Underwater multi-view stereoscopic vision accurate fish body segmentation system and segmentation method

Abstract

The invention discloses an underwater multi-eye stereoscopic vision accurate fish body segmentation system and segmentation method, and belongs to the field of underwater fish monitoring, the segmentation system comprises at least one underwater trinocular imaging module and a data processing unit; wherein the underwater trinocular imaging module is fixed in a culture water body of the culture tank body, and the underwater trinocular imaging module is used for collecting a multi-view underwater image in the culture tank body; and the data processing unit is used for identifying and segmenting the multi-view underwater image to obtain a segmentation result and space angle information of the single fish body. According to the invention, the precision and the stability of a fish body segmentation result are significantly improved by combining the multi-view imaging module with geometric layout design and depth estimation technologies.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater fish monitoring, and in particular relates to an underwater multi-eye stereoscopic vision accurate fish body segmentation system and segmentation method. Background Art

[0002] In modern aquaculture, accurate segmentation of fish targets is one of the key technologies for health assessment, behavior analysis, and intelligent aquaculture management. However, the underwater environment is complex, and fish targets are easily affected by background interference, light refraction, and water scattering, resulting in segmentation accuracy that is difficult to meet the needs of actual applications. In recent years, deep learning-based target detection and segmentation methods have gradually attracted attention in underwater scenes. They can extract fine features of fish and combine image processing technology to remove background interference. However, most existing solutions are based on monocular or binocular imaging, which can only provide two-dimensional segmentation results of fish targets, and it is difficult to further establish the precise position relationship of the fish in three-dimensional space. In addition, since fish swimming in water is often accompanied by posture changes, existing solutions lack effective means to obtain spatial posture information such as roll angle, resulting in obvious errors in segmentation results in complex aquaculture scenes.

[0003] At present, there have been some studies using binocular or multi-eye imaging systems, such as the invention patent "An underwater fish school observation device and method" with application number CN112465778A, which obtains three-dimensional information of fish bodies through binocular parallax calculation; the invention patent "Biological automatic measurement system based on stereoscopic vision and its measurement method" with application number CN112131921B, which realizes multi-view observation of underwater fish bodies. However, these methods mainly focus on the content related to fish body measurement, and have not conducted in-depth research on the spatial correlation and posture correction of fish body segmentation. At the same time, in a dynamic breeding environment, the stability and robustness of segmentation are poor. Therefore, the present invention proposes an underwater multi-eye stereoscopic vision accurate fish body segmentation system and segmentation method. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes an underwater multi-view stereoscopic vision accurate fish body segmentation system and segmentation method to solve the above problems existing in the prior art.

[0005] To achieve the above object, the present invention provides an underwater multi-eye stereoscopic vision accurate fish body segmentation system, comprising:

[0006] At least one underwater trinocular imaging module and data processing unit;

[0007] Wherein, the underwater three-eye imaging module is fixed in the aquaculture water body of the aquaculture tank, and the underwater three-eye imaging module is used to collect multi-view underwater images in the aquaculture tank;

[0008] The data processing unit is used to identify and segment the multi-view underwater image to obtain the segmentation result and spatial angle information of a single fish body.

[0009] Optionally, the three-eye imaging module includes: a first underwater imaging unit, a second underwater imaging unit, a third underwater imaging unit and a pool wall slide rail;

[0010] The first underwater imaging unit, the second underwater imaging unit and the third underwater imaging unit are waterproof cameras, and the first underwater imaging unit, the second underwater imaging unit and the third underwater imaging unit are arranged in an equilateral triangle;

[0011] Three waterproof cameras arranged in an equilateral triangle are installed on the inner wall of the breeding tank via pool wall slide rails.

[0012] Optionally, the data processing unit includes: a target recognition and segmentation module, a three-eye depth calculation module and a roll angle calculation module;

[0013] Wherein, the target recognition and segmentation module detects a single fish body based on the multi-view underwater image and segments the fish body to obtain a segmentation result;

[0014] The three-eye depth calculation module is used to establish the position relationship of the fish body in three-dimensional space;

[0015] The roll angle calculation module is used to estimate the fish body roll angle based on the segmentation result and the position relationship of the fish body in three-dimensional space.

