A method, system, equipment and medium for managing aquaculture fish

Abstract

The invention discloses an aquaculture fish management method, system, equipment and medium, and relates to the technical field of fish management, comprising the steps of collecting real-time images of targets to be identified in a breeding pond, wherein the targets to be identified include individual fish and bait residues; embedding a separable stepwise convolution module into a YOLOv8n model to obtain an LSS-YOLOv8 model; inputting the real-time image into the LSS-YOLOv8 model, learning the contextual relationship between the target to be identified and the environment in the real-time image, and obtaining the number of bait residues or the number and state of individual fish; performing text vector mapping on the recognition result to obtain a text vector mapping result; inputting the text vector mapping result into different large language models to obtain different answer strategies and management strategies. The invention combines the YOLO v8n model with the large language model, and uses the rich knowledge base and huge parameter amount of the large language model to provide reasonable suggestions for feeding strategies.

Description

Technical Field

[0001] The present invention relates to the technical field of fish management, and in particular to an aquaculture fish management method, system, equipment and medium. Background Art

[0002] Aquatic products provide important guarantees for my country's economic development and food security. In recent years, the total amount of aquatic products in China has continued to rise, of which aquaculture accounts for 80%. Aquaculture has formed a comprehensive industrial chain while bringing economic benefits to farmers and increasing employment opportunities. By combining emerging technologies and methods, it is possible to improve resource utilization, increase aquaculture efficiency, and increase aquaculture output.

[0003] In recent years, the domestic demand for aquatic products has been growing, and the current output of aquatic products still cannot meet the demand. How to increase aquaculture output and increase the output per unit area has become an urgent problem to be solved. In the fish farming and feeding management of smart fisheries, fish individual identification can observe the health, physiological state and reproductive behavior of individual fish, and formulate aquaculture strategies in a timely manner, which is particularly important to reduce cost investment.

[0004] At present, in smart fisheries, fish management adopts a breeding method based on traditional experience to identify fish individuals, and adopts a fixed feeding strategy during the fish breeding process. This makes it impossible to meet the feeding needs of individual fish under different growth conditions, resulting in increased feed costs, water resource pollution and other problems. Summary of the invention

[0005] The purpose of the present invention is to provide an aquaculture fish management method, system and equipment to address the shortcomings of the above-mentioned prior art, so as to solve the problems in the prior art of adopting a fixed bait feeding strategy during the breeding process, which makes it impossible to meet the feeding needs of individual fish under different growth conditions, resulting in increased bait costs, water resource pollution and other problems.

[0006] The present invention specifically provides the following technical solution: a method for managing aquaculture fish, comprising the following steps:

[0007] Collect real-time images of the targets to be identified in the aquaculture pond, including individual fish and bait residues;

[0008] Embedding a separable stepwise convolution module into a YOLOv8n model to obtain an LSS-YOLOv8 model; the separable stepwise convolution module includes a depthwise convolution and a depthwise atrous convolution;

[0009] The real-time image is input into the LSS-YOLOv8 model, and the contextual relationship between the target to be identified and the environment in the real-time image is learned through a separable stepwise convolution module, and the amount of bait residue, the number of fish individuals and the status of the fish individuals in the current breeding pond are identified according to the contextual relationship;

[0010] Mapping the amount of bait residue or the amount and state of individual fish in the current breeding pond to a text vector to obtain a text vector mapping result in a large language model recognition language format;

[0011] The text vector mapping results are input into the large language model corresponding to different prediction tasks to obtain different answer strategies, and the management strategy of the breeding pond is obtained through different answer strategies.

[0012] Preferably, the step of collecting a real-time image of a target to be identified in a breeding pond comprises the following steps:

[0013] Collect video data of bait residues or fish individuals in the aquaculture pond at different time periods, different lighting conditions, and different viewing angles;

[0014] The acquired video is frame-extracted to obtain image data of each frame, and the obtained image data is screened to eliminate image data with background noise higher than a set noise threshold and viewing angle clarity lower than a set clarity threshold.

