An intelligent observation device and method for shrimp culture

Abstract

The application discloses a kind of intelligent observation device and observation method to shrimp culture, including support;Transparent feed tray, transparent feed tray can be placed in feed for shrimp to feed;Drive unit is fixed on support, drive unit is used to drive transparent feed tray to rise or drop;Shrimp image acquisition module is used to collect the video in the process of shrimp feeding;Shrimp image processing module is connected with shrimp image acquisition module, and shrimp image processing module can identify and analyze the feeding intensity, active state, hepatopancreas of shrimp according to the video in the process of shrimp feeding;Hydrophone module is arranged directly below transparent feed tray;Stress sensor module is installed in the middle position inside transparent feed tray.The application realizes remote, intelligent observation of the growth condition of shrimp, compares the collected data with the sample data of normal growth of shrimp, makes corresponding adjustment measures, reduces the consumption of manpower, material resources and financial resources.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, and in particular to an intelligent observation device and method for shrimp farming. Background Technology

[0002] With the rapid development of the global economy and the improvement of people's living standards, the consumption of aquaculture products has been growing rapidly over the past two decades. To meet people's demand for aquaculture products, we need to closely monitor the growth status of aquatic organisms. Due to climate change, the global aquaculture industry has suffered huge losses.

[0003] To ensure the healthy growth of shrimp and a bountiful harvest for fishermen, we urgently need an intelligent monitoring method and device for shrimp farming.

[0004] Traditional methods often rely on manual labor or the experience of shrimp farmers to harvest shrimp from ponds or net cages and visually assess their body length. This traditional method is time-consuming and labor-intensive, depending on the farmer's experience and the ability to visually inspect the shrimp for color purity and hepatopancreatic lesions. In this process, a transparent feeding tray is submerged near the water surface. Feed is placed on the tray for a period of time. The number of shrimp on the tray is then recorded and measured, along with their body length and the amount of uneaten feed, to assess their growth and appetite. While this method can be reliable and frequently used, it is time-consuming and subjective. Most importantly, shrimp are benthic animals, and this process can disrupt their feeding process, potentially failing to reflect their actual behavior. Therefore, a suitable and automated method is needed to estimate shrimp body length. Furthermore, manual harvesting can negatively impact shrimp growth; frequent harvesting can cause excessive stress and even death, resulting in losses for shrimp farmers.

[0005] Currently, with the development of artificial intelligence, we can use machine vision image processing technology to process shrimp feeding videos captured by cameras frame by frame to obtain data such as shrimp body length, hepatopancreas, and uneaten feed. This allows for real-time monitoring of shrimp growth. By comparing the collected data with data from normally growing shrimp samples, appropriate adjustments can be made to reduce losses. Therefore, an intelligent monitoring device and method for shrimp farming are proposed. Summary of the Invention

[0006] Therefore, it is necessary to provide an intelligent observation device and method for shrimp farming to address the aforementioned technical problems, so as to realize remote and intelligent observation of shrimp growth, obtain shrimp body length, quantity and uneaten feed amount, improve feed utilization, bring great convenience to farmers and reduce costs.

[0007] An intelligent monitoring device for shrimp farming includes:

[0008] support;

[0009] A transparent feeding tray, wherein feed for shrimp to consume is placed inside the transparent feeding tray;

[0010] A drive unit, fixed on the bracket, is used to drive the transparent feed tray to rise or fall.

[0011] A shrimp image acquisition module is located above the transparent feed tray and is used to acquire video of the shrimp feeding process.

[0012] The shrimp image processing module is connected to the shrimp image acquisition module. The shrimp image processing module can identify and analyze the shrimp's feeding intensity, activity state, and hepatopancreas based on the video of the shrimp feeding process.

[0013] A hydrophone module is located directly below the transparent feed tray, and the hydrophone module is used to determine the feeding intensity of the shrimp.

