Intelligent fish calling method and system based on deep learning

Abstract

The application discloses an intelligent fish protection and fish returning method and system based on deep learning, which obtains fish body images and total weight information, adopts an object recognition model to analyze fish body contour feature vectors in real time, and obtains detection quantity and species results by fusing total weight data; then, a fusion model is used to process multi-modal fusion of contour size and total weight data to determine single fish weight distribution, and if the deviation exceeds a threshold value, the feature weight is iteratively adjusted to obtain an accurate single fish weight value; subsequently, data transmission generates a visual interface, a fish returning instruction sequence is triggered according to user selection, a local processing unit executes an opening operation and records the number, weight and timestamp of the fish returned, and an encrypted log file is formed; finally, log analysis judges environmental compliance parameters and updates behavior records. The application realizes accurate and automatic management of fish catching, improves the efficiency of fishing activities and environmental sustainability, and ensures the integrity of compliance behavior feedback.

Description

Technical Field

[0001] This invention relates to the field of fishing technology, and in particular discloses a method and system for weighing and retrieving fish using a deep learning-based intelligent fish catcher. Background Technology

[0002] In modern fishing activities, the application of intelligent technology is gradually becoming an important means to enhance the experience and protect fishing resources. Whether it's competitive or recreational fishing, accurately recording catch information and implementing standardized management are crucial to ensuring fairness and sustainable development. Research in this field is not only about technological innovation but also closely related to ecological protection and user needs, demonstrating extremely high application value and practical significance. However, current tools and methods on the market have revealed many problems in practical use.

[0003] Many solutions are ill-suited to complex outdoor aquatic environments, with equipment susceptible to water disturbances or weather conditions, leading to unstable data acquisition. More importantly, existing technologies often struggle to make comprehensive judgments when processing fish catch information, particularly in combining fish identification and weight estimation. The lack of effective technical means results in insufficient information accuracy, impacting user trust in fish catch management.

[0004] Focusing on the technical challenges, the core issue lies in how to efficiently capture and accurately analyze fish information in a dynamic and complex underwater environment. Especially in comprehensively determining fish size and weight, the system often struggles to process information quickly and accurately due to variations in underwater light, the unpredictable posture of the fish, and limitations in the equipment's computing power.

[0005] This technological bottleneck directly leads to deviations in the collection and analysis of fish catch data in actual fishing scenarios, thus affecting subsequent decision-making. For example, in a fishing activity, after anglers put multiple fish into their baskets, the system needs to capture images of the fish using underwater equipment and analyze them in conjunction with weight data. However, due to water currents and fish obscuring each other, the image information is incomplete. Simultaneously, the equipment's processing of weight data cannot effectively correspond to the image information, ultimately resulting in significant errors in the estimation of the number and weight of the fish. This situation is particularly prominent in multi-person competitive fishing, where data deviations can directly affect the fairness of the competition results.

[0006] Therefore, how to accurately fuse fish body information and weight data in complex underwater environments and ensure the system's rapid response under limited computing resources has become a key problem that this study urgently needs to solve. Summary of the Invention

[0007] This invention provides a deep learning-based intelligent fishkeeping method and system for weighing and retrieving fish, aiming to solve at least one of the defects in the prior art.

[0008] One aspect of the present invention relates to a deep learning-based intelligent fishkeeping method for weighing and retrieving fish, comprising the following steps:

[0009] S100: Acquire fish images acquired by the image acquisition device and total weight data acquired by the weight sensor; use an object recognition model to perform real-time analysis of the fish images; extract feature vectors from the fish outline; and fuse the total weight data to obtain the number of fish detected and the species identification results.

[0010] S200. Based on the number of fish detected and the species identification results, a fusion model is used to process the multimodal fusion of fish outline size and total weight data to determine the weight distribution of a single fish. If the deviation after fusion exceeds the threshold, the feature weights are iteratively adjusted to obtain the accurate weight value of a single fish.

[0011] S300: After obtaining the accurate weight value of a single fish, the accurate weight value of a single fish and the number of fish detected are transmitted through the terminal device to generate a visual fish catch information interface. The user's input is judged. If the user chooses to put the fish back, the control signal is triggered to obtain the fish return instruction sequence.

[0012] S400: For the fish return instruction sequence, the local processing unit performs the start operation, while recording the number and weight of the fish returned, and integrating the timestamp information to determine the fish return event log entries;

[0013] S500: Extract key fields from the fish return event log entries, process the fish return event log entries using an encryption algorithm, and generate a complete log file;

[0014] S600 parses persistent data through complete log files to determine environmental compliance parameters. If the environmental compliance parameters meet the specifications, it updates the behavior records and obtains the behavior feedback sequence.

[0015] Further, step S100 includes:

[0016] S110. Acquire the fish image captured by the image acquisition device and the total weight data recorded by the weight sensor. Perform noise reduction and edge enhancement on the fish image using a pre-established image processing tool to obtain a clear fish outline image and determine the position and boundary range of the fish in the image.

[0017] S120. For the fish outline image, use an image segmentation tool to extract the feature vector of the fish outline, and compare it with a preset feature matching database. If the matching degree of the feature vector of the fish outline is higher than the preset threshold, it is judged as a known fish species, and a preliminary species classification result is obtained.

[0018] S130. Based on the preliminary species classification results and total weight data, the number of fish is estimated using a data fusion tool. If the estimated number is inconsistent with the number identified in the fish outline image, the feature matching parameters are adjusted and recalculated to determine the final fish count.

[0019] S140. By linking the final fish count with the preliminary species classification results, a complete fish detection report is generated using a data recording tool to obtain the fish count and species identification results for subsequent processing.

[0020] Further, step S200 includes:

[0021] S210. Obtain the corresponding fish body outline size information from the fish body detection quantity and species identification results, and use image processing tools to standardize the outline size information to obtain unified size feature data.

[0022] S220. Based on the size characteristic data and the total weight data, a preliminary weight distribution estimate is made using data fusion tools to determine the average weight range for each species.

[0023] S230. If the average weight range of each species is determined to be outside the preset deviation threshold, the feature weights are iteratively corrected using a parameter adjustment tool to determine the adjusted weight distribution information.

[0024] S240. Based on the adjusted weight distribution information, generate the final single-fish weight report using a data recording tool to obtain the accurate single-fish weight value.

[0025] Further, step S300 includes:

[0026] S310. Obtain the precise single fish weight value and fish detection quantity from the terminal device through a data transmission tool, and use a data verification tool to check the integrity of the transmitted content to determine the accuracy of the transmitted data.

[0027] S320. Based on the accuracy of the transmitted data, an interface generation tool is used to construct a visual fish catch information interface, which is categorized and displayed according to the accurate single fish weight value and the number of fish detected, resulting in user-interactive interface content.

[0028] S330. If the user inputs selection information through the visual fish catch information interface, the input parsing tool is used to identify the content of the selection information, and combined with the preset threshold, it is determined whether it is to release the fish back into the body, and the classification result of the user's intention is determined.

[0029] S340. Based on the classification results of user intent, a signal triggering tool is used to generate corresponding control signals, and a return instruction sequence is constructed using an instruction generation tool to obtain the instruction content for subsequent execution.

[0030] Further, step S400 includes:

[0031] S410. After obtaining the fish return command sequence, use the command parsing tool to decompose the command content of the fish return command sequence, classify the decomposition results, and determine the operation type and execution order contained in the fish return command sequence.

[0032] S420. Based on the operation type and execution order, the corresponding start operation is executed using a local processing tool. During the execution process, the number and weight of the fish released are obtained in real time through a data acquisition tool.

[0033] S430. For the data on the number and weight of the returned fish, a time stamping tool is used to generate corresponding timestamp information, which is then linked and bound using a data integration tool to obtain the merged fish return event log entries.

[0034] Further, step S500 includes:

[0035] S510. Obtain the original record from the fish return event log entry, parse the content of the original record, match at least one key field through the preset field rules, and use a word segmentation tool to perform text segmentation on the original record to obtain a structured log field set.

[0036] S520. Based on the structured log field set, classify and process the content of each field. If the field content meets the preset sensitive information standard, use encryption tools to protect the field content and generate an encrypted log field combination.

[0037] S530. By combining the encrypted log fields, obtain the corresponding log entry identifier, combine the encrypted log field combination with the corresponding log entry identifier using a preset format template, and use a generation tool to construct new log file data.

[0038] S540. Based on the new log file data, allocate a storage path for each file item, and write the new log file data to the specified location using the storage management tool to obtain the complete log file.

[0039] Further, step S600 includes:

[0040] S610. Extract persistent data from the complete log file, format the persistent data, match at least one environment-related parameter according to preset field rules, and use a parsing tool to compare the persistent data item by item to obtain the preliminary classification results of the environment parameters.

[0041] S620. Based on the preliminary classification results of environmental parameters, compare each parameter value with the preset compliance threshold. If the parameter value exceeds the preset compliance threshold, mark it as an anomaly. Use a recording tool to generate an anomaly parameter list and determine the specific distribution of the anomaly parameters.

[0042] S630. Obtain the corresponding behavior record data through the abnormal parameter list, match the correlation between the abnormal parameters and the behavior record data, and if the matching result shows a correlation, use the update tool to adjust the behavior record to obtain the behavior feedback sequence.

[0043] Another aspect of the present invention relates to a deep learning-based intelligent fish-weighing and fish-retrieving system for performing the aforementioned deep learning-based intelligent fish-weighing and fish-retrieving method, comprising:

[0044] The fish body detection data acquisition module is used to acquire fish body images collected by the image acquisition device and total weight data collected by the weight sensor. It uses an object recognition model to perform real-time analysis of the fish body images, extracts feature vectors from the fish body contours, and fuses the total weight data to obtain the number of fish detected and the species identification results.

[0045] The module for obtaining accurate single fish weight value is used to process the multimodal fusion of fish outline size and total weight data based on the number of fish detected and the species identification results, and to determine the weight distribution of single fish. If the deviation after fusion exceeds the threshold, the feature weights are iteratively adjusted to obtain the accurate single fish weight value.

[0046] The fish return instruction sequence acquisition module is used to obtain the accurate single fish weight value, transmit the accurate single fish weight value and the number of fish detected through the terminal device, generate a visual fish catch information interface, determine the user's input, and if the user chooses to put the fish back, trigger a control signal to obtain the fish return instruction sequence.

[0047] The fish return event log entry determination module is used to perform an enabling operation using a local processing unit for a fish return command sequence, while recording the number and weight of the returned fish, and integrating timestamp information to determine the fish return event log entry.

[0048] The complete log file generation module is used to extract key fields from the return fish event log entries, process the return fish event log entries with encryption algorithms, and generate a complete log file.