[0016] The present invention also provides an underwater multi-eye stereoscopic vision accurate fish body segmentation method, comprising the following steps:

[0017] Collect multi-view underwater images based on underwater trinocular imaging module;

[0018] Inputting the multi-view underwater image into a data processing unit to obtain a segmentation result;

[0019] Based on the depth information, the segmentation result is converted into coordinates to obtain the position coordinates of the fish body parts in three-dimensional space;

[0020] Based on the position coordinates of the fish body parts in the three-dimensional space, the spatial correlation information of the fish body is constructed to obtain the roll angle;

[0021] The segmentation results are corrected and integrated based on the roll angle to obtain a final segmentation result.

[0022] Optionally, the process of inputting the multi-view underwater image into a data processing unit to obtain a segmentation result includes:

[0023] Constructing a fish target detection dataset based on the multi-view underwater images;

[0024] Using the fish body target detection data set to train the target detection model to obtain a trained target detection model;

[0025] Inputting the actual collected image into the trained target detection model to obtain a single fish marked by a positioning frame;

[0026] Constructing a single semantic segmentation dataset based on the single fish image;

[0027] Using the single semantic segmentation dataset to train the semantic segmentation model to obtain a trained semantic segmentation model;

[0028] Inputting the actual collected image into the trained semantic segmentation model to obtain a two-dimensional image of the fish body part;

[0029] A segmentation result is obtained based on the two-dimensional image of the fish body part.

[0030] Optionally, the process of segmenting the single fish image includes:

[0031] Scaling the single fish image to a size of 256×72 to obtain a scaled single fish image;

[0032] Downsampling the scaled single fish image to obtain a downsampled feature map;

[0033] The downsampled feature map is subjected to a deconvolution operation and a skip connection is performed to obtain a segmentation result.

[0034] Optionally, an expression for performing coordinate transformation on the segmentation result based on the depth information is:

[0035]

[0036] In the formula, Represents the coordinates of the pixel point. represents the world coordinates, K, R, T are the intrinsic parameter matrix, rotation matrix and translation matrix obtained by Zhang's calibration rule, and d represents the depth value between the camera optical center and the 3D scene point corresponding to the pixel point.

[0037] Optionally, the expression for calculating the roll angle is:

[0038]

[0039] In the formula, is the horizontal normal vector, is the center axis direction vector, and θ represents the roll angle.

[0040] Optionally, an expression for correcting the segmentation result based on the roll angle is:

[0041] S c =S o α(θ)

[0042] In the formula, S c and S o are the corrected segmentation boundary and the original boundary respectively, and α(θ) is the proportional coefficient.

[0043] Optionally, the expression for obtaining the final segmentation result by fusing the segmentation boundaries corrected by multiple perspectives is:

[0044]

[0045] In the formula, S final represents the final segmentation result, ω i represents the segmentation boundary weight of the i-th view.

[0046] Compared with the prior art, the present invention has the following advantages and technical effects:

[0047] The present invention provides an underwater multi-eye stereoscopic vision accurate fish segmentation system, which is fixed in the aquaculture tank through at least one underwater three-eye imaging module, and can collect multi-view underwater images in the tank. These images are then sent to the data processing unit for recognition and segmentation driven by deep learning algorithms, thereby obtaining accurate segmentation results of single fish and spatial angle information. This technological breakthrough significantly improves the accuracy and robustness of underwater fish segmentation, especially in complex underwater environments, and can effectively identify and segment fish.

[0048] The present invention also proposes an underwater multi-eye stereoscopic vision accurate fish segmentation method, which collects multi-view underwater images through an underwater three-eye imaging module, uses a data processing unit to perform image segmentation, and obtains the three-dimensional spatial position coordinates of the fish body parts. Further, based on these coordinates, the spatial correlation information of the fish body is constructed, and the roll angle of the fish body is calculated. Finally, the segmentation results are corrected and fused according to the roll angle to obtain the final accurate segmentation result. This method effectively overcomes the challenges brought by the complexity of the underwater environment and the changes in the posture of the fish body, improves the accuracy and robustness of the segmentation, and provides strong technical support for fish monitoring and management in aquaculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0050] Figure 1 Schematic diagram of an underwater multi-view stereoscopic accurate fish segmentation system and segmentation method according to an embodiment of the present invention;

[0051] Figure 2 The camera position structure and imaging field of view of the underwater multi-eye fish segmentation system according to an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the scene structure of an underwater multi-eye fish segmentation system according to an embodiment of the present invention;

