4 This is a fusion diagram of the transfer indicators during the feeding stage of this invention; Figure 5 This is a statistical chart of closed-loop sample classification in this invention. Detailed Implementation

[0026] 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.

[0027] Please see Figures 1-5 This invention provides a technical solution: a precise feeding decision-making method for aquaculture based on visual behavior analysis, comprising: S1, synchronously observe the activity process of fish and the operation process of feeding equipment in the aquaculture area to form visual observation data, and perform uniform scale processing to establish a baseline observation window; S2, divide the observation phase around the single feeding process, analyze the source of the current visual changes, identify the switching of the current observation segment, and construct a set of observation segments of purification behavior; S3, based on the observation fragment set of purification behavior, identifies the fish migration process, performs data interpretation on the current aquaculture unit, and obtains a feeding interpretation report; S4 generates feeding control parameters based on data interpretation conclusions, corrects the current control process, forms closed-loop samples, and corrects the overall control process.

[0028] Specifically, the process of simultaneously observing the activity of fish schools and the operation of feeding equipment within the aquaculture area to generate visual observation data, and then performing standardized scaling to establish a baseline observation window, is as follows: Surface cameras, underwater cameras, or fixed visual acquisition devices are deployed within the aquaculture unit to observe the activity of fish. The start-up, running, and stop statuses of feeding equipment, along with equipment number information, are simultaneously accessed to obtain equipment action information corresponding to the current feeding process. Image sequences, sampling time information, aquaculture unit identifiers, camera device identifiers, feeding trigger information, feeding equipment action information, and post-feeding feedback images are collected and written into the visual observation data. Feeding equipment action information includes start-up, running, and stop statuses, and equipment number information. Image sequences and equipment action information are aligned by timestamp. Missing frames are marked for sampling times with missing frames. Image sequences with changes in field of view are registered to a fixed reference area. Reflection suppression is applied to areas with brightness values ​​exceeding the reflectivity threshold. Occlusion segments with a continuous sampling period number less than the occlusion period threshold are marked. Background noise suppression is performed on areas with suspended noise in the water. Brightness distribution, contrast distribution, and background areas in the image sequences are standardized.

[0029] Based on image sequences, a fish swarm target extraction method combining background modeling and inter-frame differencing is used to separate the fish movement regions. Connectivity filtering and morphological correction are then applied to the separated fish movement regions to obtain the target fish region. Based on the centroid position, circumscribed region position, and region overlap of the target fish region at adjacent sampling times, cross-frame correlation tracking is performed to construct fish movement trajectory segments. Fish swarm target detection, trajectory correlation, and continuous behavior analysis methods are used to characterize the changes in the fish swarm's regional aggregation state, swimming change state, directional migration state, and follow-response state during the observation period. Behavioral features corresponding to the fish swarm aggregation range, swimming change rate, directional vector discrete value, and follow-migration rate are extracted, respectively.

[0030] Spatially merge all target areas of fish schools within the same sampling period, extract the overall coverage and regional shrinkage changes of the target areas of fish schools, and obtain the behavioral feature quantity corresponding to the fish school aggregation range; based on the displacement changes, velocity changes, and trajectory transfer frequencies of each target area of ​​fish schools between adjacent sampling times, extract the statistical changes in the target displacement velocity of fish schools within a unit observation period, and obtain the behavioral feature quantity corresponding to the swimming change rate; based on the direction vector distribution of fish school movement trajectory segments, statistically analyze the proportion of targets moving towards the material drop area and the discrete value of the direction vector within the same sampling period, and obtain the behavioral feature quantity corresponding to the discrete value of the direction vector; the discrete value of the direction vector is the variance of the angle between the movement direction of each fish school target and the center direction of the material drop area within the same sampling period; based on the decreasing relationship of the distance sequence from the target area to the material drop area, the accompanying migration relationship, and the number of sampling periods that meet the corresponding judgment conditions within adjacent sampling periods of the fish school movement trajectory segments relative to the material drop area or particle distribution area, extract the continuous following change of fish schools towards the target area, and obtain the behavioral feature quantity corresponding to the following migration rate. The migration rate is the proportion of fish targets that maintain their movement toward the feeding area and whose distance decreases between adjacent periods within a continuous sampling period, relative to the total number of active fish during that period.

[0031] A unified scaling process, using linear normalization, is applied to all behavioral characteristics to ensure that different types of behavioral characteristics are expressed within the same numerical range. This eliminates differences in numerical scale between characteristics, facilitating stage switching discrimination, feeding behavior interpretation, and closed-loop feedback correction. The minimum and maximum values ​​of each behavioral characteristic are collected from the undisturbed baseline observation window, or the mean and standard deviation of each behavioral characteristic are calculated within the initial static and stable observation window of the current breeding unit. These minimum and maximum values ​​are used as the upper and lower bounds for normalization, or the mean and standard deviation are standardized and mapped to the [0,1] interval. If no data is available, fixed empirical upper and lower bounds are used as temporary benchmarks, and the benchmarks are dynamically updated after obtaining sufficient static and stable window samples.

[0032] Before the feeding action is initiated, a static stability screening is performed on the image sequence. This screening includes: counting the number of newly appearing bait drop trajectories, the area of ​​water surface texture changes, the area of ​​foam areas, the area of ​​particle stripe areas, and the number of fish sprint trajectories within N consecutive sampling periods. When the feeding device is not activated, and the number of newly appearing bait drop trajectories is less than or equal to a trajectory number threshold, the area of ​​water surface texture changes, the area of ​​foam areas, and the area of ​​particle stripe areas are all less than or equal to their corresponding area thresholds, and the number of fish sprint trajectories is less than or equal to a sprint trajectory threshold, the current observation segment is determined to meet the static stability condition. Only when the current observation segment meets the static stability condition is it included in the baseline observation window; if it does not meet the static stability condition, the current observation segment is discarded, and the screening is repeated backwards to generate a baseline observation report for the current aquaculture unit. N is read from the parameter configuration record before the current feeding task begins and written to the processing task record corresponding to the current aquaculture unit.

