CN-121999351-A - Non-invasive monitoring method and system for fish health status

CN121999351ACN 121999351 ACN121999351 ACN 121999351ACN-121999351-A

Abstract

The invention discloses a non-invasive monitoring method and a system for fish health status, which relate to the technical field of image processing and pattern recognition and comprise the following steps: the image acquisition module is used for acquiring continuous image sequence data of fishes in the aquaculture water body, and performing time marking and basic image normalization processing on the continuous image sequence data to generate time sequence image data. According to the invention, unified acquisition time information is introduced into a continuous image sequence, and cross-frame correlation analysis and maintenance identification allocation are performed between adjacent time frames, so that continuous maintenance and track level modeling of the same fish body in multiple time frames are realized, the problem of individual identity confusion caused by shielding and cross swimming in a multiple-fish-body scene is effectively avoided, continuous and stable health state monitoring can be realized on an individual level, and continuous monitoring of the health state of fish facing a complex culture environment is realized on the premise of not contacting the fish body and not introducing an invasive sensor.

Inventors

KE GUOHUA
Ke Chuanduan
DUAN YUANXING
HUANG LE

Assignees

江西豪海绿色农业发展有限公司
九江星祥水产养殖有限公司
瑞昌市洋乐家庭农场

Dates

Publication Date: 20260508
Application Date: 20260202

Claims (9)