[0049] The behavior feedback sequence acquisition module is used to parse persistent data from the complete log file, determine the environment compliance parameters, and update the behavior records to obtain the behavior feedback sequence if the environment compliance parameters meet the specifications.

[0050] The beneficial effects achieved by this invention are as follows:

[0051] This invention provides a deep learning-based intelligent fishkeeping method and system for weighing and retrieving fish. It addresses the challenges of accurately identifying the number, species, and weight of fish caught during fishing activities, the lack of automated recording of release operations, and the difficulty in real-time monitoring of environmental compliance. By acquiring fish images and total weight information, an object recognition model is used to analyze the fish's contour feature vectors in real time, and the total weight data is fused to obtain the detected number and species results. Furthermore, a fusion model is used to process the multimodal fusion of contour size and total weight data to determine the weight distribution of individual fish. If the deviation exceeds a threshold, the feature weights are iteratively adjusted to obtain an accurate individual fish weight value. Subsequently, data is transmitted to generate a visual interface. Based on the user's selection of a fish-retrieval command sequence, the local processing unit executes the activation operation and records the number, weight, and timestamp of fish released, forming an encrypted log file. Finally, the log is parsed to determine environmental compliance parameters and update the behavior record. This invention achieves precise and automated management of fish catches, improves the efficiency of fishing activities and environmental sustainability, and ensures the integrity of compliance behavior feedback. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating an embodiment of an intelligent fish-keeping and fish-retrieving method based on deep learning according to the present invention.

[0053] Figure 2 This is a functional block diagram of an embodiment of an intelligent fish-keeping and fish-retrieving system based on deep learning according to the present invention.

[0054] Explanation of icon numbers:

[0055] 10. Fish body detection data acquisition module; 20. Precise single fish weight value acquisition module; 30. Fish return command sequence acquisition module; 40. Fish return event log entry determination module; 50. Complete log file generation module; 60. Behavior feedback sequence acquisition module. Detailed Implementation

[0056] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0057] like Figure 1 As shown, the first embodiment of the present invention proposes a deep learning-based intelligent fishkeeping method for weighing and retrieving fish, comprising the following steps:

[0058] Step S100: Obtain the fish body image collected by the image acquisition device and the total weight data collected by the weight sensor. Use an object recognition model to perform real-time analysis on the fish body image, extract feature vectors for the fish body contour, and fuse the total weight data to obtain the number of fish detected and the species identification results.

[0059] The system simultaneously acquires real-time dynamic images of the fish bodies captured by the built-in image acquisition device (such as a waterproof camera) in the fish basket, as well as the total weight data of the fish bodies inside the fish basket collected by a weight sensor (such as a high-precision strain gauge sensor). A lightweight object recognition model (such as an improved YOLOv8-nano, adapted for edge computing scenarios, with a detection speed ≥25fps) is used to analyze the fish body images. Fish body regions are located through object detection, and geometric features of the fish body contour (such as body length, body width, contour perimeter, and key skeletal point coordinates) are extracted by semantic segmentation and encoded into high-dimensional feature vectors (dimension ≥512). A feature fusion module (such as attention-based weighted fusion) associates the contour feature vectors with the total weight data (e.g., using the total weight range to constrain the possible range of fish numbers), outputting the number of fish detected (e.g., "3 fish") and species identification results (e.g., "2 crucian carp, 1 bass"). The accuracy rate for quantity detection is required to be ≥95%, and the accuracy rate for species identification is required to be ≥92% (for common freshwater fish), providing multi-dimensional basic data of "quantity-species-total weight-morphology" for subsequent single-fish weight calculation.

[0060] Step S200: Based on the number of fish detected and the species identification results, a fusion model is used to process the multimodal fusion of fish outline size and total weight data to determine the weight distribution of a single fish. If the deviation after fusion exceeds the threshold, the feature weights are iteratively adjusted to obtain the accurate weight value of a single fish.

[0061] Based on the number of fish detected (e.g., "3 fish") and species identification results (e.g., "2 crucian carp, 1 bass") output in step S100, a fusion model (e.g., a multimodal deep neural network with visual feature branches and weight feature branches) is invoked to process the fish outline size data (e.g., the body length, body width, and outline area of ​​each fish, converted into body shape parameters through a regression model) and total weight data (e.g., "total weight 1.2kg"): First, a pre-trained "body length-weight" baseline model is loaded according to species characteristics (e.g., the weight of crucian carp ≈ a × body length² + b × body width, the weight of bass ≈ c × body length³, parameters a, b, and c vary depending on the species), combined with the wheel The initial weight distribution of a single fish is determined by its outline size. Then, a multimodal fusion layer (such as a cross-attention mechanism) is used to calculate the deviation between the sum of the weights of the single fish and the total weight data. If the deviation exceeds a preset threshold (such as ±5% or ±10g, taking the maximum of the two), the weight ratio of the outline features and the total weight data is iteratively adjusted (such as increasing the weight of the total weight data in the allocation of large individual fish, or strengthening the influence of the outline size on small fish), and the weight distribution of the single fish is recalculated. This adjustment is repeated until the deviation is ≤ the threshold, and finally, the accurate weight value of each fish is output (accuracy ≤ ±2g). The accuracy of the single fish weight calculation is required to be ≥96%, providing a quantitative basis for subsequent fish catch display and repatriation decisions.

[0062] Step S300: After obtaining the accurate weight value of a single fish, the accurate weight value of a single fish and the number of fish detected are transmitted through the terminal device to generate a visual fish catch information interface. The user's input is judged. If the user chooses to put the fish back, a control signal is triggered to obtain the fish return instruction sequence.

[0063] After obtaining the precise single fish weight value output in step S200 (e.g., "Crucian Carp 1: 250g, Crucian Carp 2: 280g, Bass: 420g"), the single fish weight, total number of fish detected (e.g., "3 fish"), and species details are transmitted to the user terminal device (e.g., mobile APP or smart bracelet) via a wireless communication module (e.g., Bluetooth 5.0 or Wi-Fi). The terminal generates a visual fish catch information interface based on the received data, including a fish catch list (species, weight, and image thumbnail for each fish), total weight statistics (e.g., "950g"), and operation options ("Keep", "Release" button). The system includes a button that supports single or batch selection; it uses interface interaction logic to determine user input (e.g., clicking "Release Bass"). If a fish return selection is detected, the terminal generates a control signal containing the target fish identifier (e.g., ID number) and operation type ("Single Return"). This signal is encrypted and transmitted to the local control unit of the fish basket, where it is parsed into a fish return instruction sequence (e.g., "Identify ID3 fish location → Drive corresponding area bottom plate to unlock → Delay 2 seconds to close → Confirm return complete"). The instruction transmission delay is required to be ≤500ms, and the interface operation response accuracy is ≥99%, providing accurate instruction basis for subsequent fish return execution.

[0064] Step S400: For the fish return instruction sequence, the local processing unit performs the start operation, while recording the number and weight of the returned fish, and combining the timestamp information to determine the fish return event log entry.

[0065] After receiving the fish return instruction sequence (e.g., "Return ID3 sea bass, weight 420g") output in step S300, the local processing unit (e.g., MCU microcontroller) built into the fish basket starts the execution logic, driving the automatic opening and closing mechanism (e.g., waterproof servo motor control base plate) of the corresponding area to complete the opening operation. At the same time, the image acquisition device confirms for the second time that the fish has left the fish basket (to avoid accidental operation); key data of fish return are recorded simultaneously: the number of fish returned (e.g., "1 fish"), single fish / total weight (e.g., "420g"), and the built-in clock module is called to obtain the accurate timestamp (e.g., "2025-11-03 15:30:22"); the above information is integrated with the fish species tag (e.g., "sea bass") and the device identifier (e.g., fish basket unique ID) through the data integration module to generate a structured fish return event log entry (fields include: event ID, timestamp, number of fish returned, total weight, species details, device ID), requiring 100% data recording accuracy and timestamp error ≤1 second, providing a raw and complete event data foundation for subsequent log encryption and compliance analysis.

[0066] Step S500: Extract key fields from the return event log entries, process the return event log entries using an encryption algorithm, and generate a complete log file.

[0067] From the fish return event log entries generated in step S400, extract key fields with core value (including unique event identifier, timestamp, species / number / total weight of the returned fish, device identifier, etc., and filter redundant format information); use a lightweight encryption algorithm (such as AES-128 symmetric encryption, adapted to edge device computing power) to encrypt the key fields and complete log entries to ensure that the data is not tampered with or leaked during local storage or transmission; through the log integration module, aggregate the encrypted single log entries in chronological order, add a file header (including encryption algorithm identifier, log start / end time) and a checksum (such as CRC32, used to verify file integrity), and finally generate a structured complete log file (format can be encrypted binary or encrypted JSON). The encryption process is required to take ≤100ms, and the log file integrity verification pass rate is 100%, providing a secure and reliable data source for subsequent data parsing and compliance checks.

[0068] Step S600: Parse the persistent data through the complete log file to determine the environment compliance parameters. If the environment compliance parameters meet the specifications, update the behavior record and obtain the behavior feedback sequence.

[0069] For the complete log file generated in step S500, persistent data (including the time, species, quantity, weight, and device association information of historical fish return events) is parsed through the authorized decryption module (such as user terminal key verification); the environmental compliance parameter database (including local fishing activity regulations, such as closed fishing periods, protected species lists, minimum catch size / weight limits, etc.) is called, and the data is compared and parsed with compliance parameters through the rule engine (such as determining whether the released species is a protected animal, whether the fish return time is within the closed fishing period, and whether the retained fish body meets the size standard); if the comparison results show that all behaviors comply with the regulations (such as not involving protected species, not operating during the closed fishing period), the user behavior record database is updated (compliance mark and points reward are added), and the behavior feedback sequence is output through the natural language generation model (such as "This fish return complies with ecological protection regulations, cumulative compliance times +1, 5 points of 'ecological angler' are obtained"); if there are non-compliant items, the specific violation points are fed back (such as "The retained crucian carp is less than 15cm in length, it is recommended to release it to protect the juvenile fish"). The system requires a compliance judgment accuracy rate of ≥99% (based on the timeliness of the regulatory database) and a feedback sequence generation response time of ≤1 second, providing users with real-time and accurate behavioral guidance to help protect fishing resources sustainably.

[0070] Furthermore, the deep learning-based intelligent fishkeeping and fish return method proposed in this embodiment includes step S100:

[0071] Step S110: Obtain the fish image captured by the image acquisition device and the total weight data recorded by the weight sensor. Use a pre-established image processing tool to perform noise reduction and edge enhancement on the fish image to obtain a clear fish outline and determine the position and boundary range of the fish in the image.