[0053] Marking Description:

[0054] 1. Aquaculture tank; 11. Aquaculture water; 2. Underwater trinocular imaging module; 21. First underwater imaging unit; 22. Second underwater imaging unit; 23. Third underwater imaging unit; 24. Pool wall slide rail. DETAILED DESCRIPTION

[0055] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0056] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0057] Embodiment 1

[0058] The present invention combines the semantic segmentation model with the depth estimation technology to establish the spatial association of the fish body and clarify the tilt angle of the fish body, thereby optimizing the multi-eye segmentation results, improving the segmentation accuracy and robustness in complex underwater environments, and providing support for subsequent precise analysis and aquaculture management.

[0059] like Figure 1 As shown, a specific implementation example of the present invention provides an underwater fish body segmentation method using a trinocular camera. The method comprises at least the following steps: collecting multi-view underwater images based on an underwater trinocular imaging module 2; inputting the multi-view underwater images into a data processing unit to obtain a segmentation result; performing coordinate conversion on the segmentation result based on depth information to obtain the position coordinates of the fish body parts in three-dimensional space; constructing the spatial correlation information of the fish body based on the position coordinates of the fish body parts in three-dimensional space to obtain the roll angle; and correcting and fusing the segmentation result based on the roll angle to obtain the final segmentation result.

[0060] As a specific implementation of this embodiment, the following steps are included:

[0061] Step 1): Start the trinocular camera to collect underwater data from three perspectives and transmit the collected data to the data processing module.

[0062] Step 2): Construct a fish target detection dataset and train a target detection model based on it so that the system can accurately identify and locate individual fish from high-density underwater fish schools during real-time operation.

[0063] 2.1) Fish identification and positioning data collection and labeling: Collect real-time video stream data from the breeding pond, select key frame images with higher definition, and label all visible fish in the field of view. The labeling box (x, y, w, h) should cover the maximum contour area of ​​the fish as much as possible.

[0064] 2.2) Fish body positioning detection model training and reasoning: YOLOv8 target detection algorithm is used to locate the loss L loc , classification loss L cls , confidence loss L conf The model is trained as a joint loss function. The input image is first compressed to 640×640 size, and the initial feature map is extracted through the convolution layer. Assume that the output of the first layer is The feature map is halved layer by layerC n+1 =f conv (C n ),n∈{1,…4}, and finally make bounding box and classification predictions on the C5 layer feature map. Through the anchor-based method, each anchor outputs an offset value (Δx, Δy, Δw, Δh), and the final prediction box can be obtained by combining the prior box calculation: The softmax layer outputs the probability distribution of each category p = [p1, ..., p c ].

[0065] Step 3): construct a single semantic segmentation dataset and train the target detection model based on it so that the carcass and non-carcass parts of the fish can be segmented from the single positioning frame obtained in step 2) during real-time operation of the system.

[0066] 3.1) Fish body semantic segmentation data collection and annotation: The fish body area output by the target detection is annotated at the pixel level, and the annotated categories include the fish body, tail fin, dorsal fin, etc. Data enhancement operations are performed, including rotation (±15°), random noise addition, and image flipping to improve the robustness of the model.

[0067] 3.2) Fish segmentation model training and reasoning: Based on the UNet model, the model is trained with Dice loss and CE loss as the joint loss function. In the actual reasoning process, the individual fish data is first cropped based on the positioning box result predicted in 2.2), and the original image is scaled to 256×72 (to match the aspect ratio of the fish body). Subsequently, the size of the downsampled convolution feature map is successively reduced to 256×72, 128×36, 64×18, and 32×9 to obtain the downsampled feature map. Subsequently, the feature map size is restored through deconvolution operation, and skip-connection is performed at the same time, and finally the original segmentation map size is reconstructed and the final segmentation result is obtained through full convolution operation.

[0068] Step 4): Based on the bit depth obtained by the trinocular imaging system, realize the mapping of pixel coordinates to the world coordinate system. Assume that the coordinates of a pixel point are Then its world coordinate system is Among them, K, R, and T are the intrinsic parameter matrix, rotation matrix, and translation matrix obtained by Zhang's calibration rule.

[0069] Step 5): Combine the multi-eye segmentation results and multi-eye depth estimation to perform spatial association and estimate the fish body inclination angle.

