An underwater intelligent recognition and diagnosis system for leopard coral grouper

Abstract

This invention discloses an underwater intelligent identification and diagnostic system for leopard-gill sea bass, relating to the field of aquaculture technology. It includes an optical feature analysis module, a particle aggregation detection module, a feature separation and comparison module, a transmission compensation and correction module, and a dynamic reliability control module. The optical feature analysis module extracts the temporal evolution features of brightness gradient, edge sharpness, and texture details in the perceived image, constructs a stable optical transmission reference band, and extracts the abrupt change amplitude at its end to form a mutation marker. This invention introduces a breathing-like adjustment structure and a slow-release channel to achieve rhythmic control and delayed output of low-reliability identification results, avoiding continuous transmission of misjudgments. After image quality recovery, the delayed information is released smoothly, ensuring the integrity and continuity of the identification results and improving the identification stability and health diagnosis reliability in disturbed underwater environments.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, specifically to an underwater intelligent identification and diagnostic system for leopard-gill spiny perch. Background Technology

[0002] The Leopard Gill Spinach Underwater Intelligent Identification and Diagnosis System refers to a technical system that continuously acquires information about fish activity, body surface characteristics, and behavioral states in actual Leopard Gill Spinach aquaculture waters by deploying underwater sensing and visual detection devices. Under the constraints of underwater environmental conditions, the system performs structured analysis and state discrimination on the collected visual detection information and multi-source sensing data, thereby achieving individual identification, physiological abnormality recognition, and health status diagnosis of Leopard Gill Spinach. Based on real-time underwater sensing and visual detection, the system transforms changes in fish morphology, body surface texture features, swimming posture, and group distribution into interpretable state characteristic information. Through an intelligent analysis mechanism, it comprehensively judges abnormal signs that deviate from normal aquaculture conditions, transforming the observation behavior dominated by human experience in the aquaculture process into a continuous, objective, and quantifiable underwater intelligent visual diagnosis process. This provides technical support for the refined aquaculture management and early health risk identification of Leopard Gill Spinach.

[0003] Existing technologies have the following shortcomings: Under current conditions, the underwater intelligent identification and diagnosis system for leopard-gill sea bass typically assumes that the spatial distribution of suspended particles in the water is relatively uniform or changes slowly, and that the optical transmission environment within the sensing area is considered continuous and stable. However, in actual aquaculture scenarios, underwater suspended particles may vortex-like aggregate due to local water flow disturbances within a short period, causing sudden changes in the optical transmission characteristics within the sensing area. Because existing technologies struggle to promptly identify such abrupt changes in the optical environment and effectively correct the sensing results, the system continues to output seemingly stable and internally consistent fish condition assessment information during this stage, even though these assessments deviate from the actual fish condition. Without intuitive anomaly alerts, aquaculture managers who rely on this distorted diagnostic information for long-term aquaculture control decisions are prone to gradually amplifying management deviations, ultimately leading to a loss of control over the health of the leopard-gill sea bass population when anomalies become concentrated.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an underwater intelligent identification and diagnostic system for leopard gill spiny perch to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an underwater intelligent identification and diagnostic system for leopard-gill sea bass, comprising an optical feature analysis module, a particle aggregation detection module, a feature separation and comparison module, a transmission compensation and correction module, and a dynamic reliability control module. The optical feature analysis module extracts the temporal evolution features of brightness gradient, edge sharpness and texture details in the perceived image, constructs an optical transmission stable reference band, and extracts the abrupt change amplitude at its end to form a mutation marker. The particle aggregation detection module, based on mutation markers, performs temporal aggregation analysis on the particle density texture within the sensing area, extracts the circumferential stretching trajectory, generates a vortex aggregation pointing band, and identifies candidate aggregation center regions at its tail. The feature separation and comparison module constructs a dual-channel sampling comparison surface in the candidate region of the aggregation center, maps the fish body contour features and background transmission texture, generates a feature separation comparison frame, and extracts the transmission distortion factor. The transmission compensation and correction module performs inverse compensation on the separation feature comparison frame based on the transmission distortion factor, generates the distortion-free recognition input frame, and extracts the confidence anchor point sequence. The dynamic credibility control module, based on the credibility anchor point sequence, activates the breathing vortex gate control mechanism to execute the periodic opening and closing of the identification entrance, delays the output of low credibility judgments to the shadow relief channel, and restores the instant output when the credibility recovers, thereby controlling the continuous transmission of identification errors.

[0007] Preferably, the mutation marker line formation steps are as follows: The brightness gradient, edge sharpness, and texture details in continuously acquired underwater images are extracted frame by frame to generate a cluster of temporal evolution feature curves. Select time segments with smaller fluctuation ranges and smoother curve changes from the cluster of time evolution characteristic curves, and combine them to form an optical transmission stable reference band. Multi-feature difference extraction is performed on the end image frame of the optical transmission stable reference band to form a multi-feature difference distribution map; In the multi-feature difference distribution map, spatial focus areas are extracted, and pixel blocks are sorted according to brightness jump rate, edge break density and texture rhythm interference degree. Feature patches with large change amplitude and concentrated spatial range are extracted and encoded to generate mutation markers.