[0033] In this implementation plan, by using a unified scale to express the fish school aggregation range, swimming change rate, direction vector discrete value, and following migration rate, the impact of data scale differences under different observation periods, camera positions, and image conditions on the interpretation process is reduced, and the consistency of fish school behavior information extraction results is improved. This also enhances the authenticity of baseline references, the stability of behavior interpretation, and the reliability of subsequent feeding control, providing support for achieving precise feeding decisions in aquaculture scenarios.

[0034] Specifically, the process of dividing the observation phases around a single feeding process and analyzing the sources of current visual changes is as follows: Based on the action information of the feeding equipment, the visual observation data is divided into the pre-feeding observation stage, the feeding process stage, and the post-feeding feedback stage. The pre-feeding observation stage corresponds to the continuous sampling period before the equipment starts, the feeding process stage corresponds to the continuous period of the equipment's throwing action, and the post-feeding feedback stage corresponds to the fish response window after the equipment stops.

[0035] Within each stage, changes in fish aggregation, local surface disturbance, bait diffusion, particle settling, and group movement transfer are extracted to form phased observation content. Each change is calculated periodically based on the inter-frame difference of the corresponding stage's images and target tracking results. Based on the image sequence and behavioral features within the feeding stage, the source of current visual changes is analyzed. The disturbance occurrence time window is determined based on the feeding equipment's action information, and significance detection is performed on the aforementioned disturbance components within the time window to form a purified segment. This segment distinguishes disturbances directly caused by equipment actions from actual fish feeding behavior, thereby identifying disturbance components directly caused by equipment actions. The start and end of the disturbance occurrence time window are aligned with the start and stop times of the feeding equipment, respectively. Statistical hypothesis testing is used for significance detection. Disturbance components include bait throwing trajectory recognition results, water surface texture change area recognition results, foam area recognition results, particle stripe area recognition results, and fish sprint trajectory recognition results. Each recognition result is output as a pixel-level mask or trajectory point set.

[0036] In this implementation scheme, by performing saliency detection on the bait throwing trajectory, water surface texture change area, foam area, particle strip area and fish sprint trajectory within the disturbance occurrence time window, and outputting the recognition results in the form of pixel-level mask or trajectory point set, it is possible to separate the image disturbance directly caused by the device action from the actual feeding behavior of the fish, and reduce misjudgments caused by bait entering the water, water surface impact and instantaneous sprint of the fish during the feeding process.

[0037] Specifically, the process of switching and judging the current observation segment is as follows: The disturbance score of the current observation segment is obtained by fusion calculation method. When the disturbance score is greater than or equal to the disturbance threshold, the current observation segment is switched. The ratio of the number of newly appearing bait falling trajectory to the trajectory number threshold is calculated. The ratios of the area of ​​water surface texture change region to the corresponding area threshold, the area of ​​foam region to the corresponding area threshold, and the area of ​​particle strip region to the corresponding area threshold are calculated. The maximum value of these three ratios is taken. Then the ratio of the number of fish sprint trajectory to the sprint trajectory threshold is calculated. The result of the weighted sum of the three ratios is the disturbance score.

[0038] When the number of newly appearing bait falling trajectories exceeds the trajectory number threshold, and at least one of the areas of water surface texture change region, foam region, and particle strip region exceeds the corresponding area threshold, while the number of fish sprint trajectories exceeds the sprint trajectory threshold, the current observation segment is determined to be in a state dominated by feeding action-induced disturbance. When the number of newly appearing bait falling trajectories is less than or equal to the trajectory number threshold, and the areas of water surface texture change region, foam region, and particle strip region are all less than or equal to the corresponding area threshold, and the proportion of fish moving towards the bait falling area exceeds the direction proportion threshold, and the proportion of fish maintaining a migration relationship towards the bait falling area exceeds the following migration rate threshold, the current observation segment is determined to be in a state where feeding can be judged. State switching is only performed when all the judgment conditions for the state dominated by feeding action-induced disturbance or the state where feeding can be judged are met within the corresponding consecutive N sampling periods. If any statistical item fails to meet the corresponding judgment condition within the consecutive N sampling periods, the current state remains unchanged. If, after the current observation segment has been determined to be in a feeding-inducible state, the number of newly detected bait falling trajectories exceeds the trajectory number threshold again within N consecutive sampling periods, or the area of ​​water surface texture change, foam area, particle strip area, and number of fish sprint trajectories again meet the determination conditions for the feeding action-induced disturbance-dominant state, then the current observation segment will be re-determined as the feeding action-induced disturbance-dominant state.

[0039] Table 1, showing the feeding decision data, records the underwater fish behavior and changes in water surface characteristics throughout the feeding process. Before feeding: the number of bait falling trajectories was 0, the area of ​​water surface texture changes was 5, the area of ​​foam was 3, the number of fish sprint trajectories was 1, the fish density was 20, and the average swimming speed was 15. In the early feeding stage: the number of bait falling trajectories was 15, the area of ​​water surface texture changes was 45, the area of ​​foam was 38, the number of fish sprint trajectories was 18, the fish density was 65, and the average swimming speed was 35. In the middle feeding stage: the number of bait falling trajectories was 22, the area of ​​water surface texture changes was 45, the area of ​​foam was 38, the number of fish sprint trajectories was 18, the fish density was 65, and the average swimming speed was 35. The area of ​​the water surface texture change zone was 78, the area of ​​the foam zone was 55, the number of fish sprint tracks was 25, the fish density was 85, and the average swimming speed was 45. In the later stage of feeding: the number of bait falling tracks was 10, the area of ​​the water surface texture change zone was 35, the area of ​​the foam zone was 28, the number of fish sprint tracks was 12, the fish density was 45, and the average swimming speed was 25. Feedback after feeding: the number of bait falling tracks was 2, the area of ​​the water surface texture change zone was 16, the area of ​​the foam zone was 12, the number of fish sprint tracks was 6, the fish density was 30, and the average swimming speed was 20.