[0070] 5.1) Map the semantic segmentation results {M1, M2, M3} to the corresponding depth map, and construct the point cloud data of fish body segmentation based on it, P i ={(x,y,z)|(u,v)∈M i}, where (x, y, z) are the coordinates of the point in three-dimensional space and (u, v) are the pixel coordinates.

[0071] 5.2) Use the camera's external parameter matrix to map the point clouds {P1, P2, P3} of each view into a unified three-dimensional coordinate system to build a multi-view spatial association. And for each point cloud point of each view, verify whether there is a corresponding view matching point: Where ∈ is the tolerance threshold.

[0072] 5.3) Use the RANSAC algorithm to fit the central axis of the fish body point cloud: Among them, p0 is the starting point of the fish body's central axis, is the center axis direction vector. Then, based on the center axis direction

[0073] The angle θ between the vector and the horizontal plane is used to calculate the fish body roll angle: in is the horizontal normal vector.

[0074] Step 6) Compensation is performed based on the estimated roll angle to fuse and optimize the segmentation results.

[0075] 6.1) Establish a collective model to correct the fin contour. Optionally, the proportional coefficient α(θ) is used to describe the change of the boundary between the fin and the fish body at different inclination angles, and the corrected boundary S c =S o ·α(θ). Where S c and S o are the corrected segmentation boundary and the original boundary respectively.

[0076] 6.2) Based on the corrected boundary in 6.1), the segmentation results of multiple perspectives are fused. Optionally, multiple perspectives are fused by weighted averaging: The final optimized segmentation result is obtained.

[0077] Through the above examples, the present invention obtains multi-view images and depth information of underwater fish bodies through a multi-eye imaging module. The fish body and its key parts are semantically segmented in combination with a deep learning model. The system uses the spatial correlation between multi-eye depth data and segmentation results to clarify the roll angle of the fish body, thereby optimizing the multi-view segmentation results and improving the segmentation accuracy and robustness. This method overcomes the problem of decreased segmentation accuracy caused by changes in fish posture under a fixed viewing angle, and can accurately segment fish bodies in a high-density breeding environment, laying the foundation for subsequent measurement tasks.

[0078] Embodiment 2

[0079] The present invention aims to collect data for calculation and processing by means of an expansion structure and an underwater trinocular camera without changing the existing aquaculture pond / warehouse structure of the aquaculture scene. When the system is first installed or the aquaculture water body changes, the trinocular camera needs to be calibrated jointly by Zhang's calibration method.

[0080] The specific embodiment of the present invention provides a system calibration system for a three-eye camera. The front view of the underwater three-eye imaging module 2 is as follows Figure 2 As shown, the example installation diagram is as follows Figure 3 As shown, it includes a first underwater imaging unit 21, a second underwater imaging unit 22 and a third underwater imaging unit 23, which are fixed in an equilateral triangle structure and are respectively used to collect underwater image data.

[0081] The present embodiment provides an underwater multi-eye stereoscopic vision accurate fish segmentation system, which comprises at least an underwater three-eye imaging module 2 and a processing unit.

[0082] The underwater three-eye imaging module 2 includes three groups of waterproof cameras, each of which has the same optical parameters and is installed according to a fixed geometric layout. It can collect multi-view images in complex water environments such as breeding ponds, and provide reliable data support for subsequent depth estimation and spatial correlation analysis. The underwater three-eye imaging module 2 is fixed in the breeding water body 11 of the breeding tank 1, and the underwater three-eye imaging module 2 is used to collect multi-view underwater images in the breeding tank 1.

[0083] The underwater three-eye imaging module 2 includes: a first underwater imaging unit 21, a second underwater imaging unit 22, a third underwater imaging unit 23 and a pool wall slide rail 24;

[0084] The first underwater imaging unit 21, the second underwater imaging unit 22, and the third underwater imaging unit 23 are waterproof cameras, and the first underwater imaging unit 21, the second underwater imaging unit 22, and the third underwater imaging unit 23 are arranged in an equilateral triangle;

[0085] Three waterproof cameras arranged in an equilateral triangle are installed on the inner wall of the breeding tank 1 via a pool wall slide rail 24 .