[0015] Preferably, embedding the separable stepwise convolution module into the YOLOv8n model to obtain the LSS-YOL Ov8 model comprises the following steps:

[0016] Split the two-dimensional weight kernels of the depthwise convolution and depthwise atrous convolution in the separable stepwise convolution module into multiple cascaded one-dimensional separable weight kernels;

[0017] Embed the multiple cascaded one-dimensional separable weight kernels into the YOLOv8n model, replace the C2f layer with the C2f-LSS layer, and obtain the LSS-YOLOv8 model;

[0018] The parameters and calculation amount of the one-dimensional separable weight kernel are as follows:

[0019]

[0020]

[0021] Among them, the input feature map is F∈R C×h×w , C is the input channel, H represents the height of the feature map, W represents the width of the feature map, k is the convolution kernel size, and d is the dilation rate of the depth-wise dilated convolution.

[0022] Preferably, in the LSS-YOLOv8 model, a sliding weighted combination loss function is used as the loss function of the LSS-YOLOv8 model, specifically:

[0023] The specific expression of the sliding weighted combination loss function f(x) is:

[0024]

[0025] Among them, e is the e-pointing function, the threshold μ is the average IOU value of all bounding boxes of the classification loss function, and x is the variable.

[0026] Preferably, the method further comprises the following steps:

[0027] Construct a self-built bait feeding knowledge base, which contains several question answer pairs for specific task areas (Q i , A i ), where Q i The input of the large language model includes the number of bait residues in the current breeding pond or the number and status of individual fish. i Answer strategies for large language model outputs;

[0028] Fine-tune different large language models using a bait-feeding knowledge base.

[0029] Preferably, the method of fine-tuning the large language model using the bait feeding knowledge base comprises the following steps:

[0030] By answering questions (Q i , A i ) construct each decision task;

[0031] The corresponding question-answer pair is used as the evaluation criterion for the large language model's answer, and each answer strategy of the large language model is compared with the set question-answer pair to evaluate the answer quality score;

[0032] By designing specific prompt words, each question and its corresponding answer text pair is vectorized, and each answer of the large language model is scored according to the accuracy of the vectorized answer and the answer hallucination to obtain the corresponding answer quality score;

[0033] The large language model is optimized according to the answer quality score to obtain increasingly standardized large language model answer standards.

[0034] Preferably, the text vector mapping result is input into a large language model corresponding to different prediction tasks to obtain different answer strategies, and the management strategy of the breeding pond is obtained through the different answer strategies, specifically:

[0035] Inputting the text vector mapping results into a large language model for different prediction tasks to obtain a first answer strategy for each prediction task, wherein the prediction tasks include bait demand prediction, fish counting, and feeding behavior assessment, and the first answer strategy includes bait consumption, fish dynamics, and feeding activity information;

[0036] The first answer strategy is transmitted to a large language model for secondary analysis for summarization, and a second answer strategy is generated by comprehensively analyzing the bait consumption, fish school dynamics and feeding activity information in the first answer strategy, and the management strategy of the aquaculture pond is obtained through the second answer strategy.

[0037] Preferably, the present invention also provides an aquaculture fish management system, comprising:

[0038] The acquisition module is used to collect real-time images of the targets to be identified in the aquaculture pond, including individual fish and bait residues;

[0039] A model building module, used for embedding a separable stepwise convolution module into a YOLOv8n model to obtain an LSS-YOLOv8 model; the separable stepwise convolution module includes a depthwise convolution and a depthwise atrous convolution;

[0040] The recognition module inputs the real-time image into the LSS-YOLOv8 model, learns the contextual relationship between the target to be recognized and the environment in the real-time image through a separable stepwise convolution module, and recognizes the amount of bait residue, the number of fish individuals, and the status of the fish individuals in the current breeding pond according to the contextual relationship;

[0041] A text mapping module is used to map the amount of bait residue or the number and status of individual fish in the current breeding pond into a text vector to obtain a text vector mapping result in a large language model recognition language format;

[0042] The strategy generation module is used to input the text vector mapping results into the large language model corresponding to different prediction tasks to obtain different answer strategies, and obtain the management strategy of the breeding pond through different answer strategies.

