An underwater disease identification method based on fish behavior supervision by laser radar

Abstract

The application discloses an underwater disease identification method based on laser radar fish behavior supervision. The method comprises the following steps: building a ranging system on the bottom of the water area where the fish are bred; using the ranging system to obtain the posture distance through three-dimensional scanning detection; obtaining a three-dimensional point cloud image according to the posture distance; inputting the T-S fuzzy neural network model for training; obtaining the three-dimensional point cloud image of the fish to be identified, inputting the trained model, and performing model matching; if the model cannot be matched, the fish is identified as a diseased fish, so that the underwater fish disease is identified. The improved PCA algorithm is used to obtain more feature points in the plane area, improve the identification accuracy near the fish tail, fish belly and the like, and through the model setting confidence interval and time for identification, the disease identification accuracy is greatly improved, and support is provided for aquatic disease prevention.

Description

Technical Field

[0001] This invention relates to a posture recognition method, specifically to an underwater disease identification method based on lidar for monitoring fish behavior. Background Technology

[0002] During the aquaculture process, fish are prone to diseases such as gill rot, red skin, and enteritis due to the proliferation of parasites and bacteria.

[0003] Currently, most fish disease prevention operations rely on manual visual observation. Disease identification depends on the personal experience of the staff, which can easily lead to missed or misdiagnosed diseased fish. Failure to regularly remove diseased fish can cause the spread of infectious diseases and a decline in water quality, resulting in mass fish deaths and reduced economic benefits.

[0004] Common fish diseases include hemorrhagic septicemia, bacterial enteritis, and bacterial gill rot. These diseases can affect fish characteristics and underwater behavior to some extent. For example, infected fish may exhibit behaviors such as "frog swimming," continuous "hiccups," and weak swimming. Farmers may not be able to detect these symptoms in time, thus missing the best opportunity for disease treatment. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides an underwater disease identification method based on lidar behavior monitoring. Abnormal fish behavior is identified through image processing, and combined with manual judgment, diseased fish are removed.

[0006] The technical solution adopted in this invention is:

[0007] The present invention provides a method for underwater fish disease identification based on lidar-based monitoring of fish behavior, comprising:

[0008] 1) Construct a distance measurement system on the bottom of the water body where a predetermined type of fish is raised.

[0009] 2) Use a ranging system to perform three-dimensional scanning detection on disease-free fish in the water. For each disease-free fish, scan to obtain the attitude distance between each part of the disease-free fish and the ranging system.

[0010] 3) For each disease-free fish, obtain a three-dimensional point cloud map of the disease-free fish based on the distances of each posture.

[0011] 4) Input the 3D point cloud images of each disease-free fish into the TS fuzzy neural network model for training to obtain the trained TS fuzzy neural network model.

[0012] 5) Perform the same operations as steps 2)-3) on the fish to be identified to obtain a three-dimensional point cloud map of the fish to be identified. Input the three-dimensional point cloud map of the fish to be identified into the trained TS fuzzy neural network model. The trained TS fuzzy neural network model performs model matching. Fish that cannot be matched with the model are identified as diseased fish, thereby realizing the identification of underwater fish diseases.

[0013] In step 1), the ranging system includes a semi-solid-state microelectronic system lidar MEMS and a mechanical rotating lidar. The semi-solid-state microelectronic system lidar MEMS is installed at the bottom center of the water area and forms a cone-shaped scanning area facing upward in the vertical direction; the mechanical rotating lidar is installed on the side of the bottom of the water area and forms a rectangular scanning area on the side in the horizontal direction.

[0014] The semi-solid-state microelectronic system lidar MEMS uses a vertical-cavity surface-emitting laser (VSEL) as its point light source; the mechanical rotating lidar uses an edge-emitting laser (EEL) as its point light source.

[0015] In step 2), the attitude distance between each part of the disease-free fish and the ranging system is obtained by scanning using a semi-solid-state microelectronic system lidar MEMS and a mechanical rotating lidar to obtain the attitude distance A between each part and the mouth, eyes, gills, dorsal fin, belly and tail of the disease-free fish.

[0016] In step 3), for each disease-free fish, the semi-solid-state microelectronic system lidar MEMS and the mechanical rotating lidar obtain the various pose distances of the disease-free fish after scanning and form two pose distance sets respectively. For each pose distance set, the initial three-dimensional point cloud map of the disease-free fish is obtained by first processing it with Principal Component Analysis (PCA). At the same time, the neighborhood centroid and its neighborhood points K are obtained. A new value K' of the neighborhood points is introduced in the PCA. When the current neighborhood points K of the neighborhood centroid is less than the new value K', the current neighborhood points K of the neighborhood centroid is expanded to the new value K'. Then, the updated PCA is used again to process each pose distance set to obtain the final three-dimensional point cloud map Q of the disease-free fish.