[0072] The following formula is used to denoise captured fish images:

[0073] (1)

[0074] In formula (1), This represents the pixel values ​​of the image after denoising. Represents the pixel values ​​of the original fish image. Indicates the original fish body image in The pixel value at the pixel coordinates. This represents the Gaussian filter kernel function. This indicates the radius of the filter kernel. and This indicates the pixel coordinates in the image.

[0075] The following formula is used to calculate the edge intensity of an image for edge enhancement processing:

[0076] (2)

[0077] In formula (2), Represents the image edge intensity value. This represents an image that has undergone noise reduction processing. Indicates the image in gradient of direction, Indicates the image in Gradient of direction.

[0078] The following formula is used to determine the position and boundary of the fish in the image:

[0079] (3)

[0080] In formula (3), The set of pixels representing the boundary of the fish's body. This represents the pixel values ​​of the image after edge enhancement. and Indicates the threshold range for boundary detection. , , , These represent the minimum and maximum coordinate boundaries of the fish in the image, respectively.

[0081] In the process of acquiring images of the fish captured by the image acquisition device and total weight data recorded by the weight sensor, the image acquisition device typically uses a high-resolution camera installed inside the fish enclosure to capture dynamic images of the fish in real time. The weight sensor, placed inside the enclosure, records the total weight of the fish as they pass through. Assuming the fish farm uses a 1920x1080 resolution CCD camera that captures 30 frames per second, and simultaneously connects to an electronic scale sensor with an accuracy of 0.1 kg, when the fish are guided through the detection area, the system simultaneously acquires images and weight data. For example, a total weight of 5.2 kg represents the cumulative weight of a batch of fish. This data serves as the input for subsequent processing, ensuring the real-time nature and accuracy of fish detection.

[0082] The process of denoising and edge enhancement of fish images using pre-established image processing tools needs to be explained in detail. The image processing tools are implemented based on the OpenCV library. First, denoising employs a Gaussian filtering algorithm. The principle is to use a Gaussian function to perform a weighted average of image pixels, reducing random noise. For example, applying a 5x5 Gaussian kernel to a fish image smooths out water ripple interference. Next, edge enhancement uses the Canny operator. The principle is to detect edges by calculating image gradients. First, the pixel gradient intensity and direction are calculated, then non-maximum suppression and double thresholding are applied to obtain a clear fish outline. Specifically, when processing an image containing multiple fish, the system determines the position and boundary range of the fish in the image. For example, the bounding box coordinates of each fish are found using a contour detection function. For instance, the boundary of fish 1 is from (100, 200) to (300, 400), thus isolating individual fish regions and providing precise boundaries for subsequent feature extraction.

[0083] Step S120: For the fish outline image, use an image segmentation tool to extract the feature vector of the fish outline and compare it with a preset feature matching database. If the matching degree of the feature vector of the fish outline is higher than the preset threshold, it is judged as a known fish species, and a preliminary species classification result is obtained.

[0084] The fish body contour after image segmentation is converted into a feature vector representation that can be used for matching using the following formula:

[0085] (4)

[0086] In formula (4), This represents the extracted fish body contour feature vector. This represents the total number of feature points on the contour. Indicates the first The weight coefficients of each feature point Indicates the first The eigenvalues ​​of each feature point.

[0087] The cosine similarity between two feature vectors is calculated as the matching degree using the following formula:

[0088] (5)

[0089] In formula (5), The score represents the matching degree of the feature vector. This represents the feature vector of the fish to be identified. This represents the feature vector of a known fish species in the database. Represents the dot product of two vectors. It represents the product of the magnitudes of two vectors.

[0090] The threshold-based binary classification decision is implemented using the following formula:

[0091] (6)

[0092] In formula (6), This indicates preliminary results of species classification. This represents the preset matching threshold. When the matching degree is greater than or equal to the threshold, the output is 1 to indicate that the fish species is known. When the matching degree is less than the threshold, the output is 0 to indicate that the fish species is unknown.

[0093] This paper focuses on the technical topic of extracting feature vectors from fish outline images using image segmentation tools. These tools are based on the U-Net neural network model, which achieves pixel-level segmentation through an encoder-decoder structure. After extracting the fish outline, Hu moments are used to calculate feature vectors. These invariants describe the geometric properties of the shape, such as area, perimeter, and eccentricity. The process of comparing the fish with a pre-set feature matching database involves calculating the matching degree using cosine similarity. For example, if the database stores feature vectors for carp [0.2, 0.5, 0.1], and the similarity of the current fish vector [0.19, 0.51, 0.09] is higher than the threshold of 0.9, it is identified as a known fish species, carp, resulting in a preliminary species classification. This method can effectively distinguish similar fish species, such as carp and grass carp, in business operations, avoiding inventory management errors caused by misjudgments.

[0094] Step S130: Based on the preliminary species classification results and total weight data, the number of fish is estimated using a data fusion tool. If the estimated number is inconsistent with the number identified in the fish outline image, the feature matching parameters are adjusted and recalculated to determine the final fish count.

[0095] The number of fish can be estimated using the following formula:

[0096] (7)

[0097] In formula (7), This indicates the number of fish estimated through data fusion. This represents the total weight data detected. This represents the total number of categories in species classification. Indicates the first The probability distribution of species in the sample Indicates the first The average weight of a single organism in a species.

[0098] The adjusted feature matching parameters are obtained using the following formula:

[0099] (8)

[0100] In formula (8), This represents the adjusted feature matching parameters. Indicates the current feature matching parameters. This represents the learning rate used to adjust the parameters. This indicates the number of fish estimated by data fusion.

[0101] The final fish count was obtained using the following formula:

[0102] (9)

[0103] In formula (9), This indicates the final count of the number of fish. This indicates the total number of candidate fish. Indicates the number of contour recognitions. arrive These represent the weighting coefficients for different constraints. This indicates the average body weight of the fish. Indicates total weight constraint. Indicates the total number of species categories. Indicates the first The number of species, Indicates the first Reference number of species.

[0104] Based on preliminary species classification results and total weight data, the number of fish is estimated using a data fusion tool. Kalman filtering is employed as the data fusion tool. The principle is to combine the quantity estimation from image recognition with weight data for state prediction and updating. For example, if the initial classification is 10 carp, each with an average weight of 0.5 kg and a total weight of 5.2 kg, the estimated number is 5.2 / 0.5 = 10.4, approximately 10 fish. If this differs from the 9 fish identified in the image, the feature matching parameters are adjusted, such as lowering the threshold to 0.85, and the calculation is repeated iteratively until a match is found, confirming the final fish count as 10 fish. This fusion improves the robustness of the counting, especially reducing errors when fish overlap.

[0105] Step S140: By associating the final fish count with the preliminary species classification results, a complete fish detection report is generated using a data recording tool to obtain the fish count and species identification results for subsequent processing.

[0106] The final fish count results are linked and integrated with the preliminary species classification results using the following formula to generate a complete test report:

[0107] (10)

[0108] In formula (10), This represents complete fish body examination report data. The quantitative value representing the quantity detection result. This indicates the accuracy score for species identification. Represents timestamp information. , , These represent the weighting coefficients for each data item.

[0109] The process involves correlating the final fish count with the preliminary species classification results and generating a complete fish detection report using a data logging tool based on an SQLite database. The resulting report includes fields such as fish species, quantity, and total weight. For example, the report might show "Carp: 10, Total Weight: 5.2kg". This information is used for subsequent processing such as inventory updates or sales forecasts, resulting in efficient fishing activity management and ensuring data traceability and decision support.

[0110] Furthermore, the deep learning-based intelligent fishkeeping and fish return method proposed in this embodiment includes step S200:

[0111] Step S210: Obtain the corresponding fish body outline size information from the fish body detection quantity and species identification results, and use image processing tools to standardize the outline size information to obtain unified size feature data.

[0112] The following formula maps the outline dimensions of fish of different sizes to a standardized numerical range between 0 and 1:

[0113] (11)

[0114] In formula (11), This represents the standardized fish body outline size characteristics. This indicates the size of the detected individual fish body outline. This represents the minimum size of the body outline of a fish of the same species. This represents the maximum size of the body outline of fish of the same species.

[0115] The following formula is used to fuse multi-dimensional fish body outline size information into a unified feature representation:

[0116] (12)

[0117] In formula (12), This represents uniform fish body size characteristic data. Indicates the characteristics of fish body length. Indicates the width characteristics of the fish body. Indicates the height characteristics of the fish body. Weighting coefficients representing length features. The weighting coefficients representing the width feature. Weighting coefficients representing highly distinctive features.

[0118] The process of obtaining corresponding fish outline size information from the fish detection count and species identification results first involves using a contour extraction algorithm to obtain the length and width data of each fish based on the previously identified number of fish (e.g., 15) and species (e.g., tilapia). For example, the outline size of a tilapia is 30 cm in length and 10 cm in width. This information is derived from pixel coordinates in the image. Next, an image processing tool is used to standardize the outline size information. This tool is implemented using the MATLAB image toolbox. The principle is to map different fish sizes to a uniform scale through normalization operations. For example, dividing all dimensions by the maximum length value yields the relative proportion; for instance, the original length of 30 cm is normalized to 0.75, thus obtaining uniform size feature data. This helps eliminate errors caused by differences in shooting distance and ensures consistency in subsequent analysis. Specifically, when processing a batch of mixed fish species images, the system first collects all outline sizes and then applies a min-max normalization method. For example, the width data is scaled from 5 cm to 15 cm, and the output is a value between 0 and 1, forming a feature vector such as [0.75, 0.33]. This data serves as the input basis.

[0119] Step S220: Based on the size characteristic data and the total weight data, perform a preliminary weight distribution estimation using a data fusion tool to determine the average weight range for each species.

[0120] A preliminary weight distribution estimate can be made using the following formula:

[0121] (13)

[0122] In formula (13), Indicates the first Estimated weight of each species, Indicates the first Volumetric characteristic parameters of a species Indicates the first Surface area characteristic parameters of a species Indicates the first Length characteristic parameters of a species , , These represent the contribution coefficients of volume, surface area, and length to weight, respectively.

[0123] The average weight range for each species is derived using the following formula:

[0124] (14)

[0125] In formula (14), Indicates the first The average weight range for each species Indicates the first Total weight data for each species This represents the total number of dimensional characteristic parameters. Indicates the first The fusion weighting factor for each size feature, Indicates the first The standard deviation of the weight distribution of each species.

[0126] The process of preliminary weight distribution estimation is performed by combining size characteristic data with total weight data using a data fusion tool. This tool employs a Bayesian network method, which integrates size and weight information through a probabilistic model. For example, it correlates standardized length with known fish species density to estimate the weight of a single fish. For instance, assuming a total weight of 7.5 kg, a size characteristic showing an average standardized length of 0.8 cm, and an empirical tilapia density of 1.2 kg / cm³, the preliminary estimated average weight is 0.5 kg per fish. Then, the average weight range for each species is determined, such as 0.4-0.6 kg for tilapia. This step, through data fusion, improves the reliability of the estimation, especially providing a more accurate distribution view when fish sizes are uneven.