1. A fish health status non-invasive monitoring system, comprising: The image acquisition module is used for acquiring continuous image sequence data of fishes in the aquaculture water body, and executing time marking and basic image normalization processing on the continuous image sequence data to generate time sequence image data; The target identification individual keeping module is used for receiving the time sequence image data, executing fish body target identification processing on the time sequence image data, and continuously keeping the image targets of the same fish body in different time frames based on cross-frame correlation analysis to generate fish body individual track data; The health associated image feature construction module is used for receiving individual fish body track data, performing image analysis processing on fish body posture change, motion rhythm change, local form change and spatial distribution change in the individual fish body track data, and generating multidimensional health associated image feature data; The periodic variation characteristic extraction module is used for receiving the multidimensional health associated image characteristic data, performing periodic variation analysis processing on the local area of the fish body image in continuous time frames and generating periodic image characteristic data reflecting the physiological associated state; The feature consistency check health judging module is used for receiving the multidimensional health associated image feature data and the periodic image feature data, executing consistency check processing on the change trend of the different types of image features in the time dimension, and generating fish health state judging data based on consistency check results; The image quality evaluation reliability control module is used for receiving the time sequence image data and the fish health state discrimination data, performing definition evaluation, effective area evaluation and abnormal interference identification processing on the time sequence image data, generating image quality evaluation data, performing reliability control processing on the fish health state discrimination data based on the image quality evaluation data and outputting health state monitoring result data.
2. The non-invasive monitoring system for fish health status according to claim 1, wherein the image acquisition module acquires current frame images in continuous image sequence data from the aquaculture water body according to a preset sampling interval, and writes corresponding acquisition time information into the current frame images to form a time-stamp frame; Performing basic image normalization processing on the time-stamp frame, performing geometric correction processing on the time-stamp frame to obtain a geometric correction frame, performing brightness normalization processing on the geometric correction frame to obtain a brightness normalization frame, performing scale normalization processing on the brightness normalization frame to obtain a scale normalization frame, and performing noise suppression processing on the scale normalization frame to obtain a normalization frame; And carrying out sequence organization on a plurality of continuous normalized frames according to the time sequence of the acquired time information to form time sequence image data.
3. The system of claim 1, wherein the target recognition individual keeping module reads the current frame image and the last frame image of the adjacent time frame according to the acquisition time sequence of the time sequence image data, performs fish body target recognition processing in the current frame image to obtain a current frame fish body target set, and reads the last frame fish body target set in the recognition result corresponding to the last frame image; Extracting a target center position and a target boundary range for each fish body target in the current frame of fish body target set, and extracting a target center position and a target boundary range for each fish body target in the previous frame of fish body target set; Calculating a frame-crossing association cost based on the overlapping proportion of the distance difference of the center position of the target in the current frame fish body target set and the target center position in the previous frame fish body target set and the target boundary range, and determining the association corresponding relation between the current frame fish body target and the previous frame fish body target according to the sequence of the frame-crossing association cost from small to large; distributing a holding identifier for the current frame fish object target according to the association correspondence and forming a cross-frame association result, distributing a new holding identifier for the current frame fish object target which does not acquire the holding identifier, and recording the last frame holding identifier which does not acquire the association correspondence in the current frame as a missing state; writing a target center position and a target boundary range corresponding to each holding mark according to the time sequence of the acquisition time information to form a holding mark track sequence; and the target identification individual keeping module is used for carrying out aggregation organization on all the keeping identification track sequences, generating individual track data of the fish body and outputting the individual track data.
4. The system for monitoring the health state of the fish is characterized in that the health-related image feature construction module reads a target center position and a target boundary range corresponding to the keeping mark every moment according to the acquisition time sequence of the keeping mark track sequence in the individual track data of the fish body as a reference, calculates displacement variation from the target center position at the adjacent moment, calculates movement direction variation from the target center position variation direction at the adjacent moment, calculates boundary scale variation from the target boundary range at the adjacent moment, and forms an original time sequence feature vector at the current moment; Normalizing and combining the displacement variation and the boundary scale variation to form a track variation amplitude sequence, performing statistical analysis processing on the track variation amplitude sequence in the time dimension, calculating the average variation level of the track variation amplitude sequence, and mapping the average variation level to a characteristic variation sensitivity coefficient ; When the original time sequence feature vector corresponding to the first acquisition time information is read, writing the original time sequence feature vector into the original feature moment and generating a last time feature moment vector representation, when the original time sequence feature vector corresponding to the subsequent acquisition time information is read, calculating a difference vector represented by the original time sequence feature vector and the last time feature moment vector, and calculating two norms for the difference vector to obtain a feature change amplitude; according to the time sequence of the acquisition time information, organizing the characteristic variation amplitude corresponding to a plurality of continuous acquisition times to form a characteristic variation amplitude historical sequence, performing statistical distribution analysis processing on the characteristic variation amplitude historical sequence to obtain a minimum variation value and a maximum variation value of the characteristic variation amplitude in the historical acquisition time range, taking the minimum variation value as a normalization lower bound and taking the maximum variation value as a normalization upper bound, and constructing a normalization interval; the characteristic change amplitude and the characteristic change sensitivity coefficient Multiplying to obtain a gating map value, performing upper limit truncation processing on the gating map value, comparing the gating map value with a normalized interval, limiting the gating map value to be the normalized interval lower bound when the gating map value is smaller than the normalized interval lower bound, limiting the gating map value to be the normalized interval upper bound when the gating map value is larger than the normalized interval upper bound, limiting the gating map value to be in the normalized interval, performing linear mapping processing on the gating map value limited to be in the normalized interval, and linearly mapping the gating map value to an interval [0,1] as an adaptive gating update coefficient Taking the mapped numerical value as an adaptive gating update coefficient corresponding to the current acquisition time ; Updating coefficients according to adaptive gating And carrying out weighted structure updating on the characteristic moment vector representation at the previous moment and the original time sequence characteristic vector at the current moment to obtain the characteristic moment vector representation at the current moment, writing the characteristic moment vector representation at the current moment into the corresponding time line position of the characteristic moment to form track time sequence characteristic moment, carrying out convergence organization on the track time sequence characteristic moment formed by each maintenance identification track sequence, and generating multidimensional health associated image characteristic data.
5. The non-invasive monitoring system for fish health status according to claim 1, wherein the periodic variation feature extraction module reads current frame images of adjacent time frames according to the acquisition time sequence corresponding to the multi-dimensional health associated image feature data, and locates local areas of the fish body images in the current frame images based on target boundary ranges in the fish body individual track data to obtain local area image sequences; Calculating the gray average value change quantity of the local area image from moment to moment for the local area image sequence, and organizing the gray average value change quantity corresponding to a plurality of continuous acquisition times according to the time sequence of the acquisition time information to form a local time sequence change sequence; Performing autocorrelation calculation on the local time sequence change sequence to obtain an autocorrelation sequence, determining a time interval corresponding to a main peak position in the autocorrelation sequence as a period length, and calculating a main peak amplitude as a period stability index; And organizing the cycle length and the cycle stability index according to the time sequence of the acquisition time information, generating and outputting the periodic image characteristic data reflecting the physiological association state.