[0040] Table 1. Feeding Decision Data Table Observation phase Number of bait drop trajectories Area of ​​water surface texture variation Foam area Number of fish sprinting trajectories Fish gathering density Average swimming speed Before feeding 0 5 3 1 20 15 Initial feeding 15 45 38 18 65 35 Mid-stage of baiting 22 78 55 25 85 45 Later stages of feeding 10 35 28 12 45 25 Feedback after feeding 2 16 12 6 30 20 like Figure 3The statistical chart showing the disturbance during the feeding process illustrates the changes in visual observation data of the fish school across five stages: before feeding, early feeding, mid-feeding, late feeding, and post-feeding feedback. The horizontal axis represents the stages of the feeding process, and the vertical axis represents the statistical values ​​corresponding to each statistical item. The chart provides the statistical results for the number of bait falling trajectories, the area of ​​water surface texture changes, the area of ​​foam areas, the number of fish sprint trajectories, the fish aggregation density, and the average swimming speed at each stage. As shown in the figure, the statistical values ​​are generally low in the pre-feeding stage, indicating that the surface disturbance of the aquaculture water is small and the fish are in a baseline observation state unaffected by the feeding action. In the early feeding stage, the number of feed falling trajectories, the area of ​​water surface texture changes, the area of ​​foam areas, and the number of fish dashing trajectories all increase simultaneously, indicating that the feeding equipment action has significantly disturbed the image, and the fish begin to gather towards the feed area. In the middle feeding stage, the statistical values ​​of various disturbances and fish behavior reach their peaks, indicating that this stage includes both the most obvious equipment-induced disturbances and the concentrated response of the fish to the feed area, representing disturbance recognition and feeding. The key period for interpretation is the period most prone to confusion. After entering the later stage of feeding, various statistical values ​​begin to decline, indicating that the impact of equipment operation is gradually weakening. The fish gathering density and average swimming speed remain at a relatively high level, reflecting that the fish are still migrating and feeding towards the feeding area. In the post-feeding feedback stage, the number of bait falling trajectories, the number of fish sprinting trajectories, the area of ​​water surface texture changes, and the area of ​​foam areas further decrease, while the fish gathering density and average swimming speed are still higher than in the pre-feeding stage. This indicates that the disturbance induced by equipment operation has basically withdrawn from the dominant state. At this time, it is more suitable to extract the real feeding behavior information of the fish and use it for interpretation of feeding control.

[0041] In this implementation plan, by calculating the disturbance score of the current observation segment, the switching between the dominant state of disturbance induced by the feeding action and the state of feeding can be judged in the observation segment, and it is determined whether the fish school has entered the interpretation interval that can be used for feeding analysis. This avoids directly judging the significant changes in the picture caused by the combination of bait entering the water, water surface impact, foam diffusion and fish sprint in the early and middle stages of feeding as real feeding enhancement.

[0042] Specifically, the process of constructing the observation fragment set of purification behavior is as follows: Continuous observation segments in a definable feeding state are retained, along with their corresponding behavioral characteristics, forming a set of purification behavior observation segments. These retained segments are sequentially spliced ​​together, and the behavioral characteristics include aggregation-contraction ratio, directional consistency coefficient, feeding zone companion rate, and follow-and-maintain similarity. Segments dominated by disturbances induced by feeding actions are not directly included in the feeding analysis but are treated as equipment action interference segments to update disturbance templates, discrimination thresholds, classifier parameters, or mask region libraries for feedback correction. Disturbance templates record typical spatiotemporal patterns in the interference segments, discrimination thresholds are dynamically adjusted based on interference intensity statistics, classifier parameters are updated through incremental learning, and the mask region library stores pixel-level markers of common disturbance regions. Disturbance template records are established for equipment action interference segments based on feeding equipment number, feeding position, and particle falling trajectory distribution. The feeding equipment number corresponds to a specific feeding machine identifier, the feeding position is the center coordinate of the feeding zone, and the particle falling trajectory distribution is stored as a set of trajectory points or parametric curves.

[0043] During the feeding cycle, the corresponding perturbation template record is invoked, and template matching and reduction processing is performed on similar perturbation areas in the current observation segment. When invoked, a matching perturbation template is retrieved based on the current device number and feeding position. Matching and reduction uses a sliding window correlation calculation, removing successfully matched areas from the image sequence. Continuous segments in the purification behavior observation segment set are structured and bound according to the aquaculture unit number, feeding stage information, sampling time information, and device number information, and written into the purification observation segment record. The aquaculture unit number uniquely identifies the aquaculture pond or cage; the feeding stage information is the pre-feeding, feeding in progress, or post-feeding feedback stage; the sampling time information is accurate to milliseconds or the sampling cycle number; and the device number distinguishes different feeding devices. After binding, the data is written into the record in the form of a structured data table or a serialized file.

[0044] In this implementation plan, by structurally binding and recording continuous segments in the purification behavior observation segment set according to the breeding unit number, feeding stage information, sampling time information and equipment number information, it can be ensured that subsequent feeding behavior interpretation, feeding control parameter generation and closed-loop feedback correction are all based on data with clear source, complete time sequence and traceability, thereby improving the authenticity, stability and consistency of feeding control interpretation.