[0014] A stress sensor module is installed in the middle of the inside of the transparent feed tray. The stress sensor module can sense the activity state of the shrimp.

[0015] In one embodiment, the feed tray includes an upper wall and a lower bottom plate, both of which are made of PMMA material.

[0016] In one embodiment, the support includes:

[0017] A vertical rod, with a horizontal bar provided on one side;

[0018] A triangular support is provided between the vertical bar and the horizontal bar.

[0019] In one embodiment, the drive unit includes:

[0020] The motor is connected to the solar panel;

[0021] The first U-shaped fixing bracket and the second U-shaped fixing bracket are fixed to the upper surface of the crossbar, and the first U-shaped fixing bracket and the second U-shaped fixing bracket are respectively equipped with a first pulley and a second pulley;

[0022] A rope, one end of which is wound around the motor, and the other end of which passes through the first pulley and the second pulley in sequence and is connected to the transparent feed tray.

[0023] An observation method for an intelligent monitoring device for shrimp farming includes the following steps:

[0024] S1. Collect videos of shrimp feeding processes;

[0025] S2. Extract images frame by frame from the acquired video, and use the improved version of the Mosaic data augmentation algorithm to obtain the original dataset;

[0026] S3. Use Labelme to divide the original dataset into training set, test set and validation set;

[0027] S4. In the target recognition stage, the training set is input into the feature extraction network to obtain the feature map;

[0028] S5. Input the feature maps into the region proposal network and the region of interest pooling, respectively, to generate candidate detection boxes and proposed feature maps.

[0029] S6. Use Faster R-CNN to identify, classify, and regress candidate detection boxes and proposed feature maps;

[0030] S7. Obtain the body length, number, and amount of uneaten feed of the shrimp.

[0031] In one embodiment, step S2 includes:

[0032] S21. Select 5 original images captured frame by frame, and randomly pick 4 images from the 5 original images to flip, scale, and change the color gamut.

[0033] S22. Arrange the original images in the following order: the first image is placed in the upper left, the second image in the lower left, the third image in the lower right, and the fourth image in the upper right.

[0034] S23. Finally, the combined image is merged with the remaining image to obtain a rich dataset as the original dataset.

[0035] In one embodiment, step S3 includes: labeling the original dataset using Labelme and dividing it into a training set, a test set, and a validation set in an 8:1:1 ratio.

[0036] In one embodiment, step S5, inputting the feature map into the region proposal network, includes:

[0037] The input region suggestion network outputs all candidate boxes that may contain the target.

[0038] Candidate boxes are extracted using a sliding window method;

[0039] A classifier is used to determine whether a candidate box is a target or background. A regressor is used to further refine the candidate box location. The target location is then completed, and the candidate boxes that contain the target and have been adjusted for location are obtained. At the same time, candidate boxes that are too small or exceed the boundary are removed.

[0040] Finally, nonmaximum suppression is used to determine candidate detection boxes containing the target.

[0041] In one embodiment, step S5, inputting the feature map into the region of interest pooling includes: collecting the coordinates of each candidate box generated by the region proposal network and annotating them on the initial feature map to generate a proposed feature map.

[0042] In one embodiment, step S7 involves obtaining the shrimp's body length data using the following method:

[0043] The video footage of shrimp feeding is input frame by frame into Faster R-CNN; Faster R-CNN detects body parts of the moving shrimp.

[0044] When a certain number of body parts of a shrimp are detected, a bounding box will be drawn to cover the entire shrimp body, and the individual node corresponding to the center of the drawn bounding box will also be assigned to each correctly detected body part.

[0045] When sufficient detection is achieved, i.e. when two adjacent parts are detected, an effective directed cycle graph (DCG) can be obtained using the shrimp's body characteristics.

[0046] The shrimp's posture is defined by key parts of the shrimp's body, resulting in a posture matrix containing a series of postures. Then, the obtained shrimp posture matrix is ​​encoded.