6. The system for monitoring the health status of fish according to claim 1, wherein the feature consistency verification health discrimination module reads corresponding time window data formed by the multi-dimensional health associated image feature data and the periodic image feature data in a historical acquisition time range according to the time sequence of the acquisition time information, reads track time sequence feature moment corresponding to a current time window in the multi-dimensional health associated image feature data from time window to time window, and reads period length and period stability index corresponding to the current time window in the periodic image feature data; in the current time window, performing in-time-window convergence processing on the track time sequence feature moment to obtain a feature convergence vector corresponding to the current time window, and performing in-time-window organization processing on the period length and the period stability index to obtain a period feature vector corresponding to the current time window; Performing consistency similarity calculation processing on the feature convergence vector and the periodic feature vector to obtain consistency similarity corresponding to the current time window; In the historical collection time range, the consistency similarity corresponding to each historical time window is converged to form a consistency similarity set, statistical analysis processing is carried out on the consistency similarity set in the time dimension, the center level of the consistency similarity set is calculated, and the center level of the consistency similarity set is determined as a reference consistency threshold value ; At the determination of the benchmark consistency threshold Then, according to the time sequence of the collected time information, a consistency similarity set corresponding to a historical time window is read, and the change amplitude of the consistency similarity corresponding to the adjacent historical time window is calculated to form a consistency similarity change sequence; Performing statistical analysis processing on the consistency similarity change sequence in the time dimension, calculating the average change level of the consistency similarity change sequence, and mapping the average change level of the consistency similarity change sequence into a threshold adaptive coefficient ; Based on a benchmark consistency threshold within a current time window Establishing an initial consistency threshold, reading the consistency similarity corresponding to the previous time window, and calculating the consistency similarity corresponding to the previous time window relative to a reference consistency threshold Is used for matching the deviation amount with the threshold adaptive coefficient Multiplying to obtain a threshold correction amount, and adding the threshold correction amount to the reference consistency threshold Obtaining a self-adaptive consistency threshold corresponding to the current time window; Comparing the consistency similarity corresponding to the current time window with the self-adaptive consistency threshold corresponding to the current time window, judging the current time window as a consistency meeting state when the consistency similarity is not smaller than the self-adaptive consistency threshold, and judging the current time window as a consistency not meeting state when the consistency similarity is smaller than the self-adaptive consistency threshold; And organizing the consistency meeting state and the consistency failing state corresponding to each time window according to the time sequence of collecting the time information, and generating fish health state discrimination data.
7. The non-invasive monitoring system for fish health status according to claim 1, wherein the image quality evaluation reliability control module reads a normalized frame sequence corresponding to a current time window in the time sequence image data time-by-time window according to the time sequence of collecting the time information, and synchronously reads fish health status discrimination data corresponding to the current time window output by the feature consistency verification health discrimination module; In the current time window, performing definition evaluation processing on the normalized frame sequence frame by frame, calculating definition evaluation values of all normalized frames, and converging the definition evaluation values in the time window to obtain a definition evaluation result corresponding to the current time window; In the current time window, performing effective area evaluation processing on the normalized frame sequence frame by frame, calculating an effective area evaluation value based on the coverage area proportion of the identifiable fish object in the normalized frame, and converging the effective area evaluation value in the time window to obtain an effective area evaluation result corresponding to the current time window; In the current time window, carrying out abnormal interference identification processing on the normalized frame sequence frame by frame, identifying abnormal frames caused by shielding, strong reflection or noise mutation, and counting the occurrence proportion of the abnormal frames in the time window to obtain an abnormal interference evaluation result corresponding to the current time window; combining the definition evaluation result, the effective area evaluation result and the abnormal interference evaluation result to generate image quality evaluation data corresponding to the current time window; According to the image quality evaluation data corresponding to the current time window, performing credibility control processing on the fish health state judgment data corresponding to the current time window, when the image quality evaluation data meets the quality effective condition, keeping the fish health state judgment data as an effective judgment result, and when the image quality evaluation data does not meet the quality effective condition, marking the fish health state judgment data as a low credibility state; And organizing the effective discrimination results and the low-reliability state discrimination results corresponding to the time windows according to the time sequence of collecting the time information, and generating health state monitoring result data.
8. The system of claim 1, wherein the image acquisition module, the target identification individual keeping module, the health-related image feature construction module, the periodic variation feature extraction module, the feature consistency verification health judgment module and the image quality evaluation reliability control module all execute time window division based on the unified acquisition time information, and complete data reading, processing and outputting in the respective corresponding time windows; When the characteristic consistency verification health judging module generates fish health state judging data, carrying out continuous constraint processing on consistency meeting states and consistency failing states corresponding to adjacent time windows so as to eliminate transient state jump caused by single time window abnormality; And when the image quality evaluation reliability control module outputs the health state monitoring result data, the low reliability state judgment result is subjected to delay confirmation processing in the time dimension, and the health state monitoring result data of the corresponding time window is marked as a low reliability output state only when a plurality of continuous time windows are judged to be in a low reliability state, so that the stability and reliability of the overall health state monitoring result are improved.
9. A method for application to a non-invasive monitoring system for fish health status according to any of claims 1-8, characterized by the steps of: Acquiring continuous image sequence data of fish in a culture water body, performing time marking and basic image normalization processing on the continuous image sequence data, and generating time sequence image data organized according to the acquisition time sequence; Based on time sequence image data, executing fish body target identification processing in adjacent time frames, executing cross-frame correlation analysis according to spatial position relation and boundary overlapping relation of fish body targets in different time frames, continuously keeping the image targets of the same fish body in different time frames, and generating fish body individual track data; Extracting a target center position and a target boundary range from moment to moment according to the acquisition time sequence of each maintenance identification track sequence in individual track data of the fish body, constructing an original time sequence feature vector, introducing a feature change sensitivity coefficient and a self-adaptive gating update coefficient based on the difference relation between the original time sequence feature vector and the representation of the last time sequence feature moment vector, performing weighting structure update on the feature moment to form a track time sequence feature moment, and converging tissues of the track time sequence feature moment corresponding to each maintenance identification to generate multidimensional health associated image feature data; Based on the multidimensional health associated image feature data, positioning a local area of the fish body image in a continuous time frame, performing periodic analysis processing on time sequence changes of the local area image, extracting period length and period stability indexes, and organizing according to the acquisition time sequence to generate periodic image feature data reflecting the physiological associated state; In a historical acquisition time range, based on multidimensional health-related image feature data and periodic image feature data, calculating consistency similarity in each time window, performing statistical analysis on distribution features of the consistency similarity in a time dimension, determining a reference consistency threshold value and a threshold value self-adaptive coefficient, performing self-adaptive correction on the reference consistency threshold value according to the consistency similarity of the last time window in a current time window to form a self-adaptive consistency threshold value corresponding to the current time window, and generating fish health state discrimination data based on a comparison result of the consistency similarity and the self-adaptive consistency threshold value; And in each time window, performing quality evaluation processing on the image definition, the coverage condition of the effective area and the abnormal interference condition based on the time sequence image data to generate image quality evaluation data, performing credibility control on the fish health state discrimination data according to the image quality evaluation data to form health state monitoring result data organized according to the acquisition time sequence and outputting the health state monitoring result data.