[0008] Preferably, during the generation of mutation markers, the spatial location, temporal location, and multidimensional variation amplitude of the feature patches are combined and encoded as the initial identification basis for the mutation state.

[0009] Preferably, the process for identifying candidate regions for cluster centers is as follows: Starting with mutation markers, a sequence of grain density texture images from multiple time frames in the perceptual region is obtained, texture density field information is extracted, and time-ordered texture clustering is performed. Based on the texture convergence trend, the spatial stretching trajectory is extracted from the mutation marker line to construct the vortex convergence direction zone of the circumferential stretching trend; In the tail region of the vortex aggregation pointing zone, the temporal evolution characteristics of the texture density field are traced to identify candidate regions of aggregation centers with density clustering states. Based on the spatial focusing characteristics of the candidate regions of the aggregation center, the mutation markers, vortex aggregation pointing bands and candidate regions of the aggregation center are summarized to form the evolution path of the perturbation structure.

[0010] Preferably, the stretching trajectory of the vortex aggregation pointing band exhibits a circumferential extension in spatial structure and forms a continuous arc structure along the direction of the maximum change gradient of the particle density field. The candidate region of the aggregation center is located at the tail of the vortex aggregation pointing band. In the last frame of the image sequence, it is an image region where the texture density continuously increases, the change slope gradually slows down, the boundary region is blurred, and the brightness gradient tends to be flat. The density evolution rhythm, boundary continuity, and spatial center stability are jointly analyzed and extracted.

[0011] Preferably, the transmission distortion factor extraction process is as follows: A dual-channel sampling control surface is constructed within the candidate region of the aggregation center to map the fish body contour features to the main body layer, and contour overlay extraction is performed in multiple consecutive image frames. A background transmission layer is established in the image channel adjacent to the main layer, and background transmission texture information in the same area is collected to form an upper and lower layered structure. The main body layer and the background layer are aligned and analyzed to form a separation feature comparison frame, and information on structural offset, morphological curvature and texture misalignment is extracted. Based on the difference trajectory structure in the separation feature comparison frame, the cumulative offset information of the image structure is extracted to generate a transmission distortion factor to describe the amplitude and direction of the disturbance.

[0012] Preferably, the transmission distortion factor includes the displacement gradient of the brightness channel, the periodic shift amplitude of the texture channel, and the deformation trend of the edge sharpness. It is used to simultaneously express the spatial offset state of the fish body contour features and the perturbation direction of the background transmission texture, and to keep the spatial reference frame of the image frame unchanged during the generation process, so as to ensure accurate compensation and rhythm restoration of the subsequent image structure.

[0013] Preferably, the steps for extracting the credibility anchor sequence are as follows: Based on the transmission distortion factor, the fish body contour structure region in the separation feature comparison frame is located and processed, and the contour rhythm is continuously unfolded according to the distortion direction and deformation amplitude. While unfolding the fish body outline, the background transmission texture area is spatially compressed and rolled back according to the transmission distortion factor to restore the background structure to the state before the disturbance. After completing the fish outline unfolding and background transmission regression, multiple feature block parameters are extracted based on the degree of structure restoration, a set of credibility anchor points is constructed, and a credibility anchor point sequence is generated. Image frames with structure unfolding, texture regression, and credibility anchor point sequences are used as input frames for distortion removal recognition and then fed into the subsequent recognition process.

[0014] Preferably, each anchor point in the confidence anchor point sequence corresponds to a feature structure block in the main body, fins, eyes, or abdomen of the fish in the image, and multidimensional parameters are generated by combining edge sharpness recovery value, brightness continuity index and background compression smoothness index to represent the recognizability of the structure block.

[0015] Preferably, based on the confidence anchor point sequence, a breathing vortex gate control mechanism is activated to periodically open and close the identification entrance, delaying the output of low confidence judgments to the shadow relief channel, and resuming output when confidence recovers. The steps are as follows: Based on the confidence anchor point sequence, the candidate region of the cluster center is classified into levels to generate anchor point confidence gradient bands; Based on the anchor point confidence gradient band, a periodic micro-opening and closing sensing and control path is constructed at the identification entry point to form a breathing-like state evolution structure. When the entry point is closed, a delayed release structure channel is constructed to receive low-confidence judgment information and maintain the original structure state. As the credibility index recovers, the control recognition entry is opened, and the information in the slow release channel is released frame by frame, so that the recognition results are output naturally in chronological order.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention, by constructing a dynamic recognition control structure, enables the perception process of fish images in underwater disturbed environments to possess rhythmic regulation capabilities. By introducing a periodically opening and closing breathing-like adjustment structure at the entrance of the recognition channel, the recognition results in low-confidence areas are processed in an orderly manner, avoiding the continuous output of erroneous judgments under abnormal perception conditions. This reduces the interference of misjudgment information on subsequent aquaculture decisions and improves the image perception process's responsiveness to local disturbance changes.