[0043] The present invention also provides a computer device, including a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the above-mentioned aquaculture fish management method.

[0044] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned aquaculture fish management method are implemented.

[0045] Compared with the prior art, the present invention has the following significant advantages:

[0046] The present invention designs a large separable stepwise convolution module to be embedded in YOLOv8n to obtain an LSS-YOLOv8 model, inputs an implementation image into the LSS-YOLOv8 model, learns the contextual relationship between bait residues or individual fish and the environment in the real-time image through the separable stepwise convolution module, identifies the number of bait residues or the number and state of individual fish in the current breeding pond according to the contextual relationship, maps the number and state to text vectors, and then transmits them to different large language models to obtain different answer strategies, and manages aquaculture fish through the answer strategies. The present invention utilizes the combination of the LSS-YOLOv8 model and the large language model, transmits the number and state of bait residues or individual fish in the breeding pond obtained by the LSS-YOLOv8 model to different large language models, and provides different answer strategies by utilizing the rich knowledge base and huge parameter amount of the large language model, thereby improving the supply effect of the feeding requirements of individual fish under different growth conditions, providing bait scientifically and accurately, and ensuring that the bait reduces the degree of water resource pollution under the premise of satisfying the growth of fish. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A basic flow chart of an aquaculture fish management method of the present invention;

[0048] Figure 2 Schematic diagram of the separable stepwise convolution module structure in the present invention;

[0049] Figure 3 A schematic diagram of strategy suggestions for the LSS-YOLOv8 model fused with a large language model in the present invention. DETAILED DESCRIPTION

[0050] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0051] like Figure 1 As shown, an embodiment of the present invention provides an aquaculture fish management method, comprising the following steps:

[0052] Step S1: A real-time image of the target to be identified in the breeding pond is collected by a data collection device, and the target to be identified includes individual fish and bait residues. In this embodiment, the data collection device is an underwater camera, and the data collection device is connected to the deep learning model training device and the real-time monitoring computing device, and the real-time monitoring computing device is also connected to the user decision interface device.

[0053] Step S1 is specifically as follows:

[0054] Video data of bait residues or individual fish were collected in the aquaculture pond at different time periods, different lighting conditions, different viewing angles, and different numbers of fish schools.

[0055] The acquired video data is extracted by using Python to obtain the image data of each frame. The obtained image data is filtered to eliminate image data with background noise higher than the set noise threshold and viewing angle clarity lower than the set clarity threshold.

[0056] Step S2: embed the separable stepwise convolution module into the YOLOv8n model, and obtain the LSS-YOLOv8 model by replacing the C2f layer with the C2f-LSS layer. The separable stepwise convolution module includes depthwise convolution and depthwise atrous convolution.

[0057] like Figure 2 As shown in the figure, it is a structural diagram of the large separable stepwise convolution module in the present invention, DW-Conv represents depthwise convolution, DW-D-Conv represents depthwise dilated convolution, k is the convolution kernel size, and d is the dilation rate of depthwise dilated convolution. It can directly use the large-size kernel of the depthwise convolution layer in the module without the need for additional modules, thus reducing memory usage and computational complexity.

[0058] Embed the separable stepwise convolution module into the YOLOv8n model to obtain the LSS-YOLOv8 model, which includes the following steps:

[0059] The 2D weight kernels of the depthwise convolution and depthwise atrous convolution in the (large) separable stepwise convolution module are split into multiple cascaded 1D separable weight kernels.

[0060] Multiple cascaded one-dimensional separable weight kernels are embedded into the YOLOv8n model, and the C2f layer is replaced by the C2f-LSS layer to obtain the LSS-YOLOv8 model.