[0045] Specifically, based on the observation fragment set of purification behavior, the specific process of identifying the fish migration process is as follows: Based on image sequences and behavioral features from the observation fragment set of purification behavior, a method combining continuous comparison of sliding analysis windows and stage transition indicators is used to identify the transition process of fish schools from a waiting state to an enhanced feeding stage, from an enhanced feeding stage to a continuous feeding stage, and from a continuous feeding stage to a feeding decline stage. The window length of the sliding analysis window is set to L sampling periods, and the window sliding step size is Δ sampling periods. The stage transition determination requires that K consecutive windows meet the corresponding conditions. L, Δ, and K are all positive integers, and L is greater than Δ, K is greater than or equal to two. The specific values ​​of the window length and step size are set according to the sampling period duration, so that the physical time covered by each window is not less than the fish school response delay time. Within each sliding analysis window, the aggregation-contraction ratio, directional consistency coefficient, fish-trapping rate in the feeding area, and follow-and-maintain similarity are extracted. The aggregation-contraction ratio is defined as the ratio of the minimum convex hull area of ​​the fish group in the window to that in the previous window. The directional consistency coefficient is defined as the reciprocal or normalized value of the variance of the angle between the direction of fish movement and the center direction of the feeding area. The fish-trapping rate in the feeding area is defined as the proportion of the number of fish in the feeding area to the total number of fish in the window. The follow-and-maintain similarity is defined as the mean cosine similarity between the trajectory of the fish group in the window and the displacement direction at the previous moment. The results are then combined and judged according to the sampling time order: when each indicator changes from a missing state to a formed state, the fish group is judged to have changed from a waiting state to a feeding enhancement stage. A missing state is defined as an indicator value below the corresponding threshold or no effective fish group target is detected. A formed state is defined as an indicator value reaching or exceeding the corresponding threshold.

[0046] When all indicators remain valid for N consecutive sliding analysis windows, the fish population is determined to have transitioned from the enhanced feeding phase to the sustained feeding phase. When the aggregation-contraction ratio, directional consistency coefficient, and follow-and-maintain similarity successively decrease, and the main activity area of ​​the fish population moves away from the feeding area or particle distribution area, the fish population is determined to have transitioned from the sustained feeding phase to the decreased feeding phase. "Successively decreasing" means that the above three indicators change from a formed state to a missing state in sequence. The condition for withdrawal is that the distance from the centroid of the main activity area of ​​the fish population to the boundary of the feeding area is greater than a withdrawal distance threshold.

[0047] This study employs regional statistics, trajectory correlation, directional distribution analysis, migration continuity judgment, and stage comparison analysis to extract changes in fish school aggregation range, following status, swimming direction concentration, group migration continuity, feeding decline, and response status after disturbance exit. Regional statistics are used to calculate the minimum convex hull area and rate of change of the fish school. Trajectory correlation is used to match fish school targets across frames and generate continuous trajectories. Directional distribution analysis is used to calculate the direction vector histogram and central tendency. Migration continuity judgment is used to determine whether the fish school trajectory continuously faces the feeding area. Stage comparison analysis is used to compare the changes in various indicators between adjacent stages and arranges them in a structured manner according to the sampling time sequence to generate a feeding behavior interpretation report corresponding to the current observation period. The structured arrangement uses timestamps as the primary key, storing the indicator values, stage labels, and judgment results for each sampling period or window in rows.

[0048] like Figure 4 The fusion chart of feeding phase transfer indicators shown illustrates the changes in the fish school's aggregation-contraction ratio, directional consistency coefficient, feeding zone companion rate, and following-maintaining similarity within a continuous sliding analysis window. The horizontal axis represents the sliding window number, and the vertical axis represents the corresponding indicator value. As shown in the figure, during sliding window stages 1 to 2, all indicators were at a relatively low level, indicating that the fish had not yet formed a stable feeding response towards the feeding area, and were more exhibiting natural swimming in the baseline state or local changes in the initial stage of feeding. As the sliding window entered stage 3, all indicators increased synchronously, indicating that the fish began to show consistent changes in spatial aggregation, directional migration, and companionship and following in the feeding area, corresponding to the fish's transition from a waiting state to an enhanced feeding stage. During sliding window stages 4 to 5, the values ​​of all four indicators remained in a high range, and the trend of change was consistent, indicating that the fish had formed a continuous migration relationship and a stable feeding response around the feeding area, corresponding to the fish being in a continuous feeding stage. When the sliding window entered stage 6, the aggregation-contraction ratio, directional consistency coefficient, companionship rate in the feeding area, and similarity of following decreased synchronously, indicating that the fish's aggregation around the feeding area, directional consistent migration, and continuous following relationship began to weaken, corresponding to the fish's transition from a continuous feeding stage to a feeding decline stage. The figure illustrates that this scheme does not determine the feeding status of fish groups based on a single behavioral indicator. Instead, it identifies the stage boundaries in the feeding process of fish groups by performing joint fusion analysis on multiple types of stage transition indicators, thereby reducing the risk of misjudgment of a single indicator caused by disturbances induced by feeding actions.

[0049] In this implementation scheme, by jointly judging the aggregation-contraction ratio, directional consistency coefficient, swimming rate in the feeding area, and following-maintaining similarity of fish groups within a continuous sliding analysis window based on the set of observation fragments of purification behavior, it is possible to separate the transition process of fish groups from the waiting state to the feeding enhancement stage, from the feeding enhancement stage to the continuous feeding stage, and from the continuous feeding stage to the feeding decline stage from continuous observation data. This can reduce the impact of local short-term fluctuations, abnormal single indicators, and residual disturbances induced by feeding actions on the judgment results, and improve the consistency and stability of feeding stage identification.