[0017] This invention achieves a stable transition of recognition results during image quality fluctuations by establishing a gradual release channel driven by image confidence gradients. After the perturbation subsides, the previously delayed output is released in a temporally continuous manner, maintaining the overall rhythm balance of the recognition process. This approach improves the integrity and continuity of diagnostic information, allowing the output to seamlessly connect during the period of perceived quality recovery, thus enhancing the reliability of the system in performing health status recognition tasks in complex underwater environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a schematic diagram of the modules of an underwater intelligent identification and diagnosis system for leopard-gill spiny perch according to the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] This invention provides, for example Figure 1 The underwater intelligent identification and diagnostic system for leopard-gill sea bass, as shown, includes an optical feature analysis module, a particle aggregation detection module, a feature separation and comparison module, a transmission compensation and correction module, and a dynamic reliability control module. The optical feature analysis module extracts the evolution features of brightness gradient, edge sharpness and texture details in the underwater sensing image in the time dimension, generates an optical transmission stable reference band, and extracts the abrupt change amplitude at the end of the optical transmission stable reference band according to the change amplitude to form abrupt change marker. To effectively obtain the optical transmission stability state from continuously acquired underwater images and identify potential optical disturbances, it is necessary to rely on three typical visual information types—brightness gradient, edge sharpness, and texture detail—in the perceived image. This involves extracting evolutionary features over time to construct stable image features and extract a mutation index. This process is implemented through the following steps: For each frame of the underwater sensing image, while ensuring continuous temporal acquisition, frame-by-frame information is extracted sequentially for brightness gradient, edge sharpness, and texture detail. Brightness gradient extraction primarily depicts the overall trend of illumination distribution changes in the image; edge sharpness describes the clarity of boundary contours within the image; and texture detail reflects the structural complexity and microscopic feature continuity of local image regions. In this step, the spatial distribution curves of the three types of features in consecutive image frames are recorded, and a set of temporal evolution feature curves is generated in chronological order. Each curve reflects the dynamic change trend of the corresponding feature in the sensing image. This set of curves constitutes the basic data set for subsequent analysis of optical stability, providing a clear temporal evolution basis for continuity judgment.

[0022] Based on the obtained cluster of temporal evolution characteristic curves, a representative time window is selected sequentially, containing several consecutive image frames. Within this time window, the brightness gradient curve, edge sharpness curve, and texture detail curve are evaluated for trend consistency. The time segment with the smallest fluctuation range and the most stable curve slope is selected, and its corresponding set of image frames is used as a candidate segment for optical transmission stability. Image frames within this candidate segment are superimposed frame by frame to generate a stable image feature map. By fusing the pixel-level brightness distribution, edge contour extension, and texture repetition rhythm in the image, an optical transmission stable reference image is obtained. Then, based on this reference image, its corresponding time window is selected as a stable reference segment, and all image frames within the stable segment are combined in chronological order to form an optical transmission stable reference band, which serves as the benchmark reference for subsequent mutation identification.

[0023] Based on a constructed stable optical transmission reference band, a weighted trend analysis of the change trends of its terminal image frames is performed to identify whether abrupt changes in the current optical environment have occurred due to external disturbances. This analysis quantifies and extracts the change amplitudes of three types of temporal evolution curves—brightness gradient, edge sharpness, and texture detail—at the terminal frame of the reference band, obtaining a multi-feature difference distribution map between this terminal frame and the middle frames of the reference band. Subsequently, this difference map undergoes trend accumulation processing to comprehensively determine the spatial directionality and temporal abruptness of image feature mutations. When the three types of differences show a synchronous increase in the temporal dimension and a focusing and converging trend in spatial distribution, it can be determined that the current terminal image of the reference band has deviated from the stable optical transmission state, thereby identifying the specific value range of the mutation amplitude. This mutation amplitude reflects the strength of the current optical disturbance and the degree of damage to the visual structure in the perceived image.

[0024] Based on the specific range and spatial distribution characteristics of the abrupt change amplitude, several key feature regions are extracted from the final image frame of the optical transmission stability reference zone to locate the core region where the abrupt change occurs. Within these key regions, pixel blocks are ranked by intensity based on a composite change index of brightness jump rate, edge break density, and texture rhythm interference degree. The feature patch with the largest change amplitude and the most concentrated spatial range is selected as the abrupt change feature focal area. The spatial location, temporal location, and multidimensional change amplitude of this focal area are combined and encoded to form a abrupt change marker. This abrupt change marker records the core information of the transition from the optical transmission stability state to the abrupt change state, serving as the entry index point for subsequent image processing steps. This ensures that subsequent perception region analysis is based on the actual optical abrupt change starting point, while maintaining the consistency and logical closure of the response between the perception process and environmental disturbances.