[0061] The expressions for obtaining the parameters and calculation Flops of the one-dimensional separable weight cores in multiple cascaded one-dimensional separable weight cores are as follows:

[0062]

[0063]

[0064] Among them, the input feature map is F∈R C×h×w , C is the input channel, H represents the height of the feature map, W represents the width of the feature map, k is the convolution kernel size, and d is the dilation rate of the depth-wise dilated convolution.

[0065] The C2f layer of the original feature fusion network in the YOLOv8n model is replaced with C2f-LSS to strengthen the relationship between the target to be tested and the environmental context, thereby obtaining a larger receptive field and improving recognition accuracy.

[0066] In the LSS-YOLOv8 model, the sliding weighted combination loss function is used as the loss function of the LSS-YOLOv8 model, specifically:

[0067] Introduce the sliding weighted combination loss function: The sliding weighted combination loss function adaptively learns the positive sample threshold parameters and the negative sample threshold parameters, takes the average of the IOU values ​​of all bounding boxes as the threshold μ, and uses the weighting function to apply different weights to increase the model's attention to difficult samples while reducing the attention to simple samples. The sliding weighted combination loss function can assign different weights according to the difficulty of sample detection, increase the weight of difficult samples, and reduce the weight of simple samples, thereby improving the generalization ability and recognition accuracy of the model, solving the sample imbalance problem and improving the model generalization performance.

[0068] The sliding weighted combination loss function is used to replace the original classification loss function of the YOLOv8n model. The specific expression of the sliding weighted combination loss function f(x) is:

[0069]

[0070] Among them, e is the e-pointing function, the threshold μ is the average IOU value of all bounding boxes of the classification loss function, and x is the variable.

[0071] Step S3: The real-time image is input into the LSS-YOLOv8 model, and the contextual relationship between the target to be identified and the environment in the real-time image is learned through the separable stepwise convolution module. The amount of bait residue, the number of fish individuals, and the status of the fish individuals in the current breeding pond are identified according to the contextual relationship.

[0072] After the LSS-YOLOv8 target detection model is built, the LSS-YOLOv8 model is trained using a deep learning model training device, where the training device specifically includes a data labeling module, a data loading module, and a model training module.

[0073] The specific training steps are:

[0074] The data labeling module uses a manual method to calibrate the bait, fish individuals, and fish feeding activity status in the image obtained in step 1 using Labelimg; the labeling information includes: class: label category information, x_center and y_center: relative coordinates of the center point of the target box, width, height: relative width and height of the target box, and obtain the txt format labeling file corresponding to each image. Then use Python to randomly divide the labeled data set into training set, validation set, and test set according to 7:2:1.

[0075] The data loading module uses two image processing technologies, logarithmic transformation and adaptive histogram equalization, to effectively enhance the details of dark areas in the image and reasonably suppress the contrast of high-brightness areas. By optimizing the image contrast, the visual effect of the overall image is significantly improved.

[0076] The image is processed using logarithmic transformation and adaptive histogram equalization to obtain an enhanced image. Specifically:

[0077] Logarithmic transformation is used to perform logarithmic operations on the pixel values ​​in the image, stretching the dynamic range of dark areas and making the details that were previously difficult to detect appear.

[0078] Adaptive histogram equalization method is used to automatically adjust the histogram according to the local characteristics of the image, increase the contrast of the image, and make the transition between dark areas and high-brightness areas more natural.

[0079] The enhanced image is constructed by stretching the dynamic range of dark areas and increasing the contrast of the image.

[0080] In order to enrich the diversity of original data, the data loading module also uses Mosaic data enhancement method, specifically:

[0081] By randomly cropping and splicing multiple images to generate new training samples, the original data set is expanded. This data augmentation method not only improves the generalization ability of the model, but also helps reduce the risk of overfitting, enabling the model to better learn the essential characteristics of the image during training.

[0082] The model training module uses the annotated and enhanced images as input data and inputs them into the LSS-YOLOv8 model for training, and finally obtains the trained LSS-YOLOv8 model.

[0083] Step S4: Perform text vector mapping on the amount of bait residue or the number and status of individual fish in the current breeding pond to obtain a text vector mapping result in a large language model recognition language format.