[0086] The data processing unit includes: a target recognition and segmentation module, a three-eye depth calculation module and a roll angle calculation module; wherein the target recognition and segmentation module detects a single fish body based on a multi-view underwater image and segments the fish body to obtain a segmentation result; the three-eye depth calculation module is used to establish a position relationship of the fish body in three-dimensional space; and the roll angle calculation module is used to estimate the fish body inclination angle based on the segmentation result and the position relationship of the fish body in three-dimensional space.

[0087] The processing unit is used to process the images collected by the multi-camera imaging module, including the construction of semantic segmentation model, depth estimation and segmentation result optimization, and specifically implements the following functions:

[0088] Extract the main area and key parts of the fish body (such as tail fin, dorsal fin, pelvic fin, etc.) based on the semantic segmentation model;

[0089] Combined with the depth estimation results of multi-view images, the three-dimensional spatial association of the fish segmentation area is established to clarify the tilt angle of the fish underwater.

[0090] The roll angle information is used to optimize the multi-view segmentation results to ensure the consistency and robustness of the segmentation results in different postures and dense environments.

[0091] Embodiment 3

[0092] This embodiment provides a fish body segmentation method based on underwater multi-view imaging, which at least includes the following steps:

[0093] S1 multi-camera imaging module installation and data acquisition: Fix the multi-camera imaging module in the aquaculture pond environment to collect clear multi-view underwater images.

[0094] S2 Construction of underwater fish target recognition model: Construct a target recognition dataset for aquaculture scenarios, and use a target detection algorithm to automatically identify fish in a dense underwater environment.

[0095] S3 Construction of underwater single fish semantic segmentation model: Construct pixel-level semantic segmentation annotation of single fish according to the full-length segmentation requirement, and use semantic segmentation algorithm to realize semantic segmentation of single fish obtained in S2.

[0096] S4 Segmentation result optimization: Depth information is obtained through multi-view parallax calculation, the segmentation result is mapped to three-dimensional space, and the roll angle of the fish body is calculated in combination with the depth estimation data. The roll angle information is used to optimize the segmentation result to improve the robustness of the segmented area in complex postures and dynamic environments.

[0097] Furthermore, S2 comprises the following steps:

[0098] S2.1 Target recognition data set construction, real-time image data collection in the aquaculture pond environment, and labeling of the horizontal circumscribed rectangle of the fish body.

[0099] S2.2 The target detection algorithm needs to identify the fish body area and return the pixel coordinates and confidence score of the fish body's external bounding box to facilitate subsequent individual fish segmentation processing.

[0100] Preferably, in S2.1, the collected image can be subjected to denoising, light correction and color enhancement processing.

[0101] Preferably, the target detection algorithm in S2.2 is based on the YOLOv11 target recognition model.

[0102] Further, S3 includes the following steps:

[0103] S3.1 The semantic segmentation dataset construction should include but not be limited to the main parts of the fish body and the non-carcass parts of the fish body (tail fin, dorsal fin, pelvic fin, etc.).

[0104] S3.2 The semantic segmentation algorithm needs to be consistent with the labeled data to segment the main body and non-carcass parts.

[0105] Preferably, in S3.1, the robustness of the model can be improved by data enhancement (such as rotation, flipping, and noise addition).

[0106] Preferably, the semantic segmentation model in S3.2 uses a UNet model.

[0107] Preferably, the depth calculation result in S4 can be optimized by filtering technology to reduce the interference of noise on the size segmentation result.

[0108] Compared with monocular or binocular imaging, the present invention significantly improves the accuracy and stability of fish segmentation results by combining multi-camera imaging modules with geometric layout design and depth estimation technology, especially in environments with complex lighting or changeable fish postures. Different from the existing underwater fish measurement methods based on target detection, the present invention uses semantic segmentation technology to accurately segment fish body parts, and can accurately identify the main body, tail fin and other parts of the fish. By integrating semantic segmentation with multi-camera depth estimation, the spatial correlation and roll angle of the fish body are clarified, and the segmentation results are optimized.

[0109] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An underwater multi-eye stereoscopic accurate fish segmentation system, characterized in that: include: At least one underwater three-eye imaging module (2) and a data processing unit; The underwater three-eye imaging module (2) is fixed in the aquaculture water body (11) of the aquaculture tank (1), and the underwater three-eye imaging module (2) is used to collect multi-view underwater images in the aquaculture tank (1); The data processing unit is used to identify and segment the multi-view underwater image to obtain the segmentation result and spatial angle information of a single fish body.

