Research on underwater feature recognition and tracking based on deep learning

Abstract

The present application relates to the technical field of underwater image processing, and more particularly to underwater feature recognition and tracking research based on deep learning, which improves image processing quality by analyzing from the perspective of image information processing, and adopts an algorithm combining adaptive contrast enhancement and limited contrast adaptive histogram equalization to improve image contrast, highlight feature edge information, and denoising the enhanced feature image through a denoising template method to effectively suppress noise interference in underwater images, so as to provide data support for subsequent feature recognition, help improve underwater feature recognition accuracy, and extract underwater weld feature information from the filtered and optimized feature image based on the optimal YOLOv8 network model to further improve feature recognition accuracy and accuracy, and calibrate the laser sensor posture through feature information mapping to improve feature tracking accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater image processing, and in particular to underwater feature recognition and tracking research based on deep learning. BACKGROUND

[0002] Welding defects are the most difficult to detect defect forms of ship surface damage, and their features include cracks, pores, slag inclusions, etc., which are extremely difficult to detect and identify. When detecting underwater targets, especially underwater ship bodies, considering the effect of water, the information collected by a single sensor alone cannot obtain accurate detection results of welding defects.

[0003] Among them, underwater optical images are often affected by noise interference, blurred texture features, low contrast and color distortion, etc. These problems seriously affect the image recognition effect. However, in the prior art, underwater optical images cannot be effectively processed, which leads to poor image processing effect, is not conducive to the extraction and identification of underwater features, and reduces the extraction and identification efficiency of underwater features, and also affects the subsequent feature tracking trajectory generation accuracy, and reduces the automation efficiency of underwater machines and the safety of feature tracking trajectories.

[0004] In view of the above technical defects, a solution is proposed. SUMMARY

[0005] The purpose of the present application is to provide underwater feature recognition and tracking research based on deep learning to solve the above technical defects. The present application analyzes from the perspective of image information processing to improve image processing quality, extracts underwater weld feature information from the filtered and optimized feature image based on the YOLOv8 network model, and improves feature recognition accuracy and accuracy. The posture of the laser sensor is calibrated through the feature information mapping method to provide technical support for subsequent feature tracking to improve feature tracking accuracy. The generated feature tracking trajectory is actually tracked and self-checked through the information feedback method to adjust the real-time movement trajectory of the welding robot in time to improve the movement accuracy of the welding robot. At the same time, the feature tracking trajectory generation scheme is further managed and adjusted to reduce the risk of obstacle avoidance failure of the welding robot.

[0006] The purpose of the present application can be achieved by the following technical scheme: underwater feature recognition and tracking research based on deep learning, comprising the following steps:

[0007] S1, scanning underwater features, capturing underwater feature contour information, and setting the underwater feature contour information as a feature image to be processed;

[0008] S2, the industrial computer filters: the collected feature image to be processed is subjected to image enhancement processing by an algorithm combining adaptive contrast enhancement and limited contrast adaptive histogram equalization (CLAHE), to obtain an enhanced feature image;

[0009] S3, denoising processing: the enhanced feature image is subjected to denoising processing by a denoising template, to obtain a filter-optimized feature image;

[0010] S4: a deep learning model is built and model training parameters are set, and a YOLOv8 network model is run to extract underwater weld feature information from the filter-optimized feature image;

[0011] S5: the underwater weld feature information is accurately mapped into the coordinate system of the underwater machine, so as to calibrate the pose of the laser sensor;

[0012] S6: based on the pose calibration, an RRT (rapidly-exploring random tree) path planning algorithm is used to generate a feature tracking trajectory, a controller drives the welding robot to move the welding torch according to the feature tracking trajectory, deviation warning feedback analysis is performed on the feature tracking trajectory of the welding robot movement, the obtained avoidance deviation index is judged and processed, and the obtained adjustment instruction is outputted and fed back.

[0013] Preferably, the acquisition process of the enhanced feature image is as follows:

[0014] The to-be-processed feature image is divided into a high-frequency image and a low-frequency image through template filtering operation, the high-frequency image is further enhanced, and the enhanced high-frequency image and the low-frequency image are fused through an image enhancement algorithm to obtain a processed enhanced feature image; the image enhancement algorithm is mathematically expressed as follows: g(x,y)=m d (x,y)+G(x,y)[f(x,y)-m d (x,y)], wherein g(x,y) represents the gray value of the processed image at point (x,y), f(x,y) represents the gray value of the image at (x,y) before processing; m d (x,y) is the low-frequency image part obtained after filtering, and G(x,y) is a dynamic gain function, the value of G(x,y) can change with the change of the processed pixel point.

[0015] The mean filtering method is used as a means to obtain the low-frequency image, and the specific mathematical expression is as follows: The ratio of the image gray mean value to the neighborhood standard deviation of the image pixel point is set as the dynamic gain, and the following formula is established: Wherein σ(x,y) is defined as follows: The template radius size n is artificially set to 2.

[0016] Preferably, the method for obtaining the denoising template is as follows:

[0017] The denoised template image is obtained using the following formula: Where M(x,y) is the image element value of the denoising template, C(x,y) is the gray value of the pixel in the image after the closing operation, and k is the noise judgment threshold, and k is a constant. The purpose of this operation is to compare the difference in gray values ​​of the pixels in the enhanced feature image and the denoised enhanced feature image after the closing operation, and to mark the granular black spots in the enhanced feature image where the difference in gray values ​​of the pixels in the enhanced feature image and the denoised enhanced feature image after the closing operation is greater than a preset threshold as potential noise points. The gray values ​​of the potential noise points are replaced by the gray value information of other points using the median filtering method, and the original enhanced feature image is updated after each denoising operation on a potential noise point, finally obtaining the filtered optimized feature image.