[0127] Step S230: If the average weight range of each species exceeds the preset deviation threshold, the feature weights are iteratively corrected using a parameter adjustment tool to determine the adjusted weight distribution information.

[0128] The following formula is used to calculate the species weight deviation judgment index:

[0129] (15)

[0130] In formula (15), Indicates the first The degree of deviation for each species Indicates the first Average weight of each species Indicates the first Preset reference weight for each species This indicates the preset deviation threshold. A value greater than 1 indicates that the weight deviation of the species exceeds the allowable range.

[0131] The iterative correction process of feature weights is described by the following formula:

[0132] (16)

[0133] In formula (16), Indicates the first The feature in the first The weight values ​​after the next iteration Indicates the first The feature in the first The weight value at the next iteration This represents the learning rate parameter. The loss function represents the loss function on the th The partial derivatives of each feature weight are used to continuously adjust the feature weights using the gradient descent method to reduce prediction error.

[0134] The adjusted weight distribution information is represented by the following formula:

[0135] (17)

[0136] In formula (17), Indicates the adjusted weight The probability density distribution, Indicates the number of components in a mixed distribution. Indicates the first The weight coefficients of each distributed component, The mean is variance is The normal distribution is used to describe the final distribution characteristics of species weight through a weighted combination of multiple normal distributions.

[0137] If the determined average weight range exceeds the preset deviation threshold, a parameter adjustment tool iteratively corrects the feature weights. This tool, based on the gradient descent algorithm, works by gradually optimizing the weights by minimizing the error function. For example, if the preset threshold is 0.1 kg, and the estimated range is 0.3-0.7 kg, the weights of the size features are iteratively adjusted from 0.6 to 0.7, and the distribution is recalculated until the error converges, determining the adjusted weight distribution information, such as a precise range of 0.45-0.55 kg. This correction mechanism can handle deviations caused by environmental variables such as water temperature, ensuring the robustness of weight estimation.

[0138] Step S240: Based on the adjusted weight distribution information, generate the final single fish weight report using a data recording tool to obtain the accurate single fish weight value.

[0139] The adjusted precise weight of a single fish is calculated using the following formula through a weighted average method:

[0140] (18)

[0141] In formula (18), This indicates the precise weight of a single fish. Indicates the first Fish weight data at each weight measurement point Indicates the first The adjustment weighting coefficients corresponding to each measurement point This indicates the total number of weight measurement points.

[0142] The following formula is used to assess the dispersion of the weight distribution and generate a distribution report:

[0143] (19)

[0144] In formula (19), The standard deviation index represents the weight distribution report. Indicates the first The weight measurement of the fish. This represents the average weight of all the fish. This indicates the total number of fish.

[0145] The process involves generating a final single-fish weight report based on the adjusted weight distribution information using a data logging tool. This tool is based on a MySQL database and works by converting the distribution information into a report format through query and aggregation operations. For example, the adjusted distribution of tilapia is associated with a quantity of 15 fish, and the precise weight value of each fish, such as 0.5 kg, is calculated. The report fields include species, single-fish weight, and total. This provides accurate tracking in fishing activity inventory management and supports subsequent feeding optimization.

[0146] Furthermore, the deep learning-based intelligent fishkeeping and fish return method proposed in this embodiment includes step S300:

[0147] Step S310: Obtain the accurate single fish weight value and fish detection quantity from the terminal device through the data transmission tool, and use the data verification tool to check the integrity of the transmitted content to determine the accuracy of the transmitted data.

[0148] The following formula is used to count the total number of fish detected by the terminal device and to consider the reliability of the detection:

[0149] (20)

[0150] In formula (20), This indicates the total number of fish tested. Indicates the first The number of fish identified within a detection area. Indicates the first Confidence coefficient of each detection area This indicates the total number of areas detected.

[0151] The following formula is used by data verification tools to calculate the integrity of the transmitted content, ensuring the accuracy of data transmission:

[0152] (twenty one)

[0153] In formula (21), Indicates data transmission integrity metrics. Indicates the first The number of data packets that are incorrect or lost. Indicates the total amount of data transmitted. This indicates the number of packets that were detected as errors.

[0154] The process of obtaining precise single-fish weight and fish detection count from terminal devices via data transmission tools. These tools utilize wireless network protocols such as Wi-Fi or Bluetooth. The principle involves sending data from fishing equipment to a central server via an encrypted channel. For example, a terminal device installed on a fishing boat might collect a single-fish weight of 0.5 kg and a detection count of 20 fish. This data, derived from sensor measurements, is packaged into JSON format for transmission, ensuring real-time performance and security. In fishing scenarios, terminal devices such as smart scales, combined with cameras, first calculate the weight. Then, the data transmission tool automatically initiates a connection request. After the receiver verifies its identity, it downloads the data packet to obtain complete information. This step bridges the gap between data acquisition and subsequent processing.

[0155] The process of using data verification tools to check the integrity of transmitted content and determine the accuracy of transmitted data involves employing hash algorithms such as MD5 to calculate checksums. The principle is to detect tampering or loss by comparing the hash values ​​at the sending and receiving ends. For example, if a transmitted data packet contains the weight value 0.5kg and the quantity 20, the data verification tool generates a hash code such as "abc123". If the receiving end's calculation matches, accuracy is confirmed; otherwise, a retransmission mechanism is triggered. In actual business applications, for batch data from phishing activities, this prevents errors caused by signal interference, ensuring that subsequent analysis is based on a reliable foundation.

[0156] Step S320: Based on the accuracy of the transmitted data, a visual fish catch information interface is constructed using an interface generation tool. The interface is categorized and displayed according to the accurate single fish weight value and the number of fish detected, resulting in user-interactive interface content.

[0157] The following formula is used to evaluate the overall accuracy of the fish catch information display in the visual fish catch information interface:

[0158] (twenty two)

[0159] In formula (22), This indicates that the interface displays accuracy metrics. This indicates the total number of categories of fish catch information. Indicates the first The weight value of the type of fish caught. Indicates the first The number of fish caught tested, Indicates the total amount of data transmitted. This represents the accuracy coefficient of data transmission.

[0160] To ensure the accuracy of the transmitted data, a visual interface for fish catch information was constructed using an interface generation tool. This tool, based on a web framework like React, works by dynamically rendering components to categorize and display the data. For example, precise individual fish weight values ​​are grouped by species (e.g., bass 0.6kg and carp 0.4kg) and combined with the number of fish caught (e.g., 10 bass) to generate charts and lists. Users can interactively click to view details. On the fishing activity management platform, this creates a dashboard interface that allows users to zoom in and out to view distribution maps, thus intuitively presenting catch information and connecting to the visual output after data verification.

[0161] The system categorizes and displays precise individual fish weight values ​​and the number of fish detected, creating an interactive user interface. This categorization involves data aggregation algorithms, achieved through SQL queries and grouping statistics. For example, it averages weight values ​​and sorts them by quantity, displaying the results in an interactive table. Clicking on a row opens a detailed window, such as a photo of the individual fish. For mixed fish catches, this clearly categorized interface allows users to quickly decide whether to release the fish.

[0162] Step S330: If the user inputs selection information through the visual fish catch information interface, the input parsing tool is used to identify the content of the selection information, and combined with the preset threshold, it is determined whether it is to release the fish back into the body, thus determining the classification result of the user's intention.

[0163] The following formula is used to define the criteria for releasing the fish:

[0164] (twenty three)

[0165] In formula (23), This indicates a binary judgment result indicating whether the fish should be returned to its original body. Indicates the first The weight coefficients of each identification feature Indicates the first The recognition value of each input feature, This represents the total number of recognizable features. This indicates the preset threshold for determining whether to return the fish to its original body.

[0166] The classification results of user intent are obtained using the following formula:

[0167] (twenty four)

[0168] In formula (24), This represents the final determined user intent classification result. Represents the set of all possible intent categories. Indicates the category of intent Content recognition confidence, Indicates the category of intent Threshold matching degree, and These represent the weight parameters for confidence and matching degree, respectively.

[0169] If a user inputs selection information through a visual fish catch information interface, an input parsing tool performs content recognition on the selection information. This tool uses Natural Language Processing (NLP) models to analyze the text input. The principle is to determine the intent by combining keyword matching (e.g., "release") with the context. For example, if a user inputs "select small fish to release," the input parsing tool categorizes it as a release. It then uses a preset threshold to determine whether the fish is being released. If the weight is less than 0.3kg, it is automatically classified as a release, thus determining the user's intent. This logically extends the interface interaction to intent recognition.

[0170] Step S340: Based on the classification results of user intent, a signal triggering tool is used to generate corresponding control signals, and a return instruction sequence is constructed using an instruction generation tool to obtain the instruction content for subsequent execution.

[0171] The corresponding control signal generated based on the classification results of user intent is derived using the following formula:

[0172] (25)

[0173] In formula (25), This indicates a classification based on user intent. The strength of the generated control signal, This indicates the total number of intent categories. Indicates the first The weight coefficients of each category, Indicates the first The signal amplitude corresponding to each category, Indicates the first A classification identifier for user intent. This indicates that the Dirac function is used to match a specific classification. .

[0174] The instructions to be executed subsequently are derived using the following formula:

[0175] (26)

[0176] In formula (26), Indicates the first The instruction content in the sequence of return fish instructions. Indicates the first The trigger signal type of the step, Indicates the first Step parameter configuration vector, This indicates the execution result status of the previous instruction. This indicates the instruction generation function.

[0177] Based on the classification results of user intentions, a signal triggering tool is used to generate corresponding control signals. This tool is based on an event-driven mechanism, responding to classifications using if-then rules. For example, a release intention triggers a "release" signal, which is sent to the execution device. A command generation tool constructs a fish-returning command sequence to obtain the instructions for subsequent execution. This tool uses a scripting language such as Python to generate the fish-returning command sequence, filling parameters with templates. For example, the fish-returning command sequence might include "open the hatch - release the fish - record the log," generating specific instructions for each release category, such as "release fish numbered 1-5." This ensures automation in sustainable fishing activities.

[0178] Preferably, the deep learning-based intelligent fish-keeping and fish-retrieving method proposed in this embodiment includes step S400:

[0179] S410. After obtaining the fish return command sequence, use a command parsing tool to decompose the command content of the fish return command sequence, classify the decomposition results, and determine the operation types and execution order contained in the fish return command sequence.

[0180] The following formula is used to classify the decomposed instruction content and determine the operation type of each instruction:

[0181] (27)

[0182] In formula (27), Indication of instructions Belongs to operation type The probability, Indicates the first Weight vector of class operations transpose, Indicates the first The feature vector of the instruction Indicates the first Bias terms for class operations, Indicates the total number of operation types. Indicates the first Weight vector of class operations transpose, Indicates the first Bias terms for class operations.