[0025] The particle aggregation detection module, based on mutation markers, performs temporal aggregation analysis of particle density texture within the current sensing area, extracts the circumferential stretching trajectory guided by mutation markers, generates a vortex aggregation pointing band, and identifies candidate aggregation center regions at the tail of the vortex aggregation pointing band. A stable optical transmission reference band is obtained through preprocessing and solidified into a mutation marker. Based on the mutation marker, the dynamic evolution of particle density texture in the current sensing area is tracked and analyzed to reveal the local particle aggregation behavior triggered by mutations and its spatial orientation. A vortex structure guidance path is then constructed to locate the aggregation core of the perturbed region, providing a basis for subsequent feature decoupling. The specific implementation steps are as follows: Using the spatial location and temporal index marked by the mutation marker as the starting point, a sequence of particle density texture images of the corresponding sensing region is acquired across multiple consecutive time frames following the mutation. In these image frames, relying on the joint sensing features derived from previously extracted brightness gradients, edge sharpness, and texture details, texture density field information of particle distribution is extracted frame by frame. This information reflects the spatial microscopic distribution of particles in the water body and the local concentration variation trend. Subsequently, using the mutation marker as the reference center, temporally ordered texture clustering processing is performed on each frame after the mutation, extracting the particle density reconstruction phenomenon around the mutation point. In consecutive frames, if particles exhibit a density enhancement trend converging from the periphery to the center, it indicates that local disturbance-induced aggregation behavior is forming in this region. This inter-frame aggregation pattern is expressed through the superposition of texture density changes, enabling the mutation marker not only to indicate the starting point of the disturbance but also to guide the dynamic capture of the subsequent spatial focusing process.

[0026] Based on the trend trajectory of texture aggregation information in the time series, starting from the mutation marker line, the spatial stretching direction of particle density change is tracked, and its propagation path within the sensing area is extracted. This path exhibits a circumferential stretching pattern in the spatial dimension, manifested as an arc-shaped or rotating trajectory extension centered on the mutation marker line, with its direction consistent with the direction of the maximum change gradient of the particle density field in the image. In this process, by calculating the spatial movement center of aggregation density in consecutive time frames, a disturbance diffusion path caused by mutation can be delineated. This path is spatially represented as a ring-shaped arc, forming a clear texture density accumulation trajectory in the image. Connecting this circumferential stretching trajectory continuously constructs a vortex aggregation pointing band. This pointing band not only reflects the diffusion trend of disturbed particles but also constitutes a structural guidance path from the mutation center to the aggregation core, allowing the disturbance structure to gradually reveal overall vortex characteristics from local changes.

[0027] Based on the established vortex aggregation pointing band, the temporal evolution characteristics of the texture density field are further tracked in its tail region to capture the density clustering region formed by the stretching, agglomeration, and eventual convergence of particles. In the final time frame of the image sequence, combined with the end position of the pointing band, convergence blocks with continuously increasing texture density and gradually decreasing slope are searched in a multi-dimensional scanning manner within the spatial range to identify image regions with a particle deposition trend. This region is typically characterized by highly concentrated texture density, relatively blurred boundaries, and a gentle brightness gradient, making it a visually focused area. Among multiple candidate convergence blocks, by jointly comparing the density evolution rhythm, boundary continuity, and spatial center stability, the region with the most typical characteristics of change rate and spatial focus is selected as the final candidate region. Its spatial location, temporal index, and focusing trend information of the three types of perceptual features are recorded to constitute the core spatial range for subsequent processing.

[0028] By utilizing the spatial focusing characteristics and texture density clustering state of the candidate aggregation center region, the overall structure of the vortex aggregation pointing band is summarized, clarifying the complete path of the mutation effect from the starting point to the ending point. This path, with the mutation marker line as the source, extends along the circumferential stretching trajectory and finally converges to the candidate aggregation center region, forming a particle perturbation structure evolution path with temporal evolution logic, spatial circumferential guidance, and density focusing trend. In this process, the mutation marker line not only completes the starting point calibration function but also serves as the source beacon of the perturbation diffusion path, providing directional guidance for the spatial analysis of the entire sensing area. At the same time, the vortex aggregation pointing band transmits local perturbation to a wider area through its spatial stretching state, and its tail aggregation center candidate region bears the final manifestation area of ​​the perturbation aggregation effect. The three form a complete visual structure transformation chain from sudden perturbation to spatial aggregation, laying the foundation for the subsequent decoupling of fish body features and background perturbation.

