A precision feeding system for fishery

Abstract

The application discloses a kind of fishing precision feeding system, it is related to aquaculture technical field, including: water identification device, edge server, intelligent feeding device, water identification device, edge server, intelligent feeding device are sequentially connected;Water identification device is used to obtain fish movement containing video stream information and is sent to edge server through local area network;Edge server is used to identify and count target fish in video stream information through pre-trained YOLOv8 network, determine the real-time proportion of target fish in different regions, obtain the feed consumption data of different regions and send to corresponding intelligent feeding device;Intelligent feeding device is based on the feed consumption data received to complete feeding.The application uses deep learning computer vision technology, and more accurately feeds by the distribution of fish in each region;Through the way of intelligent precision feeding, fish feed is saved, and water pollution caused by excessive feeding is prevented.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fish farming, more particularly to a fishery precision feeding system. BACKGROUND

[0002] At present, most of the fish farms in China use artificial feeding method, and one technical personnel is needed for an average of 100 mu of land. Artificial feeding of feed mainly relies on the working experience of workers, and the required feeding weight is roughly calculated according to the rough estimation of the number of fish groups, the eating habits, speed and amount of different fish groups. Artificial feeding not only increases the production cost but also has a certain deviation from the demand of fish for feed, and has a high requirement for the experience and proficiency of the workers. At the same time, in the era of full informationization and intelligentization, a lot of products developed for intelligent feeding of feed have emerged. For example, the timing and fixed-point feeding system can be designed by the operator on the platform in advance. The feeding time and weight are designed in advance. However, the operator still needs to modify the feeding time, frequency and weight of the feed according to the growth status of the fish after a period of time, but it still cannot realize comprehensive automatic feeding. Although some feeding machines on the market can reduce the workload of workers and the use of feed bags by adjusting the feeding rate and feeding range, but the excessive feeding for the full growth of fish will cause waste of feed and increase the cost of feed and the pressure of water quality protection.

[0003] In actual fish farming, fish distribution is often uneven in fish ponds, especially in large water areas, and people usually adopt the strategy of average and excessive feeding in the water area, which will cause waste of feed and increase the cost, and the fish feed that is not eaten will also cause water eutrophication, pollute the water quality, affect the living environment of fish, and increase the mortality rate of fish.

[0004] Therefore, how to comprehensively consider the distribution of fish in each area, the growth rule of fish, the date, the temperature, the type of feed and other factors to construct an automatic system that can accurately and scientifically feed is a technical problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the present application provides a fishery precision feeding system, which uses deep learning computer vision technology to accurately feed through the distribution of fish in each area. The intelligent precision feeding method saves fish feed and prevents water pollution caused by excessive feeding.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme:

[0007] A fishery precision feeding system includes: an underwater recognition device, an edge server, and an intelligent feeding device, wherein the underwater recognition device, the edge server, and the intelligent feeding device are connected in sequence;

[0008] An underwater recognition device is used to obtain video stream information containing fish movements and send it to the edge server via the local area network;

[0009] The edge server is used to identify and count target fish in the video stream information using a pre-trained YOLOv8 network, determine the real-time proportion of target fish in different areas, obtain feed usage data for different areas, and send it to the corresponding intelligent feeding device;

[0010] The intelligent feeding device completes feeding based on the received feed usage data.

[0011] Optionally, the underwater identification device includes an above-water part for power supply and data transmission, and an underwater part for acquiring video stream information;

[0012] The above-water part includes: an above-water camera for monitoring the water surface conditions, an electric power supply module for powering the above-water camera, and a communication module for temporarily storing and wirelessly transmitting the data collected by the above-water camera;

[0013] The underwater part includes: an underwater camera for acquiring underwater image information and an anchoring device for stabilizing the underwater camera.

[0014] Optionally, the power supply module includes a solar panel, a battery and a step-down power supply module connected in sequence, and the output end of the step-down power supply module is connected to the underwater camera.

[0015] Optionally, the communication module is a Raspberry Pi PICow development board.

[0016] Optionally, the edge server includes: an image acquisition module, a model improvement module, a counting module, and a feed dosage determination module;

[0017] An image acquisition module is used to select target fish images with one or more complete outlines from the video stream information, clean the target fish images, and complete image annotation with the whole fish, fish head, and fish tail to obtain a preprocessed image;

[0018] The model improvement module is used to improve the selected YOLOv8 basic model to obtain a pre-trained YOLOv8 network;

[0019] The counting module uses a pre-trained YOLOv8 network to identify each frame of pre-processed images, obtain the number of target fish in each area, and determine the real-time ratio of target fish in different areas;

[0020] The feed dosage determination module obtains feed dosage data for each area based on the feed dosage required for the growth of target fish in different periods and a mathematical model.