Description

Non-invasive monitoring method and system for fish health status Technical Field The invention relates to the technical field of image processing and pattern recognition, in particular to a non-invasive monitoring method and system for fish health status. Background Along with the continuous improvement of the large-scale and intensive level of aquaculture, the real-time monitoring of the health state of fish becomes an important technical link for guaranteeing the aquaculture safety, reducing the disease risk and improving the aquaculture benefit. In the actual cultivation process, the health state of fish is usually closely related to the movement behavior, posture change, physiological rhythm and spatial distribution characteristics in the water body. Therefore, how to continuously, stably and reliably monitor the health state of fish on the premise of not interfering the normal growth activity of fish becomes one of the technical problems to be solved in the field of aquaculture. The existing fish health state monitoring mode mainly comprises manual inspection, water quality parameter detection and biosensor detection. The manual inspection is strong in subjectivity and high in cost, continuous monitoring is difficult, the health state of the individual fishes cannot be directly reflected by water quality parameter detection, and most of the biological sensors are contact type or invasive type, so that stress response is easy to trigger. The existing non-invasive method based on images is independent of single-frame or short-time sequence analysis, continuous maintenance and behavioral evolution modeling of fish individuals are difficult to achieve, fixed threshold judgment is often adopted, comprehensive consideration of periodic physiological characteristics, characteristic time sequence consistency and image quality is lacked, and therefore the health judgment stability and reliability are insufficient, and the method is difficult to adapt to complex culture environments. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a non-invasive monitoring method and system for the health state of fish. In order to achieve the aim, the invention adopts the following technical scheme that the fish health state non-invasive monitoring system comprises an image acquisition module, a time sequence image processing module and a control module, wherein the image acquisition module is used for acquiring continuous image sequence data of fish in a culture water body, and executing time marking and basic image normalization processing on the continuous image sequence data to generate time sequence image data; the device comprises a target identification individual keeping module for receiving time sequence image data, a periodic variation characteristic extraction module for receiving the time sequence image data, performing fish object target identification processing on the time sequence image data, continuously keeping image targets of the same fish object in different time frames based on cross-frame correlation analysis to generate fish object individual track data, a health correlation image characteristic construction module for receiving the fish object individual track data, performing image analysis processing on fish object posture variation, motion rhythm variation, local form variation and spatial distribution variation in the fish object individual track data to generate multi-dimensional health correlation image characteristic data, a periodic variation characteristic extraction module for receiving the multi-dimensional health correlation image characteristic data, performing periodic variation analysis processing on local areas of the fish object image in the continuous time frames to generate periodic image characteristic data reflecting physiological correlation states, a characteristic consistency verification health discrimination module for receiving the multi-dimensional health correlation image characteristic data and the periodic image characteristic data, performing consistency verification processing on variation trends of different types of image characteristics in the time dimension, generating fish object health state discrimination data based on consistency verification results, an image quality evaluation reliability control module for receiving the time sequence image data and the time sequence health state data, performing quality evaluation, effective evaluation and abnormal state discrimination processing on the time sequence health state data, generating image quality evaluation data, performing credibility control processing on the fish health state discrimination data based on the image quality evaluation data, and outputting health state monitoring result data. As a further description of the above technical solution: The image acquisition module acquires current frame images in continuous image sequence data from the aquaculture water