[0047] The shrimp's posture is determined by comparison, and then the shrimp's body length is estimated based on the size of the bounding box.

[0048] The aforementioned intelligent shrimp farming observation device and method combine traditional device modules with machine vision image processing technology to achieve remote and intelligent observation of shrimp growth. By comparing the collected data with sample data of shrimp growing normally, corresponding adjustment measures can be taken to reduce the consumption of manpower, material resources, and financial resources. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the intelligent shrimp farming observation device of the present invention;

[0051] Figure 2 This is a schematic diagram of the intelligent observation method for shrimp farming according to the present invention. Detailed Implementation

[0052] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0053] It should be noted that when a component is said to be "fixed to" another component, it can be directly attached to the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0055] See Figure 1 As shown, one embodiment of the present invention provides an intelligent monitoring device for shrimp farming, which includes:

[0056] support;

[0057] A transparent feeding tray 16, wherein feed for shrimp to consume is placed inside the transparent feeding tray 16;

[0058] A drive unit is fixed on the bracket, and the drive unit is used to drive the transparent feed tray 16 to rise or fall.

[0059] A shrimp image acquisition module 10 is located above the transparent feed tray 16 and is used to capture video of the shrimp feeding process. In this embodiment, the shrimp image acquisition module 10 may include a camera, which is vertically suspended above the transparent feed tray 16 to capture images of the shrimp and the amount of uneaten feed during the feeding process. Optionally, to obtain clearer images, a light source 14 can be added to the transparent feed tray 16.

[0060] The shrimp image processing module 17 is connected to the shrimp image acquisition module 10. The shrimp image processing module 17 can identify and analyze the shrimp's feeding intensity, activity state, and hepatopancreas based on the video of the shrimp feeding process.

[0061] A hydrophone module 15 is located directly below the transparent feed tray 16. The hydrophone module 15 is used to determine the feeding intensity of the shrimp.

[0062] The stress sensor module 13 is installed in the middle of the inside of the transparent feed tray 16. The stress sensor module 13 can sense the activity state of the shrimp.

[0063] In this invention, the shrimp image processing module 17 includes a shrimp activity state module, a shrimp feeding intensity module, and a shrimp hepatopancreas module. The shrimp activity state module integrates features obtained from the stress sensor module 13 and the support to further determine the shrimp's activity state. The shrimp feeding intensity module integrates features obtained from the shrimp image acquisition module 10 and the hydrophone module 15 to further determine the shrimp's feeding intensity. Furthermore, under normal conditions, the shrimp's hepatopancreas is brownish-red; when the shrimp exhibits abnormalities, the hepatopancreas may appear red, yellow, or black. The shrimp hepatopancreas module processes the acquired shrimp images to identify the color of the hepatopancreas and uses these color changes to determine whether the shrimp's growth is abnormal.

[0064] It should be noted that the stress sensor module 13 is used to detect data on stress changes inside the transparent feed tray 16. By monitoring the deformation of the bottom of the transparent feed tray 16 caused by the twisting and crowding of shrimp during feeding, the activity state of the shrimp during feeding is assessed. The hydrophone module 15 collects a large amount of sound frequency and intensity data emitted by shrimp at the beginning of feeding and when they are close to being full. It analyzes the change patterns of sound frequency and intensity during the shrimp's feeding process, and compares the real-time collected sound frequency and intensity data with the analyzed patterns to determine the feeding intensity of the shrimp.

[0065] In one embodiment of the present invention, the feed tray includes an upper wall 11 and a lower bottom plate 12, both of which are made of PMMA material. PMMA material is not only transparent but also has good light transmittance, thus enabling more accurate and effective acquisition of images of shrimp and uneaten feed.

[0066] In one embodiment of the present invention, the support includes:

[0067] A vertical rod 3 is provided with a horizontal rod 9 on one side; the vertical rod 3 and the horizontal rod 9 form a "7" shaped structure.