[0084] In this embodiment, the large language model includes: Kimi, QWen2, ChatGLM, LLaMA and other open source large language models.

[0085] These open source large language models have powerful natural language understanding and generation capabilities. Taking the above large language model as the base model, the model is fine-tuned in a supervised manner by building a bait feeding knowledge base. The prompt engineering specification large language model answers sentences, and the answers are scored according to the answer accuracy and answer hallucination to obtain the corresponding answer quality score. The model is optimized according to the answer quality score so that the fine-tuned large language model can better adapt to the task.

[0086] Text vector mapping improves the large language model's ability to understand recognition results by mapping them into commonly used conversational language formats, so as to better provide strategic recommendations: for example, "bait: 42" is mapped to "the remaining number of bait particles in the current breeding pond is 42", "fish: 23" is mapped to "the number of fish individuals in the current breeding pond is 23", and "Active / UnActive" is mapped to "the feeding activity status of fish individuals in the current fish pond is active / inactive", etc.

[0087] Step S5: Input the text vector mapping result into the large language model corresponding to different prediction tasks to obtain different answer strategies, and obtain the management strategy of the breeding pond through different answer strategies.

[0088] The large language model needs to be fine-tuned, including the following steps:

[0089] Build a self-built bait feeding knowledge base, which contains several question answer pairs for specific task areas (Q i , A i ), where the question answer is (Q i , A i ) includes the intensive fish farming knowledge base, the actual farming experience knowledge base of farmers, the fish farming water environment knowledge base, farming manuals and expert advice. These data are carefully screened and sorted to ensure that the model can learn accurate and useful information during the fine-tuning process, and the different large language models are fine-tuned using the bait feeding knowledge base.

[0090] Using the bait feeding knowledge base to fine-tune different large language models includes the following steps:

[0091] Data preparation: Extract task-related data from the bait feeding knowledge base and pre-process the data, such as cleaning, deduplication, and formatting, to ensure data quality and consistency.

[0092] Get fine-tuning strategies, including choosing appropriate fine-tuning objectives, loss functions, and learning rates, to guide the model to better adapt to specific tasks.

[0093] Custom decision task set 1 ,D 2 ...,D n}, answer the corresponding questions (Q i , A i ) to construct each decision task; where Q i The input of the large language model includes the number of bait residues in the current breeding pond or the number and status of individual fish. i is the answer strategy for the output of the large language model, D 1 , D 2 and D n They are the first task, the second task and the nth task respectively.

[0094] The corresponding question-answer pair is used as the evaluation criterion for the large language model's answer. Each answer strategy of the large language model is compared with the set question-answer pair to evaluate the answer quality score. If the quality score is lower than the target quality score, continue this iterative training.

[0095] Use prompt engineering technology to standardize the answer standards of large language models. By designing specific prompt words to vectorize each question and the corresponding answer text pair, the model is guided to generate more accurate and satisfactory answers. The large language model is scored based on the accuracy of the vectorized answers and the answer hallucinations to obtain the corresponding answer quality scores. The large language model is optimized based on the answer quality scores to obtain continuously standardized large language model answer standards.

[0096] According to the task requirements, design prompts that can guide the model to generate correct answers; first, set the task requirements of the large language model through the prompt method of role setting: "Your role is {intensive aquaculture expert}, and the prior knowledge you need is {the remaining amount of bait in the current fish pond, the number of fish individuals, and the information on the active state of fish feeding}. At the same time, you need to combine the {feeding knowledge base} to complete the {give feeding strategy suggestions task}."

[0097] Answer normalization: By setting question-answer pairs, the model's answers are made more consistent with grammatical rules, logical structure, and task requirements. The specific examples are as follows:

[0098] The first canonical answer:

[0099] Q: The number of bait particles remaining in the fish pond is 6, the number of fish in the pond is 40, and the fish feeding activity is active. Please give suggestions for the next feeding strategy.