[0017] By combining the PCA algorithm with point cloud normal vectors and curvature information, multiple distances A are transformed into a 3D point cloud map Q. In the PCA method, curvature is calculated using the following formula:

[0018]

[0019]

[0020] Where T is the calculation point p i The covariance is a K-order matrix, where K is the number of neighborhood points, and γ0, γ1, and γ2 are the eigenvalues ​​(γ0 < γ1 < γ2). For the neighborhood centroid, ω i For curvature.

[0021] However, during the transformation process, due to the irregular distribution and uneven density of fish feature points (i.e., the surfaces of the fish belly and gills are relatively smooth), there may be a shortage of feature points during the collection process, resulting in missing planar features and inaccurate calculation results for normal vectors and curvature information. Therefore, a new K' point is introduced to increase the number of feature points selected when fewer are obtained.

[0022] The new value K` of the number of neighboring points is as follows:

[0023]

[0024] Among them, K i is the number of neighborhood points of the i-th neighborhood centroid in the initial 3D point cloud map, and n is the total number of points in the point cloud.

[0025] In step 4), the TS fuzzy neural network model is specifically trained using the gradient descent method.

[0026] The TS fuzzy neural network model consists of an antecedent network and an consequent network. The antecedent network is composed of an input layer, a fuzzification layer, a rule layer, and a defuzzification layer. The output of the antecedent network is used as the connection weight of the intermediate output layer of the consequent network. The consequent network is composed of an input layer, an intermediate layer, and an output layer. The model is trained by gradient descent to establish a normal grass carp posture prediction model based on the TS fuzzy neural network.

[0027] In practice, the three-dimensional point cloud images of each disease-free fish are randomly divided into training and validation sets at a ratio of 4:1 for training.

[0028] In step 5), the trained TS fuzzy neural network model is used for model matching. Specifically, the matching time of the TS fuzzy neural network model is recorded in real time, and the confidence value output by the TS fuzzy neural network model is obtained in real time during the processing. If the confidence value obtained within a preset time period is not within the preset confidence interval, it is considered that the model cannot be matched.

[0029] In practice, if the process takes more than 20 minutes, the fish will be determined to be non-compliant with the model. The corresponding fish to be identified will be marked as diseased fish, and a warning will be issued to personnel via the network and the fish will be removed from the list.

[0030] The beneficial effects of this invention are:

[0031] This invention innovatively proposes an improved PCA algorithm for underwater posture monitoring and identification of abnormal behavior in grass carp through real-time detection by lidar. This improves the accuracy of feature point selection and reduces the curvature of the gill and belly areas, accurately identifying fish with abnormal behavior. This enables disease prevention and control during the aquaculture process, effectively reducing the incidence of fish diseases and lowering labor costs, thereby improving economic benefits and preventing the decline of aquaculture water quality.

[0032] The method of this invention utilizes an improved PCA algorithm to obtain more feature points in a planar area, thereby improving the accuracy of identifying areas such as gills and belly. Furthermore, by setting confidence intervals and time for identification through a model, the accuracy of disease identification is greatly improved, providing support for disease prevention in aquatic products. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the installation of the ranging system of the present invention;

[0034] Figure 2 This is a schematic diagram of fish identification sites according to an embodiment of the present invention, wherein, Figure 2 (a) represents the fish feature point identification site. Figure 2 (b) represents the fish feature point identification region;

[0035] Figure 3 This is a point cloud image for grass carp identification according to an embodiment of the present invention, wherein, Figure 3 (a) is the point cloud map of the ordinary PCA algorithm. Figure 3 (b) is the point cloud diagram after introducing the K' value in this invention;

[0036] Figure 4 This is a flowchart of the method of the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 4 As shown, the underwater fish disease identification method based on lidar behavior monitoring of the present invention includes:

[0039] 1) Construct a ranging system on the bottom of the waters where a predetermined species of fish are raised, such as... Figure 1 As shown.

[0040] In step 1), the ranging system includes a semi-solid-state microelectronic system lidar MEMS and a mechanical rotating lidar. The semi-solid-state microelectronic system lidar MEMS is installed at the bottom center of the water area and forms a cone-shaped scanning area facing upward in the vertical direction; the mechanical rotating lidar is installed on the side of the bottom of the water area and forms a rectangular scanning area on the side in the horizontal direction.