[0183] The following formula is used to determine the execution order of each instruction in the fish return instruction sequence:

[0184] (28)

[0185] In formula (28), Indicates the first The position of the instruction in the execution sequence, Indicates the total length of the instruction sequence. Indicates the total number of instructions. Indicates the first Instruction No. The associated weight of each instruction Indicates the first The operation type of the instruction. Indicates an indicator function, This indicates the sequence of instructions for returning to the fish.

[0186] The following formula describes the process of decomposing a fish-returning command sequence using a command parsing tool:

[0187] (29)

[0188] In formula (29), This represents the set of decomposition results obtained after instruction parsing. Indicates the first A decomposed instruction component, Indicates the total number of components in the decomposition. Represents parsing, This represents the original sequence of fish return commands. This represents the set of parameters for the command parsing tool.

[0189] After obtaining the fish-returning instruction sequence, an instruction parsing tool is used to decompose the instruction content of the fish-returning instruction sequence. The instruction parsing tool is based on a syntax analyzer. The principle is to split the sequence into independent units through word segmentation and syntax tree construction. For example, in the scenario of sustainable management of fishing activities, the fish-returning instruction sequence is such as "open the hatch - release the small fish - close the hatch - record the event". The instruction parsing tool first identifies keywords such as "open" and "release", and then decomposes them into operation steps. This ensures the structured processing of the instructions.

[0190] The decomposition results are classified to determine the operation types and execution order contained in the fish return instruction sequence. The classification process involves the application of a rule engine. The principle is to match the decomposition units by pre-defined categories such as "mechanical operation" and "data recording". For example, "opening the hatch" is classified as a mechanical type and ordered as the first execution order, and "releasing small fish" is the second. This forms an orderly operation chain in the fish protection automation system, which connects the logical basis of instruction decomposition to subsequent execution.

[0191] S420. Based on the operation type and execution order, the corresponding start operation is executed using a local processing tool. During the execution process, the number and weight of the returned fish are acquired in real time using a data acquisition tool.

[0192] The number of fish to be returned is determined by the following formula:

[0193] (30)

[0194] In formula (30), Indicates at time The number of fish released in real time. This indicates the total number of operations executed. Indicates the first The weight of the number of fish in each operation. The Dirac function is used to mark the time of operation. Indicates the first The execution time of this operation Indicates the current data collection time;

[0195] The weight of the returned fish is calculated using the following formula:

[0196] (31)

[0197] In formula (31), This indicates the total weight of the fish that were returned. This indicates the total duration of the entire operation cycle. Indicates at time fish body density function, Indicates at time The rate of return volume, This indicates a continuous cumulative calculation over the entire time period.

[0198] In the actual operation of fishing activities, the corresponding opening operation is executed by a local processing tool according to the operation type and execution sequence. Specifically, the local processing tool is implemented using an embedded controller. The principle is to activate the device by sending an electrical signal. For example, for the mechanical type of "opening the hatch", the local processing tool triggers the motor to start in sequence, opening the fish hatch for release. During the execution, the number and weight of the released fish are obtained in real time through a data acquisition tool. This data acquisition tool operates based on a sensor network. The principle is to scan the fish movement and weigh them through counting sensors. For example, when releasing 20 small bass, the local processing tool records the number as 20 and the total weight as 6kg in real time. This step directly extends from instruction execution to data monitoring, ensuring immediate feedback of the release operation.

[0199] S430. For the data on the number and weight of the returned fish, a time stamping tool is used to generate corresponding timestamp information, which is then linked and bound using a data integration tool to obtain the merged fish return event log entries.

[0200] The generated timestamp information is derived using the following formula:

[0201] (32)

[0202] In formula (32), This indicates the generated timestamp information. This refers to the timestamp generation function of the time stamping tool. This indicates the number of fish returned. This indicates the weight of the fish after it has been returned.

[0203] The merged return event log entries are derived using the following formula:

[0204] (33)

[0205] In formula (33), This represents the set of merged return fish event log entries. This indicates the total number of event entries. Indicates the first Information on the number of fish in each event. Indicates the first Fish weight information for each event. Indicates the first Each event is associated with and bound to timestamp information through a data integration tool.

[0206] For the data on the number and weight of the released fish, a time stamping tool is used to generate corresponding timestamp information. The time stamping tool uses the NTP (Network Time Protocol) to synchronize the clock. The principle is to add time tags accurate to the second, such as "2023-10-15 14:30:45", to the data. For example, a timestamp is bound to the data of the 6kg sea bass released, which provides a time dimension for event tracing.

[0207] The process of obtaining merged fish return event log entries through data integration tools involves associating and binding data. It should be noted that the data integration tools are based on database insertion operations. The principle is to bind quantity, weight, and timestamp into JSON objects through key-value pair associations. For example, a log entry {"Event": "Fish Return", "Quantity": 20, "Weight": 6.0kg, "Time": "2023-10-15 14:30:45"} is generated. This forms a complete record in the phishing activity log system, supporting subsequent auditing and report generation, and bringing efficient compliance tracking results to the business.

[0208] Furthermore, the deep learning-based intelligent fishkeeping and fish return method proposed in this embodiment includes step S500:

[0209] S510. Obtain the original record from the return event log entry, parse the content of the original record, match at least one key field according to the preset field rules, and use a word segmentation tool to perform text segmentation on the original record to obtain a structured log field set.

[0210] The key field matching process can be represented by the following formula:

[0211] (34)

[0212] In formula (34), Indicates the first Match scores for key fields, Represents the original set of records. Indicates the first One original record, Representing records With preset field rules The matching function, Indicates the first The weight coefficients of each field.

[0213] The text segmentation process of a word segmentation tool can be described by the following formula:

[0214] (35)

[0215] In formula (35), This represents the set of words obtained after word segmentation. Indicates the first Each word segmentation result This represents a word segmentation tool function used to apply word segmentation algorithm parameters. The raw record text to be processed Perform word segmentation.

[0216] The process of constructing a structured log field set can be represented by the following formula:

[0217] (36)

[0218] In formula (36), This represents the final set of structured log fields. Indicates the first Field names, This indicates the corresponding field value. This represents the total number of structured fields. The entire set is formed by combining all field value pairs through a union operation to create complete structured data.

[0219] The process of retrieving raw records from the fish return event log entries first involves accessing the stored raw data through the log reading interface. For example, in a sustainable fishing activity management system, raw records include JSON-formatted entries such as {"Event": "Fish Return", "Quantity": 15, "Weight": 4.5kg, "Time": "2023-11-20 09:15:30"}. The content of these raw records is parsed by matching preset field rules using regular expressions. For example, if the rules define key fields such as "Quantity" and "Weight", the system will scan the records and extract the matching parts to ensure that at least one key field is identified. Subsequently, a word segmentation tool was used to segment the original records into text, resulting in a structured set of log fields. This word segmentation tool is based on a natural language processing library. The principle is to split the text into words through an algorithm. For example, the sentence "Returning fish event: Released 15 small fish, total weight 4.5kg on 2023-11-20" was segmented to generate a set of fields such as {"Event type": "Returning fish", "Number of fish": 15, "Returned weight": 4.5kg, "Timestamp": "2023-11-20 09:15:30"}. This process transforms raw text into structured data in fishing boat log processing, supporting subsequent refined analysis.

[0220] S520. Based on the structured log field set, classify and process the content of each field. If the field content meets the preset sensitive information standard, use encryption tools to protect the field content and generate an encrypted log field combination.

[0221] Define the sensitivity criteria for log fields using the following formula:

[0222] (37)

[0223] In formula (37), Indicates the first Sensitivity assessment results for each log field. Indicates the field A scoring function for sensitive information detection. This indicates a preset threshold standard for sensitive information. When the sensitivity score of a field's content is greater than or equal to the threshold, it is marked as a sensitive field that needs to be encrypted; otherwise, it remains unchanged.

[0224] The encrypted log field combination is obtained using the following formula:

[0225] (38)

[0226] In formula (38), This represents the encrypted combination of log fields generated after processing. This indicates the total number of fields in the log. Indicates the first The result after processing each field Indicates the first One field, Indicates the first fields Encryption processing is performed. The function is used to determine whether a field contains sensitive information. Based on the determination result, it decides whether to encrypt the field or keep its original content. If the value is 1, then encrypt it to obtain... ;like =0, then retain the original data directly. .

[0227] Based on a structured set of log fields, the system categorizes each field's content. This categorization involves preset categories such as "numerical data" and "time data." The system checks each field individually. If a field's content meets preset sensitive information standards—for example, "number of fish" is considered sensitive due to its connection to compliance reports—then encryption tools are used to protect the field's content. These encryption tools employ the AES (Advanced Encryption Standard) algorithm, which generates an encrypted string using a key. For instance, "15" is encrypted as "U2FsdGVkX1+abc123." This generates an encrypted log field combination, which is then verified for integrity. Verification and calculation confirm that the data has not been tampered with, ensuring the secure transmission of sensitive data in phishing audit scenarios.

[0228] S530. By combining the encrypted log fields, obtain the corresponding log entry identifier, and combine it with the preset format template. Then, use the generation tool to construct new log file data.

[0229] The following formula is used to obtain a unique log entry identifier by performing an XOR operation on the encrypted field and the timestamp, taking the hash, and then taking the modulo of the identifier space:

[0230] (39)

[0231] In formula (39), Indicates the identifier of the generated log entry. Represents a hash function. This represents the encrypted combination of log fields. Represents a timestamp. Modulo operation is represented. Indicates the size of the identifier space.

[0232] The following formula describes the mapping relationship between encrypted fields and identifiers according to a preset template:

[0233] (40)

[0234] In formula (40), Indicates the integration function, This represents the encrypted combination of log fields. This indicates a preset format template. This indicates the identifier of the corresponding log entry.

[0235] The following formula represents how the integrated data is constructed into the final log file data according to the configuration parameters using the generation tool:

[0236] (41)

[0237] In formula (41), This indicates the data in the newly constructed log file. This indicates the generation of utility functions. This represents the integrated data structure. This indicates the generation of configuration parameters.

[0238] The process of obtaining the corresponding log entry identifier by combining encrypted log fields, such as a unique ID like "LOG-20231120-001", and integrating the encrypted fields with the identifier using a preset format template like an XML template, then using a generation tool to construct new log file data. This generation tool operates based on a template engine, and its principle is to generate file content by filling in templates. For example, it outputs a structured file with the following structure: {"ID": "LOG-20231120-001", "Encryption Quantity": "U2FsdGVkX1+abc123", "Weight": 4.5kg}. This connects the logical chain from data protection to file construction in the extended phishing activity monitoring.