[0100] A: The fish are still actively feeding, and the remaining bait particles in the fish pond are insufficient. Since the fish are actively feeding and there are few remaining bait particles, it indicates that the fish have a high demand for bait. You can consider increasing the amount of feed to meet the needs of the fish. Through analysis, the amount of feed this time should be increased to 1.5 times the amount of the previous feed.

[0101] The second canonical answer:

[0102] Q: The current number of remaining bait particles in the fish pond is 62, the number of fish in the fish pond is 40, the fish feeding activity status is inactive, and the last feeding time was 8:20. Please give suggestions for the next feeding time.

[0103] A: According to the current situation, it is recommended that the next feeding time can be postponed. Since the fish feeding activity is inactive, you can consider postponing the next feeding for 2-3 hours based on the current time. Therefore, it is recommended to set the next feeding time between 10:20 and 11:20. The general daily feeding rate is 1% to 5%, and feeding is 2 to 3 times a day. For your fish pond, since the fish are not active in feeding, you can consider reducing the number of feedings. If you feed once a day, it is recommended to do it at 10 to 11 am. If you choose to feed twice a day, the first time is around 9 am and the second time is 3 to 4 pm. At the same time, pay attention to the feeding situation of the fish. If the fish activity increases, you can adjust the feeding time appropriately. If the bait particles in the fish pond are consumed slowly, you can reduce the feeding amount appropriately to avoid bait waste and water quality deterioration.

[0104] like Figure 3 The figure shows a schematic diagram of strategy recommendations given by LSS-YOLOv8 combined with a large language model.

[0105] Among them, the text vector mapping results are input into the large language model corresponding to different prediction tasks to obtain different answer strategies, and the management strategy of the breeding pond is obtained through different answer strategies, which are specifically:

[0106] The text vector mapping results are input into a large language model for different prediction tasks to obtain a first answer strategy for each prediction task, which includes bait demand prediction, fish counting, and feeding behavior evaluation. The first answer strategy includes bait consumption, fish dynamics, and feeding activity information.

[0107] The first answer strategy is transmitted to a large language model for secondary analysis for summary. The second answer strategy is generated by comprehensively analyzing the information on bait consumption, fish school dynamics and feeding activities in the first answer strategy. The management strategy of the breeding pond is obtained through the second answer strategy for aquaculture fish management.

[0108] The basic equipment carried by the present invention is described below, specifically:

[0109] The deep learning model training equipment includes Labelimg data labeling module, data loading module and model training module.

[0110] The trained LSS-YOLOv8 model is deployed on the real-time monitoring computing device, and the first answer strategy and the second answer strategy in the above method are generated by the large language model and the "Zhiyu" large language model connected through the API interface.

[0111] The user decision interface device is used to receive and display the second answer feeding decision suggestions generated by the "Smart Fishing" large language model based on actual data analysis, and display the detection result screen in real time synchronously.

[0112] Among them, the user decision interface device includes a detection result module and a feeding strategy module. The detection result is used to display the detection result screen of the LSS-YOLOv8 model for real-time transmission data. The feeding decision module is used to display the second answer feeding strategy recommendation generated after the "Smart Fishing" large language model, which is convenient for farmers to make reasonable feeding decisions.

[0113] The function of the detection result module is to intuitively display the recognition results of the LSS-YOLOv8 model on the real-time transmission data, annotate the real-time image according to the recognition results, and send it to the user end for display. The user end can clearly see the model's detection results of the amount of bait residue, the number of fish individuals, and the active feeding status of fish in the video data. These images are highlighted and counted to make the detection results clear at a glance, allowing users to quickly understand the current situation of the farm.

[0114] The feeding strategy module focuses on displaying the second-answer feeding strategy recommendations generated after summary analysis by the "Smart Fishing" large language model, including key decision-making factors such as the time, quantity and frequency of feeding, and corresponding to the identified amount of residual bait, the number of fish individuals and the status of individual fish in the current breeding pond, and sent to the user end for display. These feeding strategy recommendations are based on real-time monitoring data and multi-model comprehensive analysis results, and aim to provide farmers with scientific and reasonable feeding guidance.