[0021] Optional model improvement modules include: attention mechanism introduction submodule, loss function adjustment submodule, and noise reduction submodule;

[0022] The attention mechanism introduces a submodule to add a SimAM attention module between the connection layer and the convolution layer in the Neck part of the YOLOv8 basic model to unify the weights;

[0023] The loss function adjustment submodule is used to reduce the loss weight of the small object detection layer in the YOLOv8 basic model and increase the loss weight of the medium and large object detection layers;

[0024] The anti-noise submodule obtains a pre-trained YOLOv8 network by using the Label Smoothing strategy during model training.

[0025] Optionally, the counting module includes: a local optimization submodule, an averaging submodule, and a ratio acquisition submodule;

[0026] The local optimization submodule identifies and counts the fish head, fish tail, and whole fish in each frame, and takes the maximum value of the three as the number of fish in the frame;

[0027] The averaging submodule is used to average the number of fish in the video stream and the total number of frames in the video stream to obtain the average number of fish in each local area;

[0028] The proportion acquisition submodule obtains the proportion of the total number of fish in each local area to the total number of fry in the fish pond based on the average number of fish in each local area.

[0029] Optionally, the feed usage data of each region obtained by the feed usage determination module is expressed as:

[0030] Q i =kapR i (i=1,2,3,…,m);

[0031] Where: Q i represents the feed amount in the i-th area; k represents the feed coefficient, that is, the feed amount required to produce unit mass of fish meat; a represents the feed ratio, p represents the planned fish production in the fish pond, and R i Represents the real-time proportion of target fish in the i-th area.

[0032] Optionally, the intelligent feeding device includes: a host computer and a feeding machine;

[0033] The upper computer is used to send feeding instructions to the feeding machine at the set feeding time, and control the feeding quantity and feeding position by controlling the working time and direction of the feeding machine;

[0034] The feeding machine completes the feeding operation based on the feeding instruction.

[0035] Through the above technical solutions, it can be seen that compared with the prior art, the present invention provides a fishery precision feeding system, which divides the aquaculture waters into zones and deploys underwater identification devices in each water area, obtains underwater fish activity video information through a camera and transmits the data to the edge server through a wireless network, uses a pre-trained neural network to identify and count the video stream information, determines the real-time proportion of target fish in different areas, and gives the feed dosage for different areas according to a mathematical model and transmits it to the feeder. The present invention uses deep learning computer vision technology, which can feed more accurately based on the distribution of fish in each area, and has the advantage of intelligence; through intelligent and precise feeding, fish feed is saved, and water pollution caused by excessive feeding can be prevented. The water part adopts a solar power supply mode to further save energy, which has the advantages of economy and greenness. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0037] Figure 1 This is a structural diagram of the fishery precision feeding system provided by the present invention;

[0038] Figure 2 This is a schematic diagram of image annotation provided by the present invention;

[0039] Figure 3 The approximate feeding weight of the fish feed provided by the present invention in different months. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0041] The embodiment of the present invention discloses a fishery precision feeding system, such as Figure 1As shown, it comprises: a water identification device, an edge server, and an intelligent feeding device, which are connected in sequence.

[0042] The water identification device is used to obtain video stream information containing fish movement and send it to the edge server through a local area network.

[0043] The edge server is used to identify and count target fish in the video stream information through a pre-trained YOLOv8 network, determine the real-time proportion of target fish in different areas, obtain the feed consumption data of different areas, and send it to the corresponding intelligent feeding device.

[0044] The intelligent feeding device completes feeding based on the received feed consumption data.

[0045] 1. Water identification device

[0046] The water identification device includes a water surface part for power supply and data transmission, and an underwater part for obtaining video stream information.

[0047] The water surface part includes a water surface camera for monitoring the water surface, an electric energy supply module for powering the water surface camera, and a communication module for temporarily storing and wirelessly sending data collected by the water surface camera. The electric energy supply module includes a solar panel, a storage battery, and a step-down power supply module connected in sequence, and the output end of the step-down power supply module is connected with the water surface camera. Since the distance between the water surface platform and the shore is far, it is inconvenient to draw a cable from the shore for power supply, so this embodiment uses a combination of solar panels and storage batteries to power the devices deployed on the water surface, which can ensure power supply in rainy weather and does not need maintenance for a long time. The communication module is a Raspberry Pi picow development board, which can realize the temporary storage of camera data and the wireless transmission of data, and can also be expanded to use a communication module for data transmission.