[0068] A triangular support 4 connects the vertical rod 3 and the horizontal rod 9. The triangular support frame 4 can reinforce and support the bracket.

[0069] In one embodiment of the present invention, the driving unit includes:

[0070] Motor 2 is connected to solar panel 1; specifically, motor 2 is mounted on vertical pole 3, about 0.5m above the ground, and is driven by power generated by solar panel 1.

[0071] The first U-shaped fixing frame 6 and the second U-shaped fixing frame 8 are fixed to the upper surface of the crossbar 9. The first U-shaped fixing frame 6 and the second U-shaped fixing frame 8 are respectively equipped with the first pulley 5 and the second pulley 7. The first U-shaped fixing frame 6 and the second U-shaped fixing frame 8 are used to install the first pulley 5 and the second pulley 7.

[0072] A rope, one end of which is wound around the motor 2, and the other end of which passes through the first pulley 5 and the second pulley 7 in sequence and is connected to the transparent feed tray 16.

[0073] In this embodiment, the solar panel 1 drives the motor 2 under sunlight. The first pulley 5 and the second pulley 7 cause the rope to move along the trajectory of the transparent feed tray 16, thus raising and lowering the tray. During feeding, feed is placed on the bottom plate 12 of the transparent feed tray 16. Then, the transparent feed tray 16 is lowered to the bottom of the water. After approximately 20-60 minutes, the transparent feed tray 16 is raised. Throughout the feeding process, the shrimp image acquisition module 10 continuously captures images to obtain experimental data.

[0074] See Figure 2 As shown, an embodiment of the present invention provides an observation method for an intelligent observation device for shrimp farming, comprising the following steps:

[0075] S1. Collect video of shrimp feeding process; video can be collected using shrimp image acquisition module 10 (camera);

[0076] S2. Extract images frame by frame from the acquired video, and use the improved version of the Mosaic data augmentation algorithm to obtain the original dataset;

[0077] S3. Use Labelme to divide the original dataset into training set, test set and validation set;

[0078] S4. In the target recognition stage, the training set is input into the feature extraction network (DRN) to obtain the feature map;

[0079] S5. Input the feature maps into the Region Proposal Network (RPN) and Region of Interest Pooling (ROI Pooling) respectively to generate candidate detection boxes and proposed feature maps.

[0080] S6. Use Faster R-CNN to identify, classify, and regress candidate detection boxes and proposed feature maps;

[0081] S7. Obtain the body length, number, and amount of uneaten feed of the shrimp.

[0082] In this invention, for the shrimp image acquisition module 10 acquiring video of the shrimp feeding process, the shrimp image processing module 17 uses DRN and Faster R-CNN Litopenaeus vannamei detection models for recognition and analysis. Specifically, DRN is mainly used for feature extraction, while Faster R-CNN is mainly used for candidate detection boxes for recognition, classification, and bounding box regression. The advantage of using DRN is that using multiple convolutional layers to learn the residuals between input and output makes the network easier to train. In this case, the output data of each residual module consists of the input data and the residuals. Because there is a "direct connection" between the input and output in each module, gradient propagation is easier.

[0083] In one embodiment of the present invention, step S2 includes:

[0084] S21. Select 5 original images captured frame by frame, and randomly pick 4 images from the 5 original images to flip, scale, and change the color gamut.

[0085] S22. Arrange the original images in the following order: the first image is placed in the upper left, the second image in the lower left, the third image in the lower right, and the fourth image in the upper right.

[0086] S23. Finally, the combined image is merged with the remaining image to obtain a rich dataset as the original dataset.

[0087] In one embodiment of the present invention, step S3 includes: labeling the original dataset using Labelme and dividing it into a training set, a test set, and a validation set in a ratio of 8:1:1.