[0050] Specifically, the process of interpreting data from the current aquaculture unit to obtain a feeding interpretation report is as follows: The movement status of fish in the current aquaculture unit is determined over N consecutive sampling periods. The main activity area of ​​the fish is determined to be within the feeding area or particle distribution area in each sampling period. If the main activity area of ​​the fish is within the feeding area or particle distribution area for N consecutive sampling periods, the region correspondence is considered valid. The percentage of fish moving towards the feeding area is calculated. If this percentage is greater than a directional percentage threshold for N consecutive sampling periods, the dominant migration direction is considered valid. The percentage of fish maintaining a migration relationship towards the feeding area is calculated. If this percentage is greater than a follow-migration rate threshold for N consecutive sampling periods, the follow-migration relationship is considered valid. The validity of the region correspondence, dominant migration direction, follow-migration relationship, and disturbance exit are used as feeding conditions.

[0051] The state transition record corresponding to the current observation segment is read. The state transition record includes the stage label, disturbance score and transition time mark for each sampling period. When the current observation segment is in a feeding-determinable state, and the number of bait falling trajectories, the area of ​​water surface texture change area, the area of ​​foam area, the area of ​​particle strip area and the number of fish sprint trajectories do not meet the corresponding conditions for the disturbance-dominant state induced by feeding action, the disturbance exit is determined to be established. Not meeting the corresponding conditions means that the above indicators are all less than or equal to their respective thresholds. At the same time, when the feeding interpretation report indicates that the fish are in a continuous feeding stage or a feeding enhancement stage, the current aquaculture unit is deemed to have feeding control conditions. The feeding control conditions include two sub-conditions: the feeding stage is continuous feeding or feeding enhancement and the disturbance exit is established. The feeding interpretation report is obtained. The feeding interpretation report records the feeding stage label, confidence level and corresponding time interval in a structured form.

[0052] Within the same aquaculture unit, if the feeding interpretation reports remain consistent and the feeding control conditions do not change in the opposite direction, the current data interpretation conclusion is maintained. "Consistent" means that the feeding stage labels output in consecutive segments are the same; "no reverse change" means that the two sub-conditions in the feeding control conditions do not change from being met to not being met. When the interpretation direction repeatedly switches between consecutive observation segments, the baseline observation report and stage state switching record are retrieved for verification. Repeated switching of interpretation direction is defined as the feeding stage label changing back and forth between increasing, continuous, and decreasing within N consecutive segments. The baseline observation report is a reference value for fish behavior extracted under static and stable conditions, and the stage state switching record is used to trace back the transition times of each stage and the corresponding index values.

[0053] In this implementation plan, by establishing the feeding control conditions on the basis of satisfying both the feeding phase judgment result and the disturbance exit state, the interference of feed falling, water surface texture disturbance, foam changes, particle stripes and fish sprint trajectory on the control judgment can be reduced, thereby improving the authenticity and consistency of the current feeding control basis of the aquaculture unit; and enhancing the stability and traceability of the feeding judgment report, providing a reliable basis for the generation of feeding start command, single feeding duration parameters and single feeding amount parameters.

[0054] Specifically, the process of generating feeding control parameters based on data interpretation conclusions and correcting the current control process is as follows: When the current aquaculture unit is determined to meet the conditions for feeding control, a feeding start command, a single feeding duration parameter, and a single feeding amount parameter are output to the feeding equipment. The feeding start command includes the equipment number and start timestamp. The single feeding duration parameter is in milliseconds or the number of sampling periods. The single feeding amount parameter is dynamically calculated based on a preset feeding curve or fish density. When the current aquaculture unit is determined not to meet the conditions for continued feeding control, a stop command or a wait command is output to the feeding equipment, and the start time parameter for the next feeding interpretation is generated. Conditions for not meeting the conditions for continued feeding control include the feeding phase entering the decay phase, disturbance regaining dominance, or the confidence level of the feeding interpretation report falling below the confidence threshold. The stop command triggers the equipment to immediately stop feeding, and the wait command triggers the equipment to enter standby mode. The start time parameter for the next feeding interpretation is based on the current time plus the waiting interval, and the interval is set according to the duration of the feeding decay phase or the statistical value of the feeding interval.

[0055] When the proportion of fish moving towards the feeding area or particle distribution area in N consecutive sampling periods is greater than the direction proportion threshold, the regional overlap ratio of fish participating in migration between adjacent sampling periods is greater than the regional continuity threshold, and the main activity area of ​​the fish group and the feeding area maintain a corresponding relationship, the movement is identified as a following migration behavior; when the displacement velocity of the fish group in a single sampling period is greater than the intrusion velocity threshold, the area of ​​the local aggregation area is greater than the agglomeration area threshold, and the movement corresponds to the time when the bait enters the water, the time when the water surface texture changes, or the time when the equipment is activated, the movement is identified as a sprinting intrusion behavior.

[0056] If, within N consecutive sampling periods, the number of sampling periods identified as following migration behavior is greater than the number of sampling periods identified as sprinting and rushing behavior, then the feedback phase is considered a following migration interpretation result; otherwise, the feedback phase is considered a sprinting and rushing interpretation result. The identification condition for following migration behavior is that the fish target maintains its movement towards the material drop area for multiple consecutive sampling periods and the distance from the previous period decreases; the identification condition for sprinting and rushing behavior is that the fish target's speed exceeds the sprint speed threshold and its direction points towards the material drop area within a single sampling period; the number of sampling periods is counted using a cumulative counting method, and if there are multiple targets in each sampling period, the behavior type of the period is determined by the majority behavior or weighted voting.

[0057] In this implementation plan, by distinguishing between following migration behavior and sprinting behavior during the feedback phase, and interpreting the feedback phase based on the statistical results of behavior types within a continuous sampling period, it is possible to separate the true and continuous response of the fish school to the feeding area from the instantaneous sprinting changes caused by the feeding action. This reduces control misjudgments caused by short-term high-speed gathering, disturbances corresponding to equipment actions, and local sprinting behavior, thereby improving the stability, consistency, and reliability of the feeding process.