[0239] S540. Based on the new log file data, allocate a storage path for each file item, and write the new log file data to the specified location using the storage management tool to obtain the complete log file.

[0240] The following formula is used to allocate the corresponding storage path space based on the size ratio of each log entry:

[0241] (42)

[0242] In formula (42), Indicates the first The storage path capacity allocated to the log file content. Indicates the first The size of the log file content. This indicates the total number of entries in the log file. This indicates the total available storage capacity of the system.

[0243] The following formula is used to evaluate the integrity of log files obtained after storage path allocation and write operations:

[0244] (43)

[0245] In formula (43), Indicates the integrity metrics of the complete log file. Indicates the first A validity identifier for each log data block. This indicates the total number of log data blocks. Indicates the number of standard data blocks that the target log file should contain.

[0246] Based on the new log file data, a storage path is assigned to each file item, using path generation rules such as a date-based directory " / logs / 2023 / 11 / 20 / ". Then, the log file data is written to the specified location using a storage management tool. This storage management tool is implemented using a file system API (Application Programming Interface). The principle is to persist the data to disk through write operations, resulting in the final complete log file storage result. For example, in phishing campaign operations, this supports long-term event tracing and enables rapid retrieval of encrypted logs during compliance checks, resulting in efficient data management.

[0247] Preferably, the deep learning-based intelligent fish-keeping and fish-retrieving method proposed in this embodiment includes step S600:

[0248] S610. Extract persistent data from the complete log file, format the persistent data, match at least one environment-related parameter through preset field rules, and use a parsing tool to compare the persistent data item by item to obtain the preliminary classification results of the environment parameters.

[0249] Persistent data can be extracted from a complete log file using the following formula:

[0250] (44)

[0251] In formula (44), This represents the persistent data set extracted from the complete log file. This indicates the total number of lines in the complete log file. Indicates the first Line log recording, Indicates the first The formatting function for line logs. Indicates the first Row logs and preset field rules The matching degree function.

[0252] The preliminary classification results of environmental parameters are obtained using the following formula:

[0253] (45)

[0254] In formula (45), Indicates the first Preliminary classification results of several environmental parameters Indicates the first A set of rules for class environment parameters. Indicates the first The number of class rules, Indicates the first The weight of each rule, Indicates the first One environmental parameter, Indicates parameters With rules The matching indicator function.

[0255] The process of extracting persistent data from a complete log file begins by reading the log file stored on the disk through a file access interface. For example, in a sustainable fishing activity management system, the complete log file is stored in XML format and contains multiple entries such as {"ID": "LOG-20231120-001", "Encryption Quantity": "U2FsdGVkX1+abc123", "Weight": 4.5kg, "Environmental Parameters": "Water Temperature: 22.5℃, Oxygen Content: 6.2mg / L"}. The system scans the file content and extracts persistent data. Persistent data here refers to stable records that have been written to the disk. The principle is to read the data line by line and convert it into memory objects. After ensuring data integrity, subsequent processing is performed. This forms a bridge from files to operable data in fishing activity monitoring, supporting real-time analysis of environmental data.

[0256] The persistent data is formatted, which involves unifying the data structure. For example, environmental parameters in different formats are converted into standard JSON objects. The system uses conversion tools to process the raw data. For example, "water temperature: 22.5℃" is standardized to {"parameter": "water temperature", "value": 22.5", "unit": "℃"}. This ensures data consistency and facilitates subsequent matching.

[0257] The system matches at least one environmentally relevant parameter by using preset field rules. For example, if the rules define key parameters such as "water temperature" and "oxygen content", the system uses a rule engine to compare persistent data. If the data contains these parameters, it is marked as a match. The rule engine is based on a pattern matching algorithm. The principle is to scan fields using predefined regular expressions. For example, the rule "water temperature: d+℃" will identify the value 22.5℃. In fishing scenarios, this helps to quickly locate environmental factors related to fish survival.

[0258] The persistent data is compared item by item using a parsing tool to obtain preliminary classification results of environmental parameters. The parsing tool is based on a machine learning classifier. The principle is to identify the parameter type by training the model. For example, after inputting the data "oxygen content: 6.2mg / L", the trained model outputs the classification as "chemical parameter". In actual operation, this involves data preprocessing such as normalization. Then, by extracting features and comparing them with the preset categories, results such as {"water temperature": "physical parameter", "oxygen content": "chemical parameter"} are generated, thus providing a basis for compliance inspection.

[0259] S620. Based on the preliminary classification results of environmental parameters, compare each parameter value with the preset compliance threshold. If the parameter value exceeds the preset compliance threshold, it is marked as an anomaly. Use a recording tool to generate an anomaly parameter list and determine the specific distribution of the anomaly parameters.

[0260] The abnormal marking conditions for environmental parameters are defined using the following formula:

[0261] (46)

[0262] In formula (46), Indicates the first Anomaly marker status for each environmental parameter, Indicates the first The actual measured values ​​of each environmental parameter Indicates the first Each environmental parameter has a preset compliance threshold. When the parameter value exceeds the compliance threshold, it is marked as 1 to indicate an abnormality; otherwise, it is marked as 0 to indicate normality.

[0263] The following formula is used to generate a list of abnormal parameters:

[0264] (47)

[0265] In formula (47), This represents a collection of abnormal parameters. Indicates the first One environmental parameter value, Indicates the first Anomaly markers for each environmental parameter, This indicates the total number of environmental parameters.

[0266] The following formula is used to determine the specific distribution of outlier parameters across different categories:

[0267] (48)

[0268] In formula (48), Indicates the first The percentage distribution of class-specific exception parameters. Indicates the first The number of exception parameters, This indicates the total number of abnormal parameters.

[0269] Based on the preliminary classification results of environmental parameters, each parameter value is compared with a preset compliance threshold. In one embodiment, the compliance threshold is a water temperature threshold of 18-25℃. The system checks each value one by one. If 22.5℃ is within the range, it is normal. If it exceeds the range, such as 28.0℃, it is marked as abnormal. The comparison principle here is a combination of simple numerical comparison and logical judgment. In fishing activity management, this can detect potential risks such as fish death caused by high temperature as early as possible.

[0270] If a parameter value exceeds a preset compliance threshold, it is marked as an anomaly. The process of generating an anomaly parameter list using a logging tool involves the logging tool using a log generator to create a list file. For example, the list contains {"Anomaly": "Water Temperature: 28.0℃", "Exceeded Value": 3.0, "Time": "2023-11-20"}, which determines the specific distribution of the anomaly parameters, such as by time or location, thus supporting problem tracing.

[0271] S630. Obtain the corresponding behavior record data through the abnormal parameter list, match the correlation between the abnormal parameters and the behavior record data, and if the matching result shows a correlation, use the update tool to adjust the behavior record to obtain the behavior feedback sequence.

[0272] The degree of match between abnormal parameters and behavioral record data is calculated using the following formula:

[0273] (49)

[0274] In formula (49), Indicates abnormal parameters With behavioral records The degree of matching between them Indicates the total number of parameter dimensions. Indicates the first Weight coefficients for each dimension Indicates abnormal parameters In the Values ​​of each dimension Indicates behavior record In the Values ​​of each dimension This represents a similarity measurement function.

[0275] The updated behavior log is derived using the following formula:

[0276] (50)

[0277] In formula (50), This indicates the updated behavior record. Represents the original behavior record. This indicates that the learning rate is being updated. This indicates the total number of abnormal parameters. Indicates the first The influence weight of each abnormal parameter Indicates the first Adjustment values ​​for each abnormal parameter. This indicates the corresponding number in the original record. Each component.

[0278] The behavioral feedback sequence is derived using the following formula:

[0279] (51)

[0280] In formula (51), Indicates time Behavioral feedback sequence values, The decay factor representing historical behavior. Indicates time The influence of historical behavior Indicates the number of relevant matching results. Indicates the first The contribution coefficient of each matching result. Indicates the first The confidence value of each correlation match.

[0281] The process of obtaining corresponding behavior record data through a list of abnormal parameters, such as fishing boat operation logs queried from a database (e.g., "Fishing time: 09:00-10:00, fish released: 15"), involves the system matching relevant records based on timestamps in the list. The system matches the correlation between abnormal parameters and behavior record data. If the matching results show a correlation, such as the overlap between high-temperature anomalies and fishing activity time, it is considered a correlation. The matching principle here is to calculate the intersection of time and location to ensure logical consistency. In one embodiment, if the matching results show a correlation, an update tool is used to adjust the behavior records, resulting in a behavior feedback sequence. The update tool modifies records, such as adding "Abnormal feedback: High temperature affects release efficiency," and generates a sequence of adjusted operation steps. This improves responsiveness in fishing operations, achieving closed-loop management from anomaly detection to behavior optimization, leading to more sustainable resource utilization.

[0282] Please see Figure 2This invention provides a deep learning-based intelligent fishkeeping system for weighing and retrieving fish, used to execute the aforementioned deep learning-based intelligent fishkeeping system for weighing and retrieving fish. It includes a fish body detection data acquisition module 10, a precise single fish weight value acquisition module 20, a fish retrieval command sequence acquisition module 30, a fish retrieval event log entry determination module 40, a complete log file generation module 50, and a behavior feedback sequence acquisition module 60. The fish body detection data acquisition module 10 acquires fish images collected by an image acquisition device and total weight data collected by a weight sensor. It uses an object recognition model to perform real-time analysis of the fish images, extracts feature vectors from the fish contours, and fuses the total weight data to obtain the number of fish detected and the species identification result. The precise single fish weight value acquisition module 20, based on the number of fish detected and the species identification result, uses a fusion model to process the multimodal fusion of fish contour dimensions and total weight data to determine the single fish weight distribution. If the deviation after fusion exceeds a threshold, it iterates... The system adjusts feature weights to obtain accurate single-fish weight values; the fish return instruction sequence acquisition module 30, after obtaining the accurate single-fish weight value, transmits the accurate single-fish weight value and the number of fish detected through the terminal device, generates a visual fish catch information interface, judges the user's input, and if the user chooses to return the fish, triggers a control signal to obtain the fish return instruction sequence; the fish return event log entry determination module 40, for the fish return instruction sequence, uses the local processing unit to perform the start operation, and records the number and weight data of the returned fish, and integrates timestamp information to determine the fish return event log entry; the complete log file generation module 50, for extracting key fields from the fish return event log entries, processing the fish return event log entries with encryption algorithms, and generating a complete log file; the behavior feedback sequence acquisition module 60, for parsing persistent data through the complete log file, judging environmental compliance parameters, and if the environmental compliance parameters meet the specifications, updating the behavior record to obtain the behavior feedback sequence.