[0048] The underwater part includes an underwater camera for obtaining underwater image information and an anchoring device for stabilizing the underwater camera. The anchoring device can prevent the underwater camera from moving under the action of wind and waves. The underwater camera can obtain underwater image information and collect relevant data for fish population identification and monitoring.

[0049] 2. Edge server

[0050] The edge server processes the video stream transmitted from the water through the RTSP protocol. Since multiple regions of video stream need to be processed in the same time period, this embodiment adopts a multi-process parallel method for video stream detection.

[0051] 2.1 Image acquisition module

[0052] Select one or more target fish images with complete outlines from the video stream information, improve the quality of the selected images, clean the target fish images and use the online website Make Sense to annotate the images, refer to Figure 2 The whole fish, fish head, and fish tail are labeled as “fish,” “head,” and “tail,” respectively, to obtain preprocessed images.

[0053] 2.2 Model Improvement Module

[0054] With the rapid development of deep learning in recent years, a series of image recognition models based on convolutional neural networks have emerged. In the field of fish population counting, scholars have adopted different detection methods based on the different detection objects and detection environments. The YOLO series of neural network models has achieved high mAP values ​​in fish detection. The YOLOv8 model can effectively handle various interference factors in the aquaculture environment, such as water waves, occlusions, and background clutter, improving the accuracy and robustness of fish detection and counting. Because the YOLO series is a one-stage model, it can more quickly recognize and respond to target images, providing real-time feedback to fish farmers. It is relatively mature in practical applications and relatively low in cost. Therefore, this embodiment chooses YOLOv8 as the basic model for deep learning. YOLOv8 has five basic models: n, s, m, l, and x. Since this embodiment requires real-time detection of target fish in water, the lighter s model can be selected to speed up deployment.

[0055] However, when detecting fish in multiple areas, the characteristics of fast speed and strong noise resistance are required, and most of the farmed fish targets in the water are medium and large targets. Therefore, the model improvement module of this embodiment provides the following three improvement measures. The test platform is Windows system and the graphics card is RTX3050.

[0056] (1) Attention mechanism introduces submodule

[0057] When performing real-time detection of the number of fish, multiple devices are required to send video stream signals to the server at the same time, which can easily cause information overload. However, this embodiment introduces an attention mechanism to focus the input information on more important information, thereby improving the detection efficiency of the neural network model.

[0058] SimAM is a parameter-free attention mechanism for convolutional neural networks. It is an attention module that unifies weights. Compared to other attention mechanisms, SimAM does not require parameters to the original network, but instead infers 3D attention weights from feature maps. This module avoids excessive effort in structural adjustments, making it flexible and effective in visual tasks.

[0059] In this embodiment, the SimAM attention module is added between the twelfth layer (connection layer) and the thirteenth layer (convolution layer with 3*3) in the Neck part of the YOLOv8s model. Compared with the original YOLOv8s model, the addition of SimAM can improve the detection speed to a certain extent without changing the mAP@.5, which proves the feasibility of this improvement.

[0060] (2) Loss function adjustment sub-module

[0061] Since the detection of target fish underwater will be limited by insufficient light, the shooting effect of organisms at a certain distance is poor, and the reference significance is poor. In order to further improve the detection speed, the loss weight of the small target detection layer in the YOLOv8 basic model is reduced and the loss weight of the medium target detection layer and the large target detection layer is increased in the loss function.

[0062] Specifically, in the default loss function of YOLOv8, the loss weights from the lowest layer (smallest feature size) to the highest layer (largest feature size) are 4.0, 1.0 and 0.4 respectively; after modification, the corresponding weights are 3.0, 2.0 and 0.8. Compared with the original YOLOv8s model, the mAP@.5 and running speed are basically unchanged, but the recall rate of the model after changing the loss function is improved, which is beneficial to counting fish and detecting the target position of fish as completely as possible.

[0063] The calculation formula of the recall rate is as follows:

[0064]

[0065] In the formula, TP represents a positive sample predicted as a positive sample, and FN represents a negative sample predicted as a positive sample.

[0066] (3) Anti-noise module

[0067] The present embodiment needs to detect fish underwater, and the noise for obtaining information in the underwater environment is relatively large, so the model established needs to have certain anti-noise ability, and at the same time, in order to make up for the problem of less information entropy in simple classification, the Label Smoothing strategy is used when training the model.