[0058] Specifically, the process of forming closed-loop samples and modifying the overall control process is as follows: When the feeding interpretation report, equipment execution process, and follow-migration behavior in the feedback phase are consistent, the current feeding process is recorded as a closed-loop consistent sample. When the feeding interpretation report does not match the equipment execution process, or when the number of sampling cycles for sprinting behavior in the feedback phase is greater than the number of sampling cycles for another type of behavior, or when follow-migration behavior is not formed, the current feeding process is recorded as a closed-loop deviation sample. Closed-loop consistent and closed-loop deviation samples are classified and archived, and the corresponding baseline observation reports, phase state switching records, purification observation segments, feeding behavior interpretation reports, and feedback phase behavioral characteristics are associated and stored for continuous verification of the feeding process in the same aquaculture unit. Based on the source of the deviation, closed-loop deviation samples are divided into disturbance misjudgment type deviation samples, execution deviation type deviation samples, and response lag type deviation samples, and the state switching threshold, feeding duration parameter, and single feeding amount parameter are adjusted according to the deviation type.

[0059] Based on closed-loop consistent samples and closed-loop deviation samples, the state switching threshold, direction proportion threshold, following migration rate threshold, and aggregation occupancy judgment threshold for the same aquaculture unit's observation segment are corrected, and the corrected thresholds are written into the parameter record of the current aquaculture unit. After completing multiple feeding cycles in the same aquaculture unit, the closed-loop consistent samples and closed-loop deviation samples formed in each cycle are summarized to generate a feeding closed-loop report for the corresponding aquaculture unit, and the overall control process is corrected: the feeding judgment threshold, disturbance template, stage transition criterion weight, or feeding parameter mapping function are updated using the closed-loop consistent samples and closed-loop deviation samples, respectively; and the stage transition criterion weight is updated based on the temporal characteristics of following migration behavior in the closed-loop consistent samples, so that the stage transition judgment is more consistent with the actual feeding response pattern.

[0060] like Figure 5 The closed-loop sample classification chart shown illustrates the distribution of closed-loop samples after multiple feeding cycles, categorized into closed-loop consistent samples, perturbation misjudgment bias samples, execution bias samples, and response lag bias samples. The area and corresponding percentage of each sector in the chart reflect the proportion of different types of samples in all closed-loop samples. As shown in the figure, the highest proportion of closed-loop consistent samples indicates that there is a high degree of consistency between the feeding behavior interpretation report, the equipment execution process, and the behavior recognition results in the feedback stage during most feeding cycles. The current solution can reliably complete the closed-loop processing from visual observation, disturbance isolation, feeding interpretation to feeding control. A certain proportion of disturbance misjudgment bias samples indicate that in some feeding cycles, there is still overlap between the disturbance induced by the feeding action and the actual feeding response of the fish in time or space, resulting in some observation segments being incorrectly classified into the feeding interpretation process. The proportion of execution deviation bias samples is relatively low, indicating that the feeding equipment can execute according to the feeding start command, single feeding duration parameters, and single feeding amount parameters output by the computer in most cases, but there are still a few cycles in which the equipment action is inconsistent with the expected control parameters. A certain proportion of response lag bias samples also indicate that in some cycles, the fish's following migration behavior and aggregation changes in the feeding area are not formed in time within the predetermined sampling period, resulting in a relatively delayed behavior interpretation result in the feedback stage. The figure shows that this scheme can not only identify the consistency of closed-loop samples, but also further distinguish the sources of deviation. Based on the three different types of deviations—misjudgment of disturbance, execution deviation, and response lag—the scheme can correct the state switching threshold, feeding duration parameter, and single feeding amount parameter, thereby improving the accuracy of judgment and control consistency in the feeding cycle.

[0061] In this implementation plan, by continuously accumulating closed-loop consistent samples and closed-loop deviation samples within the same aquaculture unit, and updating the feeding interpretation threshold, disturbance template, stage transition criterion weight, and feeding parameter mapping function accordingly, the feeding stage interpretation, disturbance identification, and control parameter output during the feeding cycle can better match the actual fish population response patterns of the aquaculture unit, thereby improving the overall control process's adaptability, interpretation consistency, and feeding control reliability.

[0062] Specifically, the second aspect of this invention provides a precision feeding decision system for aquaculture based on visual behavior analysis, applied to a precision feeding decision method for aquaculture based on visual behavior analysis, comprising: an observation acquisition and baseline calibration module, used to simultaneously observe the activity process of fish schools and the action process of feeding equipment within the aquaculture area; the simultaneous observation includes acquiring image sequences by surface and underwater cameras at fixed sampling periods, and simultaneously recording the start-up, operation, and stop status of the feeding equipment and the equipment number, forming visual observation data, and performing uniform scale processing to establish a baseline observation window; and a feeding action-induced disturbance identification module, used to divide the observation stage around a single feeding process and analyze the source of current visual changes. The system analyzes and distinguishes between different observation segments. The distinction is based on comparing the disturbance score with the disturbance threshold to determine whether the current segment is dominated by feeding-induced disturbances or in a state where feeding is possible, thus constructing a set of observation segments for purification behavior. A feeding demand interpretation module identifies fish migration processes based on this set of observation segments and interprets data for the current aquaculture unit. This interpretation is based on a combined chronological judgment of aggregation-contraction ratio, directional consistency coefficient, fish swimming rate in the feeding area, and follow-and-maintain similarity, generating a feeding interpretation report. An execution control and feedback correction module generates feeding control parameters based on the data interpretation conclusions, corrects the current control process, forms a closed-loop sample, and corrects the overall control process. The correction methods for the overall control process include updating the feeding interpretation threshold, disturbance template, stage transition criterion weights, or feeding parameter mapping functions. The correction rules are adaptively adjusted based on the proportion of closed-loop deviation samples in the aggregated samples.