[0283] This embodiment provides a deep learning-based intelligent fish-catching and retrieving method and system. Compared with existing technologies, it acquires fish images and total weight information, uses an object recognition model to analyze fish contour feature vectors in real time, and fuses total weight data to obtain the detection quantity and species results. Then, it uses a fusion model to process the multimodal fusion of contour size and total weight data to determine the weight distribution of individual fish. If the deviation exceeds a threshold, it iteratively adjusts the feature weights to obtain an accurate individual fish weight value. Subsequently, it transmits data to generate a visual interface, triggers a fish-retrieval command sequence based on the user's selection, and the local processing unit executes the activation operation and records the number, weight, and timestamp of fish retrieving, forming an encrypted log file. Finally, it parses the log to determine environmental compliance parameters and updates the behavior record. This embodiment achieves precise and automated management of fish catches, improves the efficiency of fishing activities and environmental sustainability, and ensures the integrity of compliant behavior feedback.

[0284] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A deep learning-based intelligent fishkeeping method for weighing and retrieving fish, characterized in that, Includes the following steps: S100: Obtain the fish body image acquired by the image acquisition device and the total weight data acquired by the weight sensor; use an object recognition model to perform real-time analysis on the fish body image; extract feature vectors from the fish body contour; and fuse the total weight data to obtain the number of fish detected and the species identification result. S200. Based on the number of fish detected and the species identification results, a fusion model is used to process the multimodal fusion of fish outline size and total weight data to determine the weight distribution of a single fish. If the deviation after fusion exceeds the threshold, the feature weights are iteratively adjusted to obtain the accurate weight value of a single fish. S300: After obtaining the precise single fish weight value, the precise single fish weight value and the number of fish detected are transmitted through the terminal device to generate a visual fish catch information interface. The user's input is judged. If the user chooses to put the fish back, a control signal is triggered to obtain the fish return instruction sequence. S400. For the fish return instruction sequence, the local processing unit performs the start operation, while recording the number and weight of the fish returned, and integrating the timestamp information to determine the fish return event log entry. S500: Extract key fields from the fish return event log entries, process the fish return event log entries using an encryption algorithm, and generate a complete log file; S600. Parse the persistent data through the complete log file to determine the environment compliance parameters. If the environment compliance parameters meet the specifications, update the behavior record to obtain the behavior feedback sequence. Step S100 includes: S110. Obtain the fish image captured by the image acquisition device and the total weight data recorded by the weight sensor. Perform noise reduction and edge enhancement processing on the fish image using a pre-established image processing tool to obtain a clear fish outline image and determine the position and boundary range of the fish in the image. The following formula is used to denoise captured fish images: ； in, This represents the pixel values ​​of the image after denoising. Represents the pixel values ​​of the original fish image. Indicates the original fish body image in The pixel value at the pixel coordinates. This represents the Gaussian filter kernel function. This indicates the radius of the filter kernel. and Indicates the pixel coordinate position in the image; The following formula is used to calculate the edge intensity of an image for edge enhancement processing: ； in, Represents the image edge intensity value. This represents an image that has undergone noise reduction processing. Indicates the image in gradient of direction, Indicates the image in Gradient of direction; The following formula is used to determine the position and boundary of the fish in the image: ； in, The set of pixels representing the boundary of the fish's body. This represents the pixel values ​​of the image after edge enhancement. and Indicates the threshold range for boundary detection. , , , These represent the minimum and maximum coordinate boundaries of the fish in the image, respectively; S120. For the fish outline image, an image segmentation tool is used to extract the feature vector of the fish outline, and the feature vector is compared with a preset feature matching database. If the matching degree of the feature vector of the fish outline is higher than the preset threshold, it is judged to be a known fish species, and a preliminary species classification result is obtained. The fish body contour after image segmentation is converted into a feature vector representation that can be used for matching using the following formula: ； in, This represents the extracted fish body contour feature vector. This represents the total number of feature points on the contour. Indicates the first The weight coefficients of each feature point Indicates the first The eigenvalues ​​of each feature point; The cosine similarity between two feature vectors is calculated as the matching degree using the following formula: ； in, The score represents the matching degree of the feature vector. This represents the feature vector of the fish to be identified. This represents the feature vector of a known fish species in the database. Represents the dot product of two vectors. This represents the product of the magnitudes of two vectors; The threshold-based binary classification decision is implemented using the following formula: ； in, This indicates preliminary results of species classification. This represents the preset matching threshold. When the matching degree is greater than or equal to the threshold, the output is 1 to indicate that the fish species is known. When the matching degree is less than the threshold, the output is 0 to indicate that the fish species is unknown. S130. Based on the preliminary species classification results and the total weight data, the number of fish is estimated using a data fusion tool. If the estimated number is inconsistent with the number identified in the fish outline image, the feature matching parameters are adjusted and recalculated to determine the final fish count result. The number of fish can be estimated using the following formula: ； in, This indicates the number of fish estimated through data fusion. This represents the total weight data detected. This represents the total number of categories in species classification. Indicates the first The probability distribution of species in the sample Indicates the first The average individual weight of the species; The adjusted feature matching parameters are obtained using the following formula: ； in, This represents the adjusted feature matching parameters. Indicates the current feature matching parameters. This represents the learning rate used to adjust the parameters. This indicates the number of fish estimated through data fusion; The final fish count was obtained using the following formula: ； in, This indicates the final count of the number of fish. This indicates the total number of candidate fish. Indicates the number of contour recognitions. arrive These represent the weighting coefficients for different constraints. This indicates the average body weight of the fish. Indicates total weight constraint. Indicates the total number of species categories. Indicates the first The number of species, Indicates the first Reference number of species; S140. By associating the final fish count with the preliminary species classification results, a complete fish detection report is generated using a data recording tool to obtain the fish count and species identification results for subsequent processing. The final fish count results are linked and integrated with the preliminary species classification results using the following formula to generate a complete test report: ； in, This represents complete fish body examination report data. The quantitative value representing the quantity detection result. This indicates the accuracy score for species identification. Represents timestamp information. , , These represent the weighting coefficients for each data item.

2. The method for weighing and retrieving fish using a deep learning-based intelligent fish cage according to claim 1, characterized in that, Step S200 includes: S210. Obtain the corresponding fish body outline size information from the fish body detection quantity and species identification results, and use image processing tools to standardize the outline size information to obtain unified size feature data. The following formula maps the outline dimensions of fish of different sizes to a standardized numerical range between 0 and 1: ； in, This represents the standardized fish body outline size characteristics. This indicates the size of the detected individual fish body outline. This represents the minimum size of the body outline of a fish of the same species. This represents the maximum size of the body outline of a fish of the same species; The following formula is used to fuse multi-dimensional fish body outline size information into a unified feature representation: ； in, This represents uniform fish body size characteristic data. Indicates the characteristics of fish body length. Indicates the width characteristics of the fish body. Indicates the height characteristics of the fish body. Weighting coefficients representing length features. The weighting coefficients representing the width feature. Weighting coefficients representing highly distinctive features; S220. Based on the size characteristic data and the total weight data, a preliminary weight distribution estimate is made using a data fusion tool to determine the average weight range of each species. A preliminary weight distribution estimate can be made using the following formula: ； in, Indicates the first Estimated weight of each species, Indicates the first Volumetric characteristic parameters of a species Indicates the first Surface area characteristic parameters of a species Indicates the first Length characteristic parameters of a species , , These represent the contribution coefficients of volume, surface area, and length to weight, respectively. The average weight range for each species is derived using the following formula: ； in, Indicates the first The average weight range for each species Indicates the first Total weight data for each species This represents the total number of dimensional characteristic parameters. Indicates the first The fusion weighting factor for each size feature, Indicates the first Standard deviation of the weight distribution of each species; S230. If the average weight range of each species is determined to be outside the preset deviation threshold, the feature weights are iteratively corrected using a parameter adjustment tool to determine the adjusted weight distribution information. The following formula is used to calculate the species weight deviation judgment index: ； in, Indicates the first The degree of deviation for each species Indicates the first Average weight of each species Indicates the first Preset reference weight for each species This indicates the preset deviation threshold. A value greater than 1 indicates that the weight deviation of this species exceeds the allowable range; The iterative correction process of feature weights is described by the following formula: ； in, Indicates the first The feature in the first The weight values ​​after the next iteration Indicates the first The feature in the first The weight value at the next iteration This represents the learning rate parameter. The loss function represents the loss function on the th The partial derivatives of each feature weight are used to continuously adjust the feature weights using the gradient descent method to reduce the prediction error. The adjusted weight distribution information is represented by the following formula: ； in, Indicates the adjusted weight The probability density distribution, Indicates the number of components in a mixed distribution. Indicates the first The weight coefficients of each distributed component, The mean is variance is The normal distribution is used to describe the final distribution characteristics of species weight through a weighted combination of multiple normal distributions; S240. Based on the adjusted weight distribution information, generate the final single fish weight report using a data recording tool to obtain the accurate single fish weight value. The adjusted precise weight of a single fish is calculated using the following formula through a weighted average method: ； in, This indicates the precise weight of a single fish. Indicates the first Fish weight data at each weight measurement point Indicates the first The adjustment weighting coefficients corresponding to each measurement point Indicates the total number of weight measurement points; The following formula is used to assess the dispersion of the weight distribution and generate a distribution report: ； in, The standard deviation index represents the weight distribution report. Indicates the first The weight measurement of the fish. This represents the average weight of all the fish. This indicates the total number of fish.