[0029] The feature separation and comparison module constructs a dual-channel sampling comparison surface within the candidate region of the aggregation center, maps the fish body contour features to the main layer of the dual-channel sampling comparison surface, extracts the background transmission texture within the same region, constructs a feature separation comparison frame, and generates a transmission distortion factor in the feature separation comparison frame. To ensure effective differentiation between fish contours and background transmission information under interference conditions, and to provide clear structural support for subsequent image correction and recognition input, a dual-channel sampling control surface with a hierarchical sampling structure can be further constructed within the candidate region of the aggregation center. This control surface can handle the spatial decoupling processing of fish and background information, forming a corresponding mapping between fish contour features and background transmission texture, and extracting the transmission structure changes caused by disturbances for subsequent transmission compensation processing. The specific implementation steps are as follows: Within the identified candidate aggregation center region, a set of sampling sections matching the fish's contour direction is set based on the image structure direction guided by the vortex aggregation direction band, ensuring that the sampling direction is consistent with the interference diffusion path. In this region, relying on the relative spatial stability of the fish structure in consecutive image frames, the main outer contour structure of the fish is extracted by analyzing the brightness boundaries, edge sharpening areas, and texture arrangement direction of the identifiable parts of the fish. This structure is then mapped to the main layer in the set sections, forming the structural baseline of the main layer. This main layer is used to centrally represent the main expression positions of the fish features in image space and is the core reference surface for subsequent comparison surface construction. Simultaneously, to ensure the integrity of the fish structure mapping, contour features should be superimposed and extracted in multiple consecutive frames to form a feature shape with a certain temporal continuity, thereby reducing the impact of structural jumps caused by instantaneous disturbances.

[0030] After mapping the fish body contour structure, background transmission texture information was acquired from the same time frame and in the same region within the same candidate aggregation center area, based on the image channel adjacent to the main layer, and a background layer parallel to the main layer was established. This background layer mainly carries information not belonging to the fish body structure, including the natural light transmission pattern in the water, the scattering structure after particle distribution, and texture noise features in the environmental background. By maintaining spatial structural consistency between the background layer and the main layer, a hierarchical structure was formed in the entire control plane, with the upper layer being the fish body contour main layer and the lower layer being the background transmission layer. The two are independent but spatially corresponding. This structure helps to clearly separate the fish body from the background in areas with interference, constructing a contrastive structural relationship between the fish body contour and background information, and providing a clear interface for subsequent difference analysis and interference factor extraction.

[0031] Based on the constructed dual-channel sampling control surface, the fish contour layer and the background transmission layer are aligned and analyzed, and then combined into a unified image frame to form a separated feature control frame. This control frame preserves the spatial distribution characteristics of the fish structure while also carrying the transmission expression of the background texture at the same location. In this process, the structural offset, morphological curvature, and texture misalignment information between the two are extracted by the differences in structural dissimilarity, brightness contrast amplitude, and texture matching degree between corresponding pixels. These differences reflect the changes in the transmission path in the current aquatic environment from a spatial dimension, and also indirectly characterize the degree of interference of vortex disturbance on image perception. As the perturbation amplitude of the background transmission texture in multiple time frames is continuously superimposed, the feature differences between the fish and the background in this control frame will gradually converge into a unified deformation trend. This trend runs through the entire control frame area, forming a perturbation trajectory structure with strong continuity.

[0032] Based on the spatial distribution characteristics of the differential trajectory structure, structural deformation trends are extracted in the end region of the separation feature comparison frame to obtain the cumulative offset information of the current transmission interference's impact on the image structure, and a transmission distortion factor is generated accordingly. This factor is used to express the amplitude and direction of the background transmission information's disturbance in the image structure, while also reflecting the offset state and deformation direction of the fish contour in the structural space. The transmission distortion factor typically includes information from multiple dimensions, such as the displacement gradient of the brightness channel, the periodic shift amplitude of the texture channel, and the deformation trend of edge sharpness. This information is combined in a joint expression to describe the image structure separation degree between the fish and the background caused by transmission anomalies. In the process of forming the transmission distortion factor, the spatial reference frame of the original image frame should be preserved so that when the factor is used to perform transmission compensation processing, the fish structure can be accurately restored to its normal rhythm state, and the background disturbance structure can be compressed simultaneously to achieve preprocessing correction of the perceived image.

[0033] The transmission compensation and correction module performs inverse transmission compensation processing on the separation feature comparison frame based on the transmission distortion factor, rhythmically unfolds the fish body contour features, compresses and backs down the background transmission texture according to the transmission distortion factor, generates a distortion-free recognition input frame, and generates a confidence anchor point sequence in the distortion-free recognition input frame. To achieve structural restoration of the fish's outline features and effective convergence of background interference information, inverse transmission compensation processing is performed on the feature separation comparison frame based on the transmission distortion factor. By adjusting the directionality of the distorted structure in the image content and restoring the rhythm of the spatial structure, the continuous unfolding of the fish's outline and the compression and regression of the background transmission texture are achieved, thereby generating a distortion-free recognition input frame with structural clarity and recognition continuity. Anchor point information for subsequent recognition confidence control is simultaneously extracted from this input frame. The specific implementation steps are as follows: Using the generated transmission distortion factor as the operational benchmark, the region where the fish contour structure is located in the separation feature comparison frame is processed according to the distortion direction and deformation amplitude described by the factor, clarifying the specific deformation state of the fish contour in the image space of the current frame. Based on this, spatial unfolding processing is performed sequentially on the node positions in the contour shape along the structural direction indicated by the transmission distortion factor. By adjusting the stretching direction and deformation amplitude of the image content point by point, the fish contour is restored from a state of local misalignment and rhythmic jumps to a structurally continuous and rhythmically balanced unfolded form. This processing is guided by the principle of temporal continuity, ensuring a natural transition relationship between the postures of the fish structure in consecutive frames and avoiding structural fragmentation or jitter. During this unfolding process, the focus is on the consistency of the curvature of the contour boundary lines and the extension of the principal axis direction, so that the contour can retain the dynamic information of the fish posture during the restoration process, providing a more stable feature structure foundation for subsequent recognition operations.