[0041] The point source of semi-solid-state microelectronic system lidar MEMS is a vertical-cavity surface-emitting laser (VSEL); the point source of mechanical rotating lidar is an edge-emitting laser (EEL).

[0042] 2) Use a ranging system to perform three-dimensional scanning detection on disease-free fish in the water. For each disease-free fish, scan to obtain the attitude distance between each part of the disease-free fish and the ranging system.

[0043] In step 2), the attitude distances between various parts of the disease-free fish and the ranging system are obtained by scanning. Specifically, a semi-solid-state microelectronic system lidar (MEMS) and a mechanical rotating lidar are used to scan and obtain the attitude distances A between each part and the mouth, eyes, gills, dorsal fin, belly, and tail of the disease-free fish. For example... Figure 2 (a) and Figure 2 As shown in (b), this is the fish identification site of the embodiment.

[0044] 3) For each disease-free fish, obtain a three-dimensional point cloud map of the disease-free fish based on the distances of each posture.

[0045] In step 3), for each disease-free fish, the semi-solid-state microelectronic system lidar MEMS and the mechanical rotating lidar obtain the various pose distances of the disease-free fish after scanning and form two pose distance sets respectively. For each pose distance set, the initial three-dimensional point cloud map of the disease-free fish is obtained by first processing it with Principal Component Analysis (PCA). At the same time, the neighborhood centroid and its neighborhood points K are obtained. A new value K' of the neighborhood points is introduced in the PCA. When the current neighborhood points K of the neighborhood centroid is less than the new value K', the current neighborhood points K of the neighborhood centroid is expanded to the new value K'. Then, the updated PCA is used again to process each pose distance set to obtain the final three-dimensional point cloud map Q of the disease-free fish.

[0046] By combining the PCA algorithm with point cloud normal vectors and curvature information, multiple distances A are transformed into a 3D point cloud map Q. In the PCA method, curvature is calculated using the following formula:

[0047]

[0048]

[0049] Where T is the calculation point p i The covariance is a K-order matrix, where K is the number of neighborhood points, and γ0, γ1, and γ2 are the eigenvalues ​​(γ0 < γ1 < γ2). For the neighborhood centroid, ω i For curvature.

[0050] However, during the transformation process, due to the irregular distribution and uneven density of fish feature points (i.e., the surfaces of the fish belly and gills are relatively smooth), there may be a shortage of feature points during the collection process, resulting in missing planar features and inaccurate calculation results for normal vectors and curvature information. Therefore, a new K' point is introduced to increase the number of feature points selected when fewer are obtained.

[0051] The new value K` of the neighborhood points is as follows:

[0052]

[0053] Among them, K i is the number of neighborhood points of the i-th neighborhood centroid in the initial 3D point cloud map, and n is the total number of points in the point cloud.

[0054] 4) Input the 3D point cloud images of each disease-free fish into the TS fuzzy neural network model for training to obtain the trained TS fuzzy neural network model.

[0055] In step 4), the TS fuzzy neural network model is specifically trained using the gradient descent method.

[0056] The TS fuzzy neural network model consists of an antecedent network and an consequent network. The antecedent network is composed of an input layer, a fuzzification layer, a rule layer, and a defuzzification layer. The output of the antecedent network is used as the connection weight of the intermediate output layer of the consequent network. The consequent network is composed of an input layer, an intermediate layer, and an output layer. The model is trained by gradient descent to establish a normal grass carp posture prediction model based on the TS fuzzy neural network.

[0057] In practice, the three-dimensional point cloud images of each disease-free fish are randomly divided into training and validation sets at a ratio of 4:1 for training.

[0058] 5) Perform the same operations as steps 2)-3) on the fish to be identified to obtain a three-dimensional point cloud map of the fish to be identified. Input the three-dimensional point cloud map of the fish to be identified into the trained TS fuzzy neural network model. The trained TS fuzzy neural network model performs model matching. Fish that cannot be matched with the model are identified as diseased fish, thereby realizing the identification of underwater fish diseases.

[0059] In step 5), the trained TS fuzzy neural network model is used for model matching. Specifically, the matching time of the TS fuzzy neural network model is recorded in real time, and the confidence value output by the TS fuzzy neural network model is obtained in real time during the processing. If the confidence value obtained within a preset time period is not within the preset confidence interval, it is considered that the model cannot be matched.

[0060] In practice, if the process takes more than 20 minutes, the fish will be determined to be non-compliant with the model. The corresponding fish to be identified will be marked as diseased fish, and a warning will be issued to personnel via the network and the fish will be removed from the list.