2. The underwater multi-eye stereoscopic accurate fish segmentation system according to claim 1, characterized in that: The underwater three-eye imaging module (2) comprises: a first underwater imaging unit (21), a second underwater imaging unit (22), a third underwater imaging unit (23) and a pool wall slide rail (24); The first underwater imaging unit (21), the second underwater imaging unit (22), and the third underwater imaging unit (23) are waterproof cameras, and the first underwater imaging unit (21), the second underwater imaging unit (22), and the third underwater imaging unit (23) are arranged in an equilateral triangle; Three waterproof cameras arranged in an equilateral triangle are installed on the inner wall of the breeding tank (1) via a pool wall slide rail (24).

3. The underwater multi-eye stereoscopic accurate fish segmentation system according to claim 1, characterized in that: The data processing unit includes: a target recognition and segmentation module, a three-eye depth calculation module and a roll angle calculation module; Wherein, the target recognition and segmentation module detects a single fish body based on the multi-view underwater image and segments the fish body to obtain a segmentation result; The three-eye depth calculation module is used to establish the position relationship of the fish body in three-dimensional space; The roll angle calculation module is used to estimate the fish body roll angle based on the segmentation result and the position relationship of the fish body in three-dimensional space.

4. An underwater multi-eye stereo vision accurate fish body segmentation method, characterized in that: The following steps are involved: Collecting multi-view underwater images based on an underwater three-eye imaging module (2); Inputting the multi-view underwater image into a data processing unit to obtain a segmentation result; Based on the depth information, the segmentation result is converted into coordinates to obtain the position coordinates of the fish body parts in three-dimensional space; Based on the position coordinates of the fish body parts in the three-dimensional space, the spatial correlation information of the fish body is constructed to obtain the roll angle; The segmentation results are corrected and integrated based on the roll angle to obtain a final segmentation result.

5. The underwater multi-eye stereoscopic accurate fish segmentation method according to claim 4, characterized in that: The process of inputting the multi-view underwater image into a data processing unit to obtain a segmentation result includes: Constructing a fish target detection dataset based on the multi-view underwater images; Using the fish body target detection data set to train the target detection model to obtain a trained target detection model; Inputting the actual collected image into the trained target detection model to obtain a single fish marked by a positioning frame; Constructing a single semantic segmentation dataset based on the single fish image; Using the single semantic segmentation dataset to train the semantic segmentation model to obtain a trained semantic segmentation model; Inputting the actual collected image into the trained semantic segmentation model to obtain a two-dimensional image of the fish body part; A segmentation result is obtained based on the two-dimensional image of the fish body part.

6. The underwater multi-eye stereoscopic accurate fish body segmentation method according to claim 5, characterized in that: The process of segmenting the single fish image includes: Scaling the single fish image to a size of 256×72 to obtain a scaled single fish image; Downsampling the scaled single fish image to obtain a downsampled feature map; The downsampled feature map is subjected to a deconvolution operation and a skip connection is performed to obtain a segmentation result.

7. The underwater multi-eye stereoscopic accurate fish body segmentation method according to claim 6, characterized in that: The expression for performing coordinate transformation on the segmentation result based on the depth information is: In the formula, Represents the coordinates of the pixel point. represents the world coordinates, K, R, T are the intrinsic parameter matrix, rotation matrix and translation matrix obtained by Zhang's calibration rule, and d represents the depth value between the camera optical center and the 3D scene point corresponding to the pixel point.

8. The underwater multi-eye stereoscopic accurate fish body segmentation method according to claim 7, characterized in that: The expression for calculating the roll angle is: In the formula, is the horizontal normal vector, is the center axis direction vector, and θ represents the roll angle.

9. The underwater multi-eye stereoscopic accurate fish body segmentation method according to claim 8, characterized in that: The expression for correcting the segmentation result based on the roll angle is: S c =S o ·a(i) In the formula, S c and S o are the corrected segmentation boundary and the original boundary respectively, and α(θ) is the proportional coefficient.

10. The underwater multi-eye stereoscopic accurate fish body segmentation method according to claim 9, characterized in that: The expression for the final segmentation result obtained by fusing the segmentation boundaries corrected by multiple perspectives is: In the formula, S final represents the final segmentation result, ω i represents the segmentation boundary weight of the i-th view.

Images