[0063] In this implementation plan, by identifying the fish migration process based on the purification behavior observation fragment set and generating a feeding interpretation report, the generation of feeding control parameters can be based on the behavior interpretation information after disturbance isolation, thereby improving the matching degree between the current feeding start time, single feeding duration and single feeding amount of the current aquaculture unit and the actual feeding needs of the fish.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A precision feeding decision-making method for aquaculture based on visual behavior analysis, characterized in that, Includes the following steps: S1, synchronously observe the activity process of fish and the operation process of feeding equipment in the aquaculture area to form visual observation data, and perform uniform scale processing to establish a baseline observation window; S2, divide the observation phase around the single feeding process, analyze the source of the current visual changes, identify the switching of the current observation segment, and construct a set of observation segments of purification behavior; S3, based on the observation fragment set of purification behavior, identifies the fish migration process, performs data interpretation on the current aquaculture unit, and obtains a feeding interpretation report; S4 generates feeding control parameters based on data interpretation conclusions, corrects the current control process, forms closed-loop samples, and corrects the overall control process.

2. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process of simultaneously observing the activity of fish schools and the operation of feeding equipment within the aquaculture area to generate visual observation data, and then performing uniform scale processing to establish a baseline observation window is as follows: Observations are conducted within the aquaculture unit; feeding equipment action information is simultaneously accessed; image sequences, sampling time information, aquaculture unit identification, camera device identification, feeding trigger information, feeding equipment action information, and post-feeding feedback images are collected and written into the visual observation data; the image sequences and equipment action information are aligned according to the timestamp. Missing frames are marked for sampling moments with missing frames; brightness distribution, contrast distribution, and background regions in the image sequence are standardized. Based on image sequences, a fish target extraction method combining background modeling and inter-frame difference is used to separate the fish movement region. The separated fish movement regions are subjected to connected component screening and morphological correction to obtain the target regions of the fish school. Based on the centroid position, circumscribed region position, and region overlap relationship of the target regions of the fish school at adjacent sampling times, cross-frame correlation tracking is performed to construct fish school movement trajectory segments. Fish school target detection, movement trajectory correlation, and continuous behavior analysis methods are used to characterize the changes in the regional aggregation state, swimming change state, directional migration state, and following response state of the fish school during the observation period. Behavioral features corresponding to the fish school aggregation range, swimming change rate, directional vector discrete value, and following migration rate are extracted respectively. The uniform scale processing of each behavioral feature is linearly normalized. Before the feeding action is initiated, static stability screening is performed on the image sequence. Only if the current observation segment meets the static stability condition is the current observation segment included in the baseline observation window; if the static stability condition is not met, the current observation segment is discarded and the screening is carried out backward to generate the baseline observation report of the current aquaculture unit.

3. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process of dividing the observation phases around a single feeding process and analyzing the source of current visual changes is as follows: Based on the feeding equipment's action information, the visual observation data is divided into three stages: pre-feeding observation, feeding in progress, and post-feeding feedback. Within each stage, changes in fish aggregation, local surface disturbance, bait diffusion, particle settling, and group movement are extracted to form phased observation content. Based on the image sequence and behavioral features within the feeding in progress stage, the source of current visual changes is analyzed. The time window for disturbance occurrence is determined based on the feeding equipment's action information, and significant detection is performed on the aforementioned disturbance components within the time window to form a purified segment. This segment distinguishes between disturbances directly caused by the equipment's actions and the actual feeding behavior of the fish, thereby identifying disturbance components directly caused by the equipment's actions. These disturbance components include the results of bait scattering trajectory recognition, water surface texture change area recognition, foam area recognition, particle stripe area recognition, and fish sprint trajectory recognition.

4. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process for switching and determining the current observation segment is as follows: The disturbance score of the current observation segment is obtained through a fusion calculation method. When the disturbance score is greater than or equal to the disturbance threshold, the current observation segment is switched. When the number of newly appearing bait falling trajectories is greater than the trajectory number threshold, and at least one of the areas of water surface texture change region, foam region, and particle strip region is greater than the corresponding area threshold, and the number of fish sprint trajectories is greater than the sprint trajectory threshold, the current observation segment is determined to be a state dominated by disturbance induced by baiting action. When the number of newly appearing bait falling trajectories is less than or equal to the trajectory number threshold, and the areas of water surface texture change region, foam region, and particle strip region are all less than or equal to the corresponding area threshold, and the proportion of fish moving towards the bait falling area is greater than the direction proportion threshold, and the proportion of fish maintaining the migration relationship towards the bait falling area is greater than the following migration rate threshold, the current observation segment is determined to be a state that can be judged by feeding action. The state switch is only performed when all the judgment conditions for the state dominated by disturbance induced by baiting action or the judgment conditions for the state that can be judged by feeding action are met within the corresponding consecutive N sampling periods. If any statistical item fails to meet the corresponding judgment condition within N consecutive sampling periods, the current state remains unchanged.

5. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process for constructing the set of observation segments of purification behavior is as follows: Continuous observation segments in a state where feeding can be determined are retained, and the corresponding behavioral feature quantities are also retained simultaneously to form a set of purification behavior observation segments. Segments in a state dominated by disturbances induced by feeding actions are not directly included in the feeding interpretation, but are used as equipment action interference segments to update the disturbance template, update the discrimination threshold, update the classifier parameters, or update the mask region library to participate in feedback correction. Continuous segments in the set of purification behavior observation segments are structured and bound according to the breeding unit number, feeding stage information, sampling time information, and equipment number information, and written into the purification observation segment record.

6. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process of identifying the fish migration process based on the set of observation fragments of purification behavior is as follows: Based on image sequences and behavioral features from the observation fragment set of purification behavior, a method combining continuous comparison of sliding analysis windows and stage transition indicators was used to identify the transition process of fish groups from a feeding state to a feeding enhancement stage, from a feeding enhancement stage to a continuous feeding stage, and from a continuous feeding stage to a feeding decline stage. Within each sliding analysis window, aggregation-contraction ratio, directional consistency coefficient, feeding zone companion rate, and following similarity were extracted and jointly judged according to the sampling time order: when each indicator changed from a missing state to a formed state, it was determined that the fish group had transitioned from a feeding state to a feeding enhancement stage. When all indicators remain valid within N consecutive sliding analysis windows, the fish population is determined to have transitioned from the enhanced feeding phase to the continuous feeding phase; when the aggregation-contraction ratio, directional consistency coefficient, and follow-and-maintain similarity successively decrease, and the main activity area of ​​the fish population moves away from the feeding area or particle distribution area, the fish population is determined to have transitioned from the continuous feeding phase to the feeding decline phase. The methods of regional statistics, trajectory correlation, directional distribution analysis, migration connection discrimination and stage comparison analysis are used to extract changes in fish school aggregation range, following status, swimming direction concentration changes, group migration continuity, feeding decline changes and response status after disturbance withdrawal. The data are then arranged in a structured manner according to the sampling time sequence to generate a feeding behavior interpretation report corresponding to the current observation period.

7. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process of interpreting data from the current aquaculture unit to obtain a feeding interpretation report is as follows: The movement status of fish in the current aquaculture unit is determined over N consecutive sampling periods. The state switching record corresponding to the current observation segment is read. When the current observation segment is in a feeding-determinable state, and the number of bait falling trajectories, the area of ​​water surface texture change area, the area of ​​foam area, the area of ​​particle strip area, and the number of fish sprint trajectories do not meet the conditions corresponding to the dominant state of disturbance induced by feeding action, the disturbance exit is determined to be established. At the same time, when the feeding interpretation report indicates that the fish are in a continuous feeding stage or a feeding enhancement stage, the current aquaculture unit is considered to have feeding control conditions, and a feeding interpretation report is obtained. Within the observation segment of the same aquaculture unit, when the feeding interpretation reports are consistent and the feeding control conditions do not change in the opposite direction, the current data interpretation conclusion is maintained. When the interpretation direction repeatedly switches between consecutive observation segments, the baseline observation report and stage status switch record are retrieved again for verification.

8. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process of generating feeding control parameters based on data interpretation conclusions and correcting the current control process is as follows: When the current aquaculture unit is determined to meet the feeding control conditions, the feeding start command, single feeding duration parameters, and single feeding amount parameters are output to the feeding equipment; when the current aquaculture unit is determined not to meet the conditions for continued feeding control, the feeding equipment is output to the feeding equipment to stop or wait, and the next feeding judgment start time parameters are generated. When the proportion of fish moving toward the material drop area or particle distribution area in N consecutive sampling periods is greater than the direction proportion threshold, the movement is identified as following migration behavior. When the displacement velocity of the fish school in a single sampling period is greater than the intrusion velocity threshold, the movement is identified as a sprinting intrusion behavior; If, within N consecutive sampling periods, the number of sampling periods identified as following migration behavior is greater than the number of sampling periods identified as sprinting behavior, then the feedback phase is considered to be a following migration interpretation result; otherwise, the feedback phase is considered to be a sprinting behavior interpretation result.

9. The aquaculture precision feeding decision-making method based on visual behavior analysis according to claim 1, characterized in that: The specific process of forming closed-loop samples and correcting the overall control process is as follows: When the feeding interpretation report, the equipment execution process, and the follow-migration behavior in the feedback phase are consistent, the current feeding process is recorded as a closed-loop consistent sample. When the feeding interpretation report does not match the equipment execution process, or when the number of sampling cycles for sprinting behavior in the feedback phase is greater than the number of sampling cycles for another type of behavior, or when follow-migration behavior is not formed, the current feeding process is recorded as a closed-loop deviation sample. The closed-loop consistent samples and closed-loop deviation samples are classified and archived. The closed-loop consistent samples and closed-loop deviation samples formed in each cycle are summarized to generate a feeding closed-loop report for the corresponding aquaculture unit, and the overall control process is corrected. The feeding interpretation threshold, disturbance template, stage transition criterion weight, or feeding parameter mapping function are updated using the closed-loop consistent samples and closed-loop deviation samples, respectively. The stage transition criterion weight is updated based on the temporal characteristics of follow-migration behavior in the closed-loop consistent samples, so that the stage transition judgment is more in line with the actual feeding response pattern.

10. A precision feeding decision system for aquaculture based on visual behavior analysis, employing the precision feeding decision method for aquaculture based on visual behavior analysis as described in any one of claims 1-9, characterized in that, include: The observation acquisition and baseline calibration module is used to simultaneously observe the activity process of fish and the operation process of feeding equipment in the aquaculture area, generate visual observation data, perform uniform scale processing, and establish a baseline observation window. The feeding action-induced disturbance recognition module is used to divide the observation stage around a single feeding process, analyze the source of the current visual changes, identify the switching of the current observation segment, and construct a set of observation segments of purification behavior. The feeding demand interpretation module is used to identify the fish migration process based on the set of observation segments of purification behavior, perform data interpretation on the current aquaculture unit, and obtain a feeding interpretation report. The execution control and feedback correction module is used to generate feeding control parameters based on data interpretation conclusions, correct the current control process, form closed-loop samples, and correct the overall control process.

Images