[0034] While completing the fish outline unfolding process, spatial compression and rollback operations are performed on the background transmission texture region in the same separation feature comparison frame, based on the transmission distortion factor. The main goal of this operation is to orderly compress the texture stretching, structural stacking, and local enhancement regions caused by perturbation according to the original background structure direction, causing the background texture structure to roll back to its state before the perturbation occurred in spatial distribution. During the compression and rollback process, the pixel structure is rolled back step by step according to the texture period offset direction in the transmission distortion factor, the gradient distribution of the brightness channel, and the local texture trend, so that the stretched area is restored to its original scale in image space and its structural interference on the fish outline region is reduced. This compression processing also needs to preserve the natural water texture features and illumination distribution features in the background to ensure that the processed image still has the ability to express real background information, thereby forming an effective environmental structure reference during the recognition process.

[0035] After completing the fish outline unfolding and background transmission compression operations, a dedistorted image with reconstructed spatial structure and balanced visual content is generated. To use this image for subsequent recognition processing, its overall feature credibility needs further evaluation and calibration, constructing a credibility anchor point sequence. This anchor point sequence describes the recognizability of different regions in the image structure after transmission compensation processing. In this step, several key feature blocks in the image are selected, including the fish's main body, fins, eyes, and abdomen. Multiple parameters, such as edge sharpness recovery value, brightness continuity index, and background compression smoothness index, are extracted from each region. Based on these parameters, an anchor point set is established in the image space. Each anchor point corresponds to a feature structure block in the image and is associated with its structural recovery integrity information. Through the spatial distribution and parameter differences of these anchor points, a multi-dimensional credibility description sequence can be constructed, providing a partitioning basis for determining the overall credibility of the image. This sequence can be used to judge the quality of the overall recognition output and also provides a basis for response delay control of subsequent recognition results.

[0036] The image frame processed as described above is used as the input frame for distortion removal recognition and is then fed into the subsequent recognition process. This image frame possesses the following processing features: First, the fish's outline structure has undergone rhythmic spatial unfolding, with a complete posture and clear contours, suitable for feature extraction of continuous actions or behaviors; second, the background transmission texture structure has undergone compression and backtracking processing, converging interference content outside the original area in spatial distribution, effectively reducing background redundancy information in the fish's recognition structure; third, the image frame contains an embedded confidence anchor point sequence, providing structured support for the confidence assessment and response control of the recognition results. Structurally, this image frame provides a stable structural benchmark after interference correction for subsequent recognition processes, ensuring the continuity and interpretability of fish feature extraction, and laying a solid foundation for image input quality for operations such as state determination, posture recognition, or behavior analysis based on fish images.

[0037] The dynamic credibility control module, based on the credibility anchor point sequence, initiates a breathing vortex gate control mechanism in the candidate region of the aggregation center, performs periodic micro-opening and closing operations of the recognition entrance, delays the recognition judgment corresponding to low credibility in the credibility anchor point sequence to the shadow relief channel, and restores the instant output of the recognition result when the credibility recovers, thereby controlling the continuous transmission process of recognition error. To ensure the stability of the identification output and the reliability of the decision response, a dynamic adjustment mechanism can be introduced within the candidate region of the aggregation center to coordinate the release rhythm and output state of the identification results. This mechanism constructs a structurally adjustable release path with periodic adjustment capabilities by referencing multi-point information in the confidence anchor sequence generated in the previous stage, combined with spatial perturbation distribution characteristics and temporal evolution trends. This avoids continuous interference from low-confidence regions on the identification process, ensuring the phased continuity of the fish state determination results. The specific implementation steps are as follows: Based on the generated confidence anchor point sequence in the dedistortion recognition input frame, the spatial regions corresponding to each anchor point within the image structure are hierarchically classified. Each anchor point in this sequence contains multiple feature parameters, such as edge sharpness restoration degree, background compression integrity, and texture consistency continuity. By classifying and integrating the distribution characteristics of these parameters in the image region, the anchor points can be divided into three categories according to confidence level: high confidence region, medium confidence region, and low confidence region. Subsequently, using the candidate region of the cluster center as the structural core, the anchor point levels covered within this region are spatially projected and mapped to form an anchor point confidence gradient band centered on the low confidence region and transitioning towards the high confidence region. This gradient band can reflect the spatial distribution trend of recognition stability in the cluster region, providing a dynamic adjustment basis for the subsequent release mechanism.