[0115] Based on the above method, the present invention provides an aquaculture fish management system, which includes a collection module, a model building module, a recognition module, a text mapping module and a strategy generation module.

[0116] Among them, the acquisition module is used to collect real-time images of the targets to be identified in the breeding pond, and the targets to be identified include fish individuals and bait residues; the model construction module is used to embed the separable stepwise convolution module into the YOLO v8n model to obtain the LSS-YOLOv8 model; the separable stepwise convolution module includes deep convolution and deep void convolution; the recognition module is used to input the real-time image into the LSS-YOLOv8 model, learn the contextual relationship between the target to be identified and the environment in the real-time image through the separable stepwise convolution module, and identify the number of bait residues, the number of fish individuals and the status of fish individuals in the current breeding pond according to the contextual relationship; the text mapping module is used to map the number of bait residues or the number and status of fish individuals in the current breeding pond to text vectors, and obtain text vector mapping results with large language model recognition language format; the strategy generation module is used to input the text vector mapping results into the large language model corresponding to different prediction tasks, obtain different answer strategies, and obtain the management strategy of the breeding pond through different answer strategies.

[0117] The present invention also provides a computer device, including a memory and a processor. The memory stores a program. When the program is executed by the processor, the processor executes the steps of an aquaculture fish management method.

[0118] In accordance with the disclosed embodiments, a computing device may communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth communications, etc.), or with any device (e.g., routers, modems, etc.) that enables a computing device to communicate with one or more other computing devices.

[0119] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, a method for monitoring transplanting operation quality information is implemented.

[0120] According to the disclosed embodiments, the storage medium may be a non-volatile computer-readable storage medium, such as but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0121] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. For technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for managing aquaculture fish, characterized in that: The steps include: Collect real-time images of the targets to be identified in the aquaculture pond, including individual fish and bait residues; Embedding a separable stepwise convolution module into a YOLOv8n model to obtain an LSS-YOLOv8 model; the separable stepwise convolution module includes a depthwise convolution and a depthwise atrous convolution; The real-time image is input into the LSS-YOLOv8 model, and the contextual relationship between the target to be identified and the environment in the real-time image is learned through a separable stepwise convolution module, and the amount of bait residue, the number of fish individuals and the status of the fish individuals in the current breeding pond are identified according to the contextual relationship; Mapping the amount of bait residue or the amount and state of individual fish in the current breeding pond to a text vector to obtain a text vector mapping result in a large language model recognition language format; The text vector mapping results are input into the large language model corresponding to different prediction tasks to obtain different answer strategies, and the management strategy of the breeding pond is obtained through different answer strategies; The method of embedding the separable stepwise convolution module into the YOLOv8n model to obtain the LSS-YOLOv8 model includes the following steps: Split the two-dimensional weight kernels of the depthwise convolution and depthwise atrous convolution in the separable stepwise convolution module into multiple cascaded one-dimensional separable weight kernels; Embedding multiple cascaded one-dimensional separable weight kernels into a YOLOv8n model to obtain an LSS-YOLOv8 model; The parameters and calculation amount of the one-dimensional separable weight kernel are as follows: Among them, the input feature map is F∈R C×h×w , C is the input channel, H represents the height of the feature map, W represents the width of the feature map, k is the convolution kernel size, and d is the dilation rate of the depth-wise dilated convolution.

2. The aquaculture fish management method according to claim 1, characterized in that: The real-time image of the target to be identified in the breeding pond is collected, comprising the following steps: Collect video data of bait residues or fish individuals in the aquaculture pond at different time periods, different lighting conditions, and different viewing angles; The acquired video is frame-extracted to obtain image data of each frame, and the obtained image data is screened to eliminate image data with background noise higher than a set noise threshold and viewing angle clarity lower than a set clarity threshold.