[0061] like Figure 3 (a) and Figure 3 As shown in (b) and Tables 1 and 2, the improved PCA algorithm was used to expand the search for nearby feature points in the relatively smooth gill and belly regions of the same target fish. The number of feature points K in the gill region increased from 35-55 to 60-75 after correction, with an average improvement rate of 45.28%; the number of feature points K in the belly region increased from 50-61 to 66-81 after correction, with an average improvement rate of 32.30%. The increased number of feature points reduced the error caused by insufficient feature point selection and improved the accuracy of the PCA algorithm in calculating curvature.

[0062] Table 1

[0063]

[0064] Table 2

[0065]

Claims

1. A method for identifying underwater fish diseases based on lidar-based monitoring of fish behavior, characterized in that, include: 1) Construct a ranging system on the bottom of the waters where a predetermined species of fish are raised; 2) Use a ranging system to perform three-dimensional scanning detection on disease-free fish in the water. For each disease-free fish, scan to obtain the attitude distance between each part of the disease-free fish and the ranging system. 3) For each disease-free fish, obtain a three-dimensional point cloud map of the disease-free fish based on the distances of each posture. 4) Input the 3D point cloud images of each disease-free fish into the TS fuzzy neural network model for training to obtain the trained TS fuzzy neural network model. 5) Perform the same operations as steps 2)-3) on the fish to be identified to obtain a three-dimensional point cloud map of the fish to be identified. Input the three-dimensional point cloud map of the fish to be identified into the trained TS fuzzy neural network model. The trained TS fuzzy neural network model performs model matching. Fish that cannot be matched with the model are identified as diseased fish, thereby realizing the identification of underwater fish diseases. In step 1), the ranging system includes a semi-solid-state microelectronic system lidar MEMS and a mechanical rotating lidar. The semi-solid-state microelectronic system lidar MEMS is installed at the bottom center of the water area and forms a cone-shaped scanning area facing upward in the vertical direction; the mechanical rotating lidar is installed on the side of the bottom of the water area and forms a rectangular scanning area on the side in the horizontal direction. In step 3), for each disease-free fish, the semi-solid-state microelectronic system lidar MEMS and the mechanical rotating lidar obtain the various posture distances of the disease-free fish after scanning and form two posture distance sets respectively. For each posture distance set, the initial three-dimensional point cloud map of the disease-free fish is obtained by first processing it with principal component analysis (PCA), and the neighborhood centroid and its neighborhood points K are obtained at the same time. A new value K' of the neighborhood points is introduced in the principal component analysis (PCA). When the current neighborhood points K of the neighborhood centroid is less than the new value K' of the neighborhood points, the current neighborhood points K of the neighborhood centroid is expanded to the new value K' of the neighborhood points. Then, the updated principal component analysis (PCA) is used again to process each posture distance set to obtain the final three-dimensional point cloud map Q of the disease-free fish.

2. The underwater fish disease identification method based on lidar behavior monitoring according to claim 1, characterized in that: The semi-solid-state microelectronic system lidar MEMS uses a vertical cavity surface-emitting laser (VSEL) as its point source; the mechanical rotating lidar uses an edge-emitting laser (EEL) as its point source.

3. The underwater fish disease identification method based on lidar behavior monitoring according to claim 1, characterized in that: In step 2), the attitude distance between each part of the disease-free fish and the ranging system is obtained by scanning using a semi-solid-state microelectronic system lidar MEMS and a mechanical rotating lidar to obtain the attitude distance A between each part and the mouth, eyes, gills, dorsal fin, belly and tail of the disease-free fish.

4. The underwater fish disease identification method based on lidar behavior monitoring according to claim 1, characterized in that: The new value K` of the number of neighborhood points is as follows: Among them, K i is the number of neighborhood points of the i-th neighborhood centroid in the initial 3D point cloud map, and n is the total number of points in the point cloud.

5. The underwater fish disease identification method based on lidar behavior monitoring according to claim 1, characterized in that: In step 4), the TS fuzzy neural network model is specifically trained using the gradient descent method.

6. The underwater fish disease identification method based on lidar behavior monitoring according to claim 1, characterized in that: In step 5), the trained TS fuzzy neural network model is used for model matching. Specifically, the matching time of the TS fuzzy neural network model is recorded in real time, and the confidence value output by the TS fuzzy neural network model is obtained in real time during the processing. If the confidence value obtained within a preset time period is not within the preset confidence interval, it is considered that the model cannot be matched.

Images