[0038] Around the anchor point confidence gradient band, a perceptual control path with a dynamic opening and closing structure is constructed at the recognition entry point of the candidate region of the aggregation center. This path is based on a periodically changing spatial transparency window, and its opening and closing rhythm is set according to the rate of change of the anchor point confidence gradient, so that the entire structure forms a breathing state evolution with inward and outward characteristics in the time dimension. This state evolution is driven by the continuous distribution of the anchor point sequence, which is manifested in the image structure as the recognition channel forming periodic micro-fluctuations in space. By adjusting the opening and closing frequency and opening amplitude of the recognition entry point, the recognition channel enters a closed phase before the recognition result in the low confidence region is about to enter the output stage. This low confidence judgment information is temporarily left in the structural channel and is not directly pushed to the judgment output chain, but waits for the confidence index of the anchor point sequence to recover before entering the next stage of processing.

[0039] To ensure a structured output path for the retention and release of recognition results, a delayed release structure channel can be constructed within the recognition path during the activation of the closed state of the recognition entry point. This channel carries the temporarily delayed recognition results and performs phased release processing on them. This release channel relies on the established perturbation path structure direction within the candidate region of the aggregation center, receiving low-confidence recognition judgment information sequentially in time series, maintaining its original structural state from premature unfolding. Simultaneously, this channel has an information release outlet, using the recovery rhythm of the confidence index in the anchor point sequence as the release condition, allowing the previously delayed information to be released in an orderly manner after the structural signal returns to a stable threshold. During this process, the timeline of the recognition results is appropriately stretched, preventing the system from directly outputting short-term consecutive misjudgments under perceptual instability, forming a structured processing rhythm with buffering characteristics.

[0040] As the confidence index of the anchor point sequence gradually recovers in subsequent image frames, the previously closed recognition entry point re-opens, allowing the information within the channel to be released frame by frame. This information then participates in the judgment process along with the new round of recognition results within the current output cycle. At this point, the recognition channel exhibits a dual structural state of synchronous release and periodic absorption operating in parallel, enabling the entire judgment and output chain to possess both input content quality perception and result output delay control capabilities. Through the aforementioned continuous cycle of closure-release-reopening-release, the recognition channel exhibits a contraction state when disturbance intensity fluctuations are significant, avoiding the accumulation of erroneous information in the output; and an expansion state when disturbances subside, allowing delayed information to enter the judgment stage in chronological order, achieving a natural connection of information rhythm. This structure forms a dynamically adaptive recognition output mechanism, establishing a flexible connection between information quality fluctuations and structural disturbance responses, improving the controllability and stability of the overall recognition process.

[0041] This invention, by constructing a dynamic recognition control structure, enables the perception process of fish images in underwater disturbed environments to possess rhythmic regulation capabilities. By introducing a periodically opening and closing breathing-like adjustment structure at the entrance of the recognition channel, the recognition results in low-confidence areas are processed in an orderly manner, avoiding the continuous output of erroneous judgments under abnormal perception conditions. This reduces the interference of misjudgment information on subsequent aquaculture decisions and improves the image perception process's responsiveness to local disturbance changes.

[0042] This invention achieves a stable transition of recognition results during image quality fluctuations by establishing a gradual release channel driven by image confidence gradients. After the perturbation subsides, the previously delayed output is released in a temporally continuous manner, maintaining the overall rhythm balance of the recognition process. This approach improves the integrity and continuity of diagnostic information, allowing the output to seamlessly connect during the period of perceived quality recovery, thus enhancing the reliability of the system in performing health status recognition tasks in complex underwater environments.

[0043] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An underwater intelligent identification and diagnostic system for leopard-gill spiny perch, characterized in that, It includes an optical feature analysis module, a particle aggregation detection module, a feature separation and comparison module, a transmission compensation and correction module, and a dynamic reliability control module. The optical feature analysis module extracts the temporal evolution features of brightness gradient, edge sharpness and texture details in the perceived image, constructs an optical transmission stable reference band, and extracts the abrupt change amplitude at its end to form a mutation marker. The particle aggregation detection module, based on mutation markers, performs temporal aggregation analysis on the particle density texture within the sensing area, extracts the circumferential stretching trajectory, generates a vortex aggregation pointing band, and identifies candidate aggregation center regions at its tail. The feature separation and comparison module constructs a dual-channel sampling comparison surface in the candidate region of the aggregation center, maps the fish body contour features and background transmission texture, generates a feature separation comparison frame, and extracts the transmission distortion factor. The transmission compensation and correction module performs inverse compensation on the separation feature comparison frame based on the transmission distortion factor, generates the distortion-free recognition input frame, and extracts the confidence anchor point sequence. The dynamic credibility control module, based on the credibility anchor point sequence, activates the breathing vortex gate control mechanism to execute the periodic opening and closing of the identification entrance, delays the output of low credibility judgments to the shadow relief channel, and restores the instant output when the credibility recovers, thereby controlling the continuous transmission of identification errors.

2. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 1, characterized in that, The steps for forming a mutation marker are as follows: The brightness gradient, edge sharpness, and texture details in continuously acquired underwater images are extracted frame by frame to generate a cluster of temporal evolution feature curves. Select time segments with smaller fluctuation ranges and smoother curve changes from the cluster of time evolution characteristic curves, and combine them to form an optical transmission stable reference band. Multi-feature difference extraction is performed on the end image frame of the optical transmission stable reference band to form a multi-feature difference distribution map; In the multi-feature difference distribution map, spatial focus areas are extracted, and pixel blocks are sorted according to brightness jump rate, edge break density and texture rhythm interference degree. Feature patches with large change amplitude and concentrated spatial range are extracted and encoded to generate mutation markers.

3. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 2, characterized in that, In the process of generating mutation markers, the spatial location, temporal location, and multidimensional variation amplitude of the feature patches are combined and encoded as the initial identification basis for mutation states.

4. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 2, characterized in that, The process of identifying candidate regions for cluster centers is as follows: Starting with mutation markers, a sequence of grain density texture images from multiple time frames in the perceptual region is obtained, texture density field information is extracted, and time-ordered texture clustering is performed. Based on the texture convergence trend, the spatial stretching trajectory is extracted from the mutation marker line to construct the vortex convergence direction zone of the circumferential stretching trend; In the tail region of the vortex aggregation pointing zone, the temporal evolution characteristics of the texture density field are traced to identify candidate regions of aggregation centers with density clustering states. Based on the spatial focusing characteristics of the candidate regions of the aggregation center, the mutation markers, vortex aggregation pointing bands and candidate regions of the aggregation center are summarized to form the evolution path of the perturbation structure.

5. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 4, characterized in that, The stretching trajectory of the vortex aggregation pointing band exhibits a circumferential extension in spatial structure and forms a continuous arc structure along the direction of the maximum gradient change in the particle density field. The candidate region of the aggregation center is located at the tail of the vortex aggregation pointing band. In the last frame of the image sequence, it is characterized by a continuously increasing texture density, a gradually decreasing slope, blurred boundary regions, and a flattening brightness gradient. The region is extracted through joint analysis of density evolution rhythm, boundary continuity, and spatial center stability.

6. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 4, characterized in that, The process for extracting the transmission distortion factor is as follows: A dual-channel sampling control surface is constructed within the candidate region of the aggregation center to map the fish body contour features to the main body layer, and contour overlay extraction is performed in multiple consecutive image frames. A background transmission layer is established in the image channel adjacent to the main layer, and background transmission texture information in the same area is collected to form an upper and lower layered structure. The main body layer and the background layer are aligned and analyzed to form a separation feature comparison frame, and information on structural offset, morphological curvature and texture misalignment is extracted. Based on the difference trajectory structure in the separation feature comparison frame, the cumulative offset information of the image structure is extracted to generate a transmission distortion factor to describe the amplitude and direction of the disturbance.

7. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 6, characterized in that, The transmission distortion factor includes the displacement gradient of the brightness channel, the periodic shift amplitude of the texture channel, and the deformation trend of edge sharpness. It is used to simultaneously express the spatial offset state of the fish body contour features and the perturbation direction of the background transmission texture. During the generation process, the spatial reference frame of the image frame remains unchanged to ensure accurate compensation and rhythm restoration of the subsequent image structure.

8. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 6, characterized in that, The steps for extracting credibility anchor sequence are as follows: Based on the transmission distortion factor, the fish body contour structure region in the separation feature comparison frame is located and processed, and the contour rhythm is continuously unfolded according to the distortion direction and deformation amplitude. While unfolding the fish body outline, the background transmission texture area is spatially compressed and rolled back according to the transmission distortion factor to restore the background structure to the state before the disturbance. After completing the fish outline unfolding and background transmission regression, multiple feature block parameters are extracted based on the degree of structure restoration, a set of credibility anchor points is constructed, and a credibility anchor point sequence is generated. Image frames with structure unfolding, texture regression, and credibility anchor point sequences are used as input frames for distortion removal recognition and then fed into the subsequent recognition process.

9. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 8, characterized in that, Each anchor point in the credibility anchor point sequence corresponds to a feature structure block in the fish's main body, fins, eyes, or abdomen in the image. Multidimensional parameters are generated by combining edge sharpness recovery value, brightness continuity index, and background compression smoothness index to represent the recognizability of the structure block.

10. The underwater intelligent identification and diagnostic system for leopard-gill spiny perch according to claim 8, characterized in that, Based on the confidence anchor point sequence, a breathing vortex gate control mechanism is activated to periodically open and close the entry point for identification. Low confidence judgments are delayed and output to the shadow relief channel. When the confidence level recovers, the output is restored. The steps are as follows: Based on the confidence anchor point sequence, the candidate region of the cluster center is classified into levels to generate anchor point confidence gradient bands; Based on the anchor point confidence gradient band, a periodic micro-opening and closing sensing and control path is constructed at the identification entry point to form a breathing-like state evolution structure. When the entry point is closed, a delayed release structure channel is constructed to receive low-confidence judgment information and maintain the original structure state. As the credibility index recovers, the control recognition entry is opened, and the information in the slow release channel is released frame by frame, so that the recognition results are output naturally in chronological order.

Images