[0068] The original YOLOv8s model may have overfitting, which is too confident in the prediction sometimes, which may cause the prediction to deviate from the true situation. After using Label Smoothing, the problem can be alleviated to a certain extent, which is equivalent to a certain degree of data enhancement, which enhances the generalization ability of the model. Using this strategy can improve mAP@.5, which proves the feasibility of this training strategy.

[0069] 2.3 Counting module

[0070] Each frame of pre-processed image is recognized through the pre-trained YOLOv8 network to obtain the number of target fish in each area and determine the real-time proportion of target fish in different areas.

[0071] (1) Local Optimization Submodule

[0072] The fish pond is divided into m areas. For each area, the video stream information of the first three minutes of feeding is extracted for inference detection. Each frame of the image is identified and counted based on the three types of fish: head, tail, and whole fish. The maximum value of the three is taken as the number of fish in the frame of video x. a (a=1,2,3,…,t), where t is the total number of frames in the three-minute video stream.

[0073] (2) Average submodule

[0074] The average number of fish in the video stream and the total number of frames t of the video stream are averaged to obtain the average number of fish in each local area n i (i=1,2,3,…,m).

[0075] (3) Ratio acquisition submodule

[0076] Based on the average number of fish in each local area, the ratio of the total number of fish in each local area to the total number of fry in the pond can be expressed as:

[0077]

[0078] 2.4 Feed dosage determination module

[0079] After determining the sum of local area fry, only need to know the linear relation of feed weight and fish, just can know how much feed will be put in this area.In addition need consider the size and month of fish, the time of putting in fry in the selected place of present embodiment is about annual March.

[0080] The feed usage data of each region obtained by the feed usage determination module is expressed as:

[0081] Q i =kapR i (i=1,2,3,…,m);

[0082] Where: Q i represents the feed amount in the i-th area; k represents the feed coefficient, that is, the feed amount required to produce unit mass of fish meat; a represents the feed ratio, p represents the planned fish production in the fish pond, and R i Represents the real-time proportion of target fish in the i-th area.

[0083] In addition, the feeding amount of fish feed accounts for the proportion of the annual feeding amount (refer to Figure 3 ) and the feeding frequency also changes with the month: feeding 1% of the annual feeding amount in March, feeding 4% of the annual feeding amount in April, feeding 8% of the annual feeding amount in May, feeding 15% of the annual feeding amount in June, feeding 20% of the annual feeding amount in July, August and September, feeding 9% of the annual feeding amount in October, and feeding 3% of the annual feeding amount in November; and feeding 4 times a day from March to June, feeding 3 times a day from July to September, and feeding 2 times a day after October.

[0084] 3. Intelligent feeding device

[0085] After the edge server identifies the collected video data, it generates a corresponding feeding scheme according to the estimated fish population and state, and sends it to the water terminal and the device of the aquaculture administrator.

[0086] The intelligent feeding device comprises a host computer and a feeding machine. The host computer is used to issue a feeding instruction to the feeding machine at a set feeding time, and controls the feeding amount and feeding position by controlling the working time and direction of the feeding machine. The feeding machine completes the feeding operation based on the feeding instruction. At the same time, this embodiment supports the functions of setting the feeding scheme by the aquaculture personnel and adding temporary feeding tasks.

[0087] In addition, this embodiment mainly uses the Socket programming interface, the TCP transport layer protocol and the RTSP application layer protocol.

[0088] This embodiment mainly needs to transmit the underwater recorded video files. Data compression is needed to reduce the data volume and improve the transmission efficiency. At the same time, factors such as image quality, transmission speed and compression ratio need to be considered when selecting the compression format. This embodiment adopts the TCP transport layer protocol. Since the TCP protocol has a limit on the data size, the compressed image data needs to be packaged. Generally, the image data is divided into fixed-size data packets for transmission. The TCP protocol is a connection-oriented protocol, and a connection needs to be established before data transmission. After the connection is established, the data packet is sent to the receiver and waits for the confirmation information of the receiver. During data transmission, the loss and retransmission of data packets need to be considered to ensure the reliability of data transmission. Solutions are as follows:

[0089] For image data transmission, this embodiment adopts the method of compressing the video data collected by the data acquisition terminal (including an underwater camera and a wireless network access device), pushing the stream to the water transfer device through RTSP, and then forwarding the video stream to the edge server by the water transfer device. Then, the corresponding deep learning model performs real-time identification processing on the video in the database, and feeds back the final result to the administrator device.