[0088] In one embodiment of the present invention, step S5, inputting the feature map into the Region Proposal Network (RPN), includes:

[0089] The input Region Proposal Network (RPN) outputs all candidate boxes that may contain the target;

[0090] Candidate boxes are extracted using a sliding window method;

[0091] A classifier is used to determine whether a candidate box is a target or background. A regressor is used to further refine the candidate box location. The target location is then completed, and the candidate boxes that contain the target and have been adjusted for location are obtained. At the same time, candidate boxes that are too small or exceed the boundary are removed.

[0092] Finally, nonmaximum suppression (NMS) is used to determine candidate detection boxes containing the target object.

[0093] In one embodiment of the present invention, step S5, inputting the feature map into the region of interest pooling, includes: collecting the coordinates of each candidate box generated by the Region Proposal Network (RPN), and annotating them on the initial feature map to generate a proposed feature map. Finally, it is fed into a subsequent Faster R-CNN (fully connected layer) for further classification and regression, ultimately detecting shrimp and uneaten bait.

[0094] In this invention, the swimming characteristics of shrimp are as follows: Although the shrimp's swimming posture is constantly changing, the relative positions of each part of the shrimp's body remain unchanged. For example, the head and tail are relative parts, as are the abdomen and abdominal legs.

[0095] In one embodiment of the present invention, step S7, wherein the method for obtaining the body length data of the shrimp includes:

[0096] The video footage of shrimp feeding is input frame by frame into Faster R-CNN; Faster R-CNN detects body parts of the moving shrimp.

[0097] When a certain number of body parts of a shrimp are detected, a bounding box will be drawn to cover the entire shrimp body, and the individual node corresponding to the center of the drawn bounding box will also be assigned to each correctly detected body part.

[0098] When sufficient detection is achieved, i.e. when two adjacent parts are detected, an effective directed cyclic graph (DCG) can be obtained using the shrimp's body characteristics. For example, detecting the head and abdomen, or the head and foot, etc. In this way, we can use dual invariance to determine the missing detections caused by the detector's defects. Only in this way can an effective DCG be obtained.

[0099] The shrimp's posture is defined by key body parts, resulting in a posture matrix containing a series of postures. This posture matrix is ​​then encoded. Here, the shrimp facing right towards the reader is defined as binary code "1," following the DCG diagram (head → ventral foot → tail → abdomen); the shrimp facing left towards the reader is defined as binary code "0," also following the DCG diagram (head → abdomen → tail → ventral foot). Sixteen template postures are then obtained at 45-degree intervals, clockwise and counterclockwise. The specific swimming direction is encoded in the range 1-8. The code for the detected shrimp posture is formed by combining the left (right) encoding of the shrimp facing the reader and the specific swimming direction encoding.

[0100] The shrimp's posture is determined by comparison, and then the shrimp's body length is estimated based on the size of the bounding box. Specifically, the shrimp's posture is determined by comparison with 16 existing templates, and then the body length of the shrimp is calculated by comparing the bounding box size with the original shrimp sample data.

[0101] In this invention, the advantage of DCG images lies in the fact that sometimes the detector may fail to detect the required parts due to shrimp overlap, obstruction by feed and excrement, or camera exposure issues. In such cases, we can rely on the dual invariance of shrimp physical characteristics to fill in the erroneously detected or undetected parts, ultimately constructing an effective DCG image. Secondly, with an effective DCG image, we can quickly determine the side of the shrimp facing the reader. This feature can reduce the computational cost of shrimp body length by half and also improve the speed of shrimp body length detection.

[0102] It should be noted that in this invention, the identification of the amount of feed remaining in the transparent feeding tray 16 is achieved by inputting frame-by-frame video of shrimp feeding. When uneaten feed is detected, the uneaten feed is covered with a bounding box, and the amount of uneaten feed is estimated by calculating the area of ​​the bounding box. Specifically, image enhancement techniques can also be used, such as histogram equalization to maximize image contrast; grayscale transformation to make the image clear and feature-clear; and image smoothing, mainly by using Gaussian filtering, median filtering, etc. to eliminate noise in the image.