3. The intelligent fish-weighing and fish-retrieving method based on deep learning according to claim 1, characterized in that, Step S300 includes: S310. Obtain the precise single fish weight value and fish detection quantity from the terminal device through a data transmission tool, and use a data verification tool to check the integrity of the transmitted content to determine the accuracy of the transmitted data. The following formula is used to count the total number of fish detected by the terminal device and to consider the reliability of the detection: ； in, This indicates the total number of fish tested. Indicates the first The number of fish identified within a detection area. Indicates the first Confidence coefficient of each detection area Indicates the total number of areas detected; The following formula is used by data verification tools to calculate the integrity of the transmitted content, ensuring the accuracy of data transmission: ； in, Indicates data transmission integrity metrics. Indicates the first The number of data packets that are incorrect or lost. Indicates the total amount of data transmitted. Indicates the number of packets that were detected as errors; S320. Based on the accuracy of the transmitted data, an interface generation tool is used to construct a visual fish catch information interface, which is then categorized and displayed according to the precise single fish weight value and the number of fish detected, resulting in user-interactive interface content. The following formula is used to evaluate the overall accuracy of the fish catch information display in the visual fish catch information interface: ； in, This indicates that the interface displays accuracy metrics. This indicates the total number of categories of fish catch information. Indicates the first The weight value of the type of fish caught. Indicates the first The number of fish caught tested, Indicates the total amount of data transmitted. This represents the accuracy coefficient of data transmission; S330. If the user inputs selection information through the visual fish catch information interface, the input parsing tool is used to perform content recognition on the selection information, and a preset threshold is used to determine whether it is to release the fish back into the body, thus determining the classification result of the user's intention. The following formula is used to define the criteria for releasing the fish: ； in, This indicates a binary judgment result indicating whether the fish should be returned to its original body. Indicates the first The weight coefficients of each identification feature Indicates the first The recognition value of each input feature, This represents the total number of recognizable features. This represents the preset threshold for determining whether to return the fish to its original body. The classification results of user intent are obtained using the following formula: ； in, This represents the final determined user intent classification result. Represents the set of all possible intent categories. Indicates the category of intent Content recognition confidence, Indicates the category of intent Threshold matching degree, and These represent the weight parameters for confidence and matching degree, respectively; S340. Based on the classification results of user intent, a signal triggering tool is used to generate corresponding control signals, and a fish return instruction sequence is constructed through an instruction generation tool to obtain the instruction content for subsequent execution. The corresponding control signal generated based on the classification results of user intent is derived using the following formula: ； in, This indicates a classification based on user intent. The strength of the generated control signal, This indicates the total number of intent categories. Indicates the first The weight coefficients of each category, Indicates the first The signal amplitude corresponding to each category, Indicates the first A classification identifier for user intent. This indicates that the Dirac function is used to match a specific classification. ; The instructions to be executed subsequently are derived using the following formula: ； in, Indicates the first The instruction content in the sequence of return fish instructions. Indicates the first The trigger signal type of the step, Indicates the first Step parameter configuration vector, This indicates the execution result status of the previous instruction. This indicates the instruction generation function.

4. The method for weighing and retrieving fish using a deep learning-based intelligent fishkeeping system according to claim 1, characterized in that, Step S400 includes: S410. After obtaining the fish return instruction sequence, use an instruction parsing tool to decompose the instruction content of the fish return instruction sequence, classify the decomposition results, and determine the operation type and execution order contained in the fish return instruction sequence. The following formula is used to classify the decomposed instruction content and determine the operation type of each instruction: ； in, Indication of instructions Belongs to operation type The probability, Indicates the first Weight vector of class operations transpose, Indicates the first The feature vector of the instruction Indicates the first Bias terms for class operations, Indicates the total number of operation types. Indicates the first Weight vector of class operations transpose, Indicates the first Bias terms for class operations; The following formula is used to determine the execution order of each instruction in the fish return instruction sequence: ； in, Indicates the first The position of the instruction in the execution sequence, Indicates the total length of the instruction sequence. Indicates the total number of instructions. Indicates the first Instruction No. The associated weight of each instruction Indicates the first The operation type of the instruction. Indicates an indicator function, This represents the sequence of instructions for returning to the fish. The following formula describes the process of decomposing a fish-returning command sequence using a command parsing tool: ； in, This represents the set of decomposition results obtained after instruction parsing. Indicates the first A decomposed instruction component, Indicates the total number of components in the decomposition. Represents parsing, This represents the original sequence of fish return commands. This represents the set of parameters for the instruction parsing tool; S420. According to the operation type and execution order, the corresponding opening operation is executed using a local processing tool. During the execution, the number of fish released and the weight of the fish released are obtained in real time through a data acquisition tool. The number of fish to be returned is determined by the following formula: ； in, Indicates at time The number of fish released in real time. This indicates the total number of operations executed. Indicates the first The weight of the number of fish in each operation. The Dirac function is used to mark the time of operation. Indicates the first The execution time of this operation Indicates the current data collection time; The weight of the returned fish is calculated using the following formula: ； in, This indicates the total weight of the fish that were returned. This indicates the total duration of the entire operation cycle. Indicates at time fish body density function, Indicates at time The rate of return volume, This indicates a continuous cumulative calculation over the entire time period; S430. For the data on the number and weight of the returned fish, a time stamping tool is used to generate corresponding timestamp information, and a data integration tool is used to associate and bind the data to obtain the merged fish return event log entries. The generated timestamp information is derived using the following formula: ； in, This indicates the generated timestamp information. This refers to the timestamp generation function of the time stamping tool. This indicates the number of fish returned. This indicates the weight of the fish that was returned to its original location. The merged return event log entries are derived using the following formula: ； in, This represents the set of merged return fish event log entries. This indicates the total number of event entries. Indicates the first Information on the number of fish in each event. Indicates the first Fish weight information for each event. Indicates the first Each event is associated with and bound to timestamp information through a data integration tool.

5. The method for weighing and retrieving fish using a deep learning-based intelligent fishkeeping system according to claim 1, characterized in that, Step S500 includes: S510. Obtain the original record from the fish return event log entry, parse the content of the original record, match at least one key field through preset field rules, and use a word segmentation tool to perform text segmentation on the original record to obtain a structured log field set. The key field matching process can be represented by the following formula: ； in, Indicates the first Match scores for key fields, Represents the original set of records. Indicates the first One original record, Representing records With preset field rules The matching function, Indicates the first Weight coefficients for each field; The text segmentation process of a word segmentation tool can be described by the following formula: ； in, This represents the set of words obtained after word segmentation. Indicates the first Each word segmentation result This represents a word segmentation tool function used to apply word segmentation algorithm parameters. The raw record text to be processed Perform word segmentation; The process of constructing a structured log field set can be represented by the following formula: ； in, This represents the final set of structured log fields. Indicates the first Field names, This indicates the corresponding field value. This represents the total number of structured fields. The entire set combines all field value pairs through a union operation to form complete structured data. S520. Based on the structured log field set, classify the content of each field. If the field content meets the preset sensitive information standard, use encryption tools to protect the field content and generate an encrypted log field combination. Define the sensitivity criteria for log fields using the following formula: ； in, Indicates the first Sensitivity assessment results for each log field. Indicates the field A scoring function for sensitive information detection. This indicates a preset threshold standard for sensitive information. When the sensitivity score of a field's content is greater than or equal to the threshold, it is marked as a sensitive field that needs to be encrypted; otherwise, it remains unchanged. The encrypted log field combination is obtained using the following formula: ； in, This represents the encrypted combination of log fields generated after processing. This indicates the total number of fields in the log. Indicates the first The result after processing each field Indicates the first One field, Indicates the first fields Encryption processing is performed. The function is used to determine whether a field contains sensitive information. Based on the determination result, it decides whether to encrypt the field or keep its original content. If the value is 1, then encrypt it to obtain... ;like =0, then retain the original data directly. ; S530. By combining the encrypted log fields, obtain the corresponding log entry identifier, combine the encrypted log field combination with the corresponding log entry identifier using a preset format template, and use a generation tool to construct new log file data. S540. Based on the new log file data, allocate a storage path for each file item, and write the new log file data to the specified location using the storage management tool to obtain the complete log file.

6. The method for weighing and retrieving fish using a deep learning-based intelligent fishkeeping system according to claim 5, characterized in that, In step S530, the unique log entry identifier is obtained by performing an XOR operation on the encrypted field and the timestamp, taking the hash, and then taking the modulo of the identifier space: ； in, Indicates the identifier of the generated log entry. Represents a hash function. This represents the encrypted combination of log fields. Represents a timestamp. Modulo operation is represented. Indicates the size of the identifier space; The following formula describes the mapping relationship between encrypted fields and identifiers according to a preset template: ； in, Indicates the integration function, This represents the encrypted combination of log fields. This indicates a preset format template. This indicates the corresponding log entry identifier; The following formula represents how the integrated data is constructed into the final log file data according to the configuration parameters using the generation tool: ； in, This indicates the data in the newly constructed log file. This indicates the generation of utility functions. This represents the integrated data structure. This indicates the generation of configuration parameters.

7. The method for weighing and retrieving fish using a deep learning-based intelligent fish cage according to claim 6, characterized in that, In step S540, the following formula is used to allocate the corresponding storage path space according to the size ratio of each log entry: ； in, Indicates the first The storage path capacity allocated to the log file content. Indicates the first The size of the log file content. This indicates the total number of entries in the log file. This indicates the total available storage capacity of the system; The following formula is used to evaluate the integrity of log files obtained after storage path allocation and write operations: ； in, Indicates the integrity metrics of the complete log file. Indicates the first A validity identifier for each log data block. This indicates the total number of log data blocks. Indicates the number of standard data blocks that the target log file should contain.

8. The method for weighing and retrieving fish using a deep learning-based intelligent fishkeeping system according to claim 1, characterized in that, Step S600 includes: S610. Extract persistent data from the complete log file, format the persistent data, match at least one environment-related parameter according to preset field rules, and use a parsing tool to compare the persistent data item by item to obtain a preliminary classification result of the environment parameters. S620. Based on the preliminary classification results of environmental parameters, each parameter value of the environmental parameters is compared with the preset compliance threshold. If the parameter value exceeds the preset compliance threshold, it is marked as an abnormal item. A record tool is used to generate an abnormal parameter list to determine the specific distribution of abnormal parameters. S630. Obtain the corresponding behavior record data through the abnormal parameter list, match the correlation between the abnormal parameters and the behavior record data, and if the matching result shows a correlation, use an update tool to adjust the behavior record to obtain a behavior feedback sequence.

9. A deep learning-based intelligent fish-weighing and fish-retrieving system, used to execute the deep learning-based intelligent fish-weighing and fish-retrieving method as described in any one of claims 1 to 8, characterized in that, include: Fish body detection data acquisition module (10) is used to acquire fish body images acquired by image acquisition device and total weight data acquired by weight sensor, and to perform real-time analysis of the fish body images using object recognition model, extract feature vectors for fish body contours, and fuse the total weight data to obtain the number of fish body detections and species identification results; The precise single fish weight value acquisition module (20) is used to process the multimodal fusion of fish body outline size and total weight data using a fusion model based on the number of fish detected and the species identification results, to determine the single fish weight distribution. If the deviation after fusion exceeds the threshold, the feature weight is iteratively adjusted to obtain the precise single fish weight value. The fish return instruction sequence acquisition module (30) is used to acquire the precise single fish weight value, transmit the precise single fish weight value and the number of fish detected through the terminal device, generate a visual fish catch information interface, determine the user's input, and if the user chooses to put the fish back, trigger a control signal to obtain the fish return instruction sequence. The fish return event log entry determination module (40) is used to perform an opening operation using a local processing unit for the fish return instruction sequence, while recording the number and weight of the returned fish and integrating timestamp information to determine the fish return event log entry; The complete log file generation module (50) is used to extract key fields from the fish return event log entries, process the fish return event log entries using an encryption algorithm, and generate a complete log file. The behavior feedback sequence acquisition module (60) is used to parse persistent data through the complete log file, determine the environmental compliance parameters, and update the behavior record if the environmental compliance parameters meet the specifications, thereby obtaining the behavior feedback sequence.

Images