3. The aquaculture fish management method according to claim 1, characterized in that: In the LSS-YOLOv8 model, a sliding weighted combination loss function is used as the loss function of the LSS-YOLOv8 model, specifically: The specific expression of the sliding weighted combination loss function f(x) is: Among them, e is the e-pointing function, the threshold μ is the average IOU value of all bounding boxes of the classification loss function, and x is the variable.

4. The aquaculture fish management method according to claim 1, characterized in that: The method further comprises the steps of: Construct a bait feeding knowledge base, which contains several question answer pairs for specific task areas (Q i , A i ), where Q i The input of the large language model includes the number of bait residues in the current breeding pond or the number and status of individual fish. i Answer strategies for large language model outputs; Fine-tune different large language models using a bait-feeding knowledge base.

5. The aquaculture fish management method according to claim 4, characterized in that: The method of fine-tuning different large language models using the bait feeding knowledge base includes the following steps: By answering questions (Q i , A i ) construct each decision task; The corresponding question-answer pair is used as the evaluation criterion for the large language model's answer, and each answer strategy of the large language model is compared with the set question-answer pair to evaluate the answer quality score; By designing specific prompt words, each question and its corresponding answer text pair is vectorized, and each answer of the large language model is scored according to the accuracy of the vectorized answer and the answer hallucination to obtain the corresponding answer quality score; The large language model is optimized according to the answer quality score to obtain increasingly standardized large language model answer standards.

6. The aquaculture fish management method according to claim 1, characterized in that: The text vector mapping result is input into the large language model corresponding to different prediction tasks to obtain different answer strategies, and the management strategy of the breeding pond is obtained through different answer strategies, which is specifically: Inputting the text vector mapping results into a large language model for different prediction tasks to obtain a first answer strategy for each prediction task, wherein the prediction tasks include bait demand prediction, fish counting, and feeding behavior assessment, and the first answer strategy includes bait consumption, fish dynamics, and feeding activity information; The first answer strategy is transmitted to a large language model for secondary analysis for summarization, and a second answer strategy is generated by comprehensively analyzing the bait consumption, fish school dynamics and feeding activity information in the first answer strategy, and the management strategy of the aquaculture pond is obtained through the second answer strategy.

7. An aquaculture fish management system, characterized in that: include: The acquisition module is used to collect real-time images of the targets to be identified in the aquaculture pond, including individual fish and bait residues; A model building module, used for embedding a separable stepwise convolution module into a YOLOv8n model to obtain an LSS-YOLOv8 model; the separable stepwise convolution module includes a depthwise convolution and a depthwise atrous convolution; The recognition module inputs the real-time image into the LSS-YOLOv8 model, learns the contextual relationship between the target to be recognized and the environment in the real-time image through a separable stepwise convolution module, and recognizes the amount of bait residue, the number of fish individuals, and the status of the fish individuals in the current breeding pond according to the contextual relationship; A text mapping module is used to map the amount of bait residue or the number and status of individual fish in the current breeding pond into a text vector to obtain a text vector mapping result in a large language model recognition language format; The strategy generation module is used to input the text vector mapping results into the large language model corresponding to different prediction tasks to obtain different answer strategies, and obtain the management strategy of the breeding pond through different answer strategies; The method of embedding the separable stepwise convolution module into the YOLOv8n model to obtain the LSS-YOLOv8 model includes the following steps: Split the two-dimensional weight kernels of the depthwise convolution and depthwise atrous convolution in the separable stepwise convolution module into multiple cascaded one-dimensional separable weight kernels; Embedding multiple cascaded one-dimensional separable weight kernels into a YOLOv8n model to obtain an LSS-YOLOv8 model; The parameters and calculation amount of the one-dimensional separable weight kernel are as follows: Among them, the input feature map is F∈R C×h×w , C is the input channel, H represents the height of the feature map, W represents the width of the feature map, k is the convolution kernel size, and d is the dilation rate of the depth-wise dilated convolution.

8. A computer device, characterized in that: It comprises a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of an aquaculture fish management method as claimed in any one of claims 1 to 6.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of an aquaculture fish management method according to any one of claims 1 to 6 are implemented.

Images