[0103] In summary, this invention combines traditional device modules with machine vision image processing technology to obtain information such as shrimp body length, quantity, hepatopancreas, and uneaten feed from shrimp feeding videos. This enables intelligent, remote monitoring of shrimp growth, greatly benefiting aquaculture farmers. The amount of uneaten feed is used to determine the shrimp's feeding intensity and the next feeding amount, effectively improving feed utilization and reducing farming costs.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The embodiments described above merely illustrate several implementations of the present invention and should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. An intelligent monitoring device for shrimp farming, characterized in that, include: support; A transparent feeding tray (16) is provided, which can hold feed for shrimp to eat; the feeding tray includes an upper wall (11) and a lower bottom plate (12), both of which are made of PMMA material. A drive unit is fixed on the bracket, and the drive unit is used to drive the transparent feed tray (16) to rise or fall; The shrimp image acquisition module (10) is located above the transparent feed tray (16) and is used to acquire video of the shrimp feeding process; it also includes a light source (14) mounted on the transparent feed tray (16) to acquire images with higher clarity. The shrimp image processing module (17) is connected to the shrimp image acquisition module (10). The shrimp image processing module (17) can identify and analyze the shrimp's feeding intensity, activity status, and hepatopancreas based on the video of the shrimp feeding process. The hydrophone module (15) is located directly below the transparent feed tray (16). The hydrophone module (15) is used to determine the feeding intensity of the shrimp. The hydrophone module collects a large amount of sound frequency and intensity data emitted by the shrimp when they are just beginning to feed and when they are close to being full. It analyzes the change pattern of sound frequency and intensity during the feeding process of the shrimp and compares the real-time collected sound frequency and intensity data with the analyzed pattern to determine the feeding intensity of the shrimp. The stress sensor module (13) is installed in the middle of the inside of the transparent feed tray (16). The stress sensor module (13) can sense the activity state of the shrimp. The stress sensor module is used to detect the data of stress change inside the transparent feed tray. By monitoring the deformation of the bottom of the transparent feed tray caused by the twisting and crowding of the shrimp during feeding, the activity state of the shrimp during feeding can be judged. The shrimp image processing module includes a shrimp activity status module, a shrimp feeding intensity module, and a shrimp hepatopancreas module. The shrimp activity status module integrates features obtained from the stress sensor module and the support to further determine the shrimp's activity status. The shrimp feeding intensity module integrates features obtained from the shrimp image acquisition module and the hydrophone module to further determine the shrimp's feeding intensity. Furthermore, under normal conditions, the shrimp's hepatopancreas is brownish-red; when the shrimp exhibits abnormalities, the hepatopancreas may appear red, yellow, or black. The shrimp hepatopancreas module processes the acquired shrimp images to identify the color of the hepatopancreas and uses these color changes to determine whether the shrimp's growth is abnormal. The observation method of the intelligent shrimp farming observation device includes the following steps: S1. Collect videos of shrimp feeding processes; S11. Perform image enhancement processing on the acquired video images, including histogram equalization, grayscale transformation, Gaussian filtering or median filtering, to improve image contrast and eliminate image noise; S2. Extract images frame by frame from the acquired video, and use the improved version of the Mosaic data augmentation algorithm to obtain the original dataset; Step S2 includes: S21. Select 5 original images captured frame by frame, and randomly pick 4 images from the 5 original images to flip, scale, and change the color gamut; S22. Arrange the original images in the following order: the first image is placed in the top left, the second image in the bottom left, the third image in the bottom right, and the fourth image in the top right. S23. Finally, merge the combined image with the remaining image to obtain a rich dataset as the original dataset; S3. Use Labelme to divide the original dataset into training set, test set and validation set; S4. In the target recognition stage, the training set is input into the feature extraction network to obtain the feature map; S5. Input the feature maps into the region proposal network and the region of interest pooling, respectively, to generate candidate detection boxes and proposed feature maps; S6. Use Faster R-CNN to identify, classify, and regress candidate detection boxes and proposed feature maps; S7. Obtain the body length, number, and amount of uneaten feed of the shrimp; In step S7, the method for obtaining the shrimp's body length data includes: The video footage of shrimp feeding is input frame by frame into Faster R-CNN; Faster R-CNN detects body parts of the moving shrimp. When a certain number of body parts of a shrimp are detected, a bounding box will be drawn to cover the entire shrimp body, and the individual node corresponding to the center of the drawn bounding box will also be assigned to each correctly detected body part. When sufficient detection is achieved, i.e. when two adjacent parts are detected, an effective directed cyclic graph (DCG) can be obtained by utilizing the shrimp's body characteristics. For example, detecting the head and abdomen, or the head and abdominal legs, etc., can utilize dual invariance to determine the missing detections caused by the detector's defects. Only in this way can an effective DCG be obtained. The shrimp's posture is defined by the key parts of the shrimp's body, resulting in a posture matrix containing a series of postures. Then, the obtained posture matrix of the shrimp is encoded; here, the shrimp facing the reader on the right is defined as binary code "1". The right-side facing the reader follows the DCG diagram, that is, from the head through the abdominal foot, tail to the abdomen. The left-side orientation of the shrimp towards the reader is defined as binary code "0". The left-side orientation follows the DCG diagram, that is, from the head through the abdomen, tail to the foot, and then 16 template postures are obtained at 45-degree intervals clockwise and counterclockwise. The specific swimming direction is encoded in the range of 1-8. The code of the shrimp's posture is formed by combining the left / right code of the shrimp facing the reader and the specific swimming direction code. The posture of the shrimp is determined by comparison, and then the body length of the shrimp is estimated based on the size of the bounding box. Specifically, the posture of the shrimp is determined by comparison with 16 existing templates, and then the body length of the shrimp is calculated by comparing with the original shrimp sample data based on the size of the bounding box.