[0090] For the image database, after receiving the video data, it will be classified and archived according to the IP address sent (corresponding to the number of data collection device) to facilitate subsequent identification processing and data display.

[0091] At the same time, in order to facilitate the use of users, visual monitoring and feeding can be carried out through the design of related system webpage, so that users can master the distribution of fish in the breeding water area and the feeding situation in real time. In order to avoid the case that the knowledge level of users is different, the webpage can be added with humanized use instruction, so as to facilitate the operation of the system, and the after-sales service can be guaranteed by adding relevant follow-up service and contact way in the webpage.

[0092] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other.

[0093] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fishery precision feeding system, characterized in that: include: The underwater identification device, the edge server, and the intelligent feeding device are connected in sequence; An underwater recognition device is used to obtain video stream information containing fish movements and send it to the edge server via the local area network; The edge server is used to identify and count target fish in the video stream information using a pre-trained YOLOv8 network, determine the real-time proportion of target fish in different areas, obtain feed usage data for different areas, and send it to the corresponding intelligent feeding device; Intelligent feeding device, which completes feeding based on the received feed dosage data; Among them, the edge server includes: image acquisition module, model improvement module, counting module, and feed dosage determination module; An image acquisition module is used to select target fish images with one or more complete outlines from the video stream information, clean the target fish images, and complete image annotation with the whole fish, fish head, and fish tail to obtain a preprocessed image; The model improvement module is used to improve the selected YOLOv8 basic model to obtain a pre-trained YOLOv8 network; The counting module uses a pre-trained YOLOv8 network to identify each frame of pre-processed images, obtain the number of target fish in each area, and determine the real-time ratio of target fish in different areas; The feed dosage determination module obtains feed dosage data for each area based on the feed dosage required for the growth of target fish at different periods and a mathematical model; The model improvement modules include: attention mechanism introduction submodule, loss function adjustment submodule, and noise reduction submodule; The attention mechanism introduces a submodule to add a SimAM attention module between the connection layer and the convolution layer in the Neck part of the YOLOv8 basic model to unify the weights; The loss function adjustment submodule is used to reduce the loss weight of the small object detection layer in the YOLOv8 basic model and increase the loss weight of the medium and large object detection layers; The anti-noise submodule obtains a pre-trained YOLOv8 network by using the Label Smoothing strategy during model training.

2. A fishery precision feeding system according to claim 1, characterized in that: The underwater identification device includes an above-water part for power supply and data transmission, and an underwater part for obtaining video stream information; The above-water part includes: an above-water camera for monitoring the water surface conditions, an electric power supply module for powering the above-water camera, and a communication module for temporarily storing and wirelessly transmitting the data collected by the above-water camera; The underwater part includes: an underwater camera for acquiring underwater image information and an anchoring device for stabilizing the underwater camera.

3. A fishery precision feeding system according to claim 2, characterized in that: The power supply module includes a solar panel, a battery and a step-down power supply module which are connected in sequence, and the output end of the step-down power supply module is connected to the water camera.

4. A fishery precision feeding system according to claim 2, characterized in that: The communication module is the Raspberry Pi picow development board.

5. The fishery precision feeding system according to claim 1, characterized in that: The counting module includes: a local optimization submodule, an average submodule, and a ratio acquisition submodule; The local optimization submodule identifies and counts the fish head, fish tail, and whole fish in each frame, and takes the maximum value of the three as the number of fish in the frame; The averaging submodule is used to average the number of fish in the video stream and the total number of frames in the video stream to obtain the average number of fish in each local area; The proportion acquisition submodule obtains the proportion of the total number of fish in each local area to the total number of fry in the fish pond based on the average number of fish in each local area.

6. A fishery precision feeding system according to claim 1, characterized in that: The feed usage data of each region obtained by the feed usage determination module is expressed as: Q i =kapR i (i=1,2,3,…,m); Where: Q i represents the feed amount in the i-th area; k represents the feed coefficient, that is, the feed amount required to produce unit mass of fish meat; a represents the feed ratio, p represents the planned fish production in the fish pond, and R i Represents the real-time proportion of target fish in the i-th area.

7. The fishery precision feeding system according to claim 1, characterized in that: The intelligent feeding device includes: host computer and feeding machine; The upper computer is used to send feeding instructions to the feeding machine at the set feeding time, and control the feeding quantity and feeding position by controlling the working time and direction of the feeding machine; The feeding machine completes the feeding operation based on the feeding instruction.

Images