2. The intelligent monitoring device for shrimp farming as described in claim 1, characterized in that, The support includes: A vertical rod (3) is provided on one side of the vertical rod (3); A triangular support (4) is connected between the vertical rod (3) and the horizontal rod (9).

3. The intelligent monitoring device for shrimp farming as described in claim 2, characterized in that, The driving unit includes: The motor (2) is connected to the solar panel (1); The first U-shaped fixing bracket (6) and the second U-shaped fixing bracket (8) are fixed on the upper surface of the crossbar (9). The first U-shaped fixing bracket (6) and the second U-shaped fixing bracket (8) are respectively equipped with a first pulley (5) and a second pulley (7). A rope, one end of which is wound around the motor (2), and the other end of which passes through the first pulley (5) and the second pulley (7) in sequence and is connected to the transparent feed tray (16).

4. The intelligent monitoring device for shrimp farming as described in claim 1, characterized in that, Step S3 includes: labeling the original dataset using Labelme and dividing it into training set, test set and validation set in an 8:1:1 ratio.

5. The intelligent monitoring device for shrimp farming as described in claim 4, characterized in that, In step S5, inputting the feature map into the region proposal network includes: The input region suggestion network outputs all candidate boxes that may contain the target. Candidate boxes are extracted using a sliding window method; A classifier is used to determine whether a candidate box is a target or background. A regressor is used to further refine the candidate box location. The target location is then completed, and the candidate boxes that contain the target and have been adjusted for location are obtained. At the same time, candidate boxes that are too small or exceed the boundary are removed. Finally, nonmaximum suppression is used to determine candidate detection boxes containing the target.

6. The intelligent monitoring device for shrimp farming as described in claim 5, characterized in that, In step S5, inputting the feature map into the region of interest pooling includes: collecting the coordinates of each candidate box generated by the region proposal network and marking them on the initial feature map to generate a proposed feature map.

Images