[0018] Preferably, the YOLOv8 network model construction process is as follows:

[0019] Set the model training parameters, construct a dataset based on the model training parameters, and divide the dataset into a training set and a validation set. The training set is used to teach the neural network model, while the validation set is responsible for verifying the effectiveness of the model's learning results.

[0020] The model is then run and full training monitoring is enabled. The main settings to be modified are the training rounds, batch size, and learning rate. The changes in training set bounding box loss, target loss, and classification loss during training are monitored to ultimately help select the best-performing YOLOv8 network model.

[0021] Preferably, the underwater weld feature information is accurately mapped onto the coordinate system of the underwater machine in order to calibrate the attitude of the laser sensor. The specific process is as follows:

[0022] Define two coordinate systems: {S} represents the sensor coordinate system, and {B} represents the reference coordinate system. The coordinates of point P in these two coordinate systems are expressed as follows: B P and S P;

[0023] Determine the transformation matrix between the sensor coordinate system {S} and the end coordinate system {E}. For any known point in its coordinate system S Q, the formula for transforming it to the machine base coordinate system is:

[0024] Preferably, the transformation matrix between the sensor coordinate system {S} and the end coordinate system {E} is... Calculate the transformation matrix. The specific steps are as follows:

[0025] SS1: determine a certain point P on the target, record the position of point P in the reference coordinate system {B} B P = (x B ,y B ,z B ,1) T ;

[0026] SS2: adjust the position of the machine to ensure that the laser line emitted by the sensor mounted thereon can pass through point P, and also record the position of point P in the sensor coordinate system {S} S P = (x S ,0,z S ,1) T ;

[0027] SS3: switch the current tool coordinate system of the machine to the end coordinate system {E}, record the pose data in the end coordinate system {E} at this time, and use the Euler rotation theorem to calculate the rotation matrix and further SS4: according to the transformation relationship of point P in space, we have

[0028] For any known point S Q in its coordinate system, the formula for transforming it to the machine base coordinate system is: where, B Q is the position coordinates of point Q in the reference coordinate system {B}, S Q is the position coordinates of point Q in the sensor coordinate system {S}.

[0029] Preferably, the deviation warning feedback analysis process is as follows:

[0030] The real-time moving trajectory of the welding robot is obtained, and the real-time moving trajectory is compared with the generated feature tracking trajectory for coincidence degree, so as to obtain the coincidence degree between the real-time moving trajectory and the generated feature tracking trajectory, and set the coincidence degree between the real-time moving trajectory and the generated feature tracking trajectory as the moving deviation warning coefficient, and the moving deviation warning coefficient is discriminated to generate a feedback instruction or an alarm instruction.

[0031] When the feedback instruction is generated, the generated feature tracking trajectory is divided into an avoidance tracking trajectory and a non-avoidance tracking trajectory, the real-time avoidance trajectory of the welding robot is obtained, the minimum avoidance distance of the avoidance tracking trajectory of the welding robot is obtained, the minimum avoidance distance represents the minimum straight-line distance between the obstacle and the welding robot, the actual minimum avoidance distance of the real-time avoidance trajectory of the welding robot is obtained, the avoidance deviation index is set as the value obtained by subtracting the minimum avoidance distance from the actual minimum avoidance distance, and the avoidance deviation index is discriminated: if the avoidance deviation index is greater than or equal to a preset avoidance deviation index threshold, no signal is generated; if the avoidance deviation index is less than the preset avoidance deviation index threshold, an adjustment instruction is generated.

[0032] The beneficial effects of the present application are as follows:

[0033] (1) The present application improves image processing quality by analyzing from the perspective of image information processing, adopts an algorithm combining adaptive contrast enhancement and limited contrast adaptive histogram equalization to improve image contrast, highlight feature edge information, and performs denoising processing on the enhanced feature image through a denoising template method to effectively suppress noise interference in underwater images, thereby providing data support for subsequent feature recognition.

[0034] (2) The present application extracts underwater weld feature information from the filtered and optimized feature image based on the YOLOv8 network model, thereby improving feature recognition accuracy and precision, and calibrates the laser sensor posture through feature information mapping to provide technical support for subsequent feature tracking and improve feature tracking accuracy. The generated feature tracking trajectory is actually tracked and self-checked through information feedback to timely adjust the real-time movement trajectory of the welding robot, thereby improving the movement accuracy of the welding robot and further managing and adjusting the feature tracking trajectory generation scheme to reduce the risk of obstacle avoidance failure of the welding robot. BRIEF DESCRIPTION OF DRAWINGS

[0035] The present application will be further described below in conjunction with the accompanying drawings;

[0036] Fig. 1 is the reference diagram of the method of the present application;

[0037] Fig. 2 is the local analysis reference diagram of the present application. DETAILED DESCRIPTION

[0038] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0039] Embodiment one:

[0040] Please refer to Figs. 1-2 The present application is a kind of underwater feature recognition and tracking research based on deep learning, comprising the following steps:

[0041] S1, scanning underwater features, capturing underwater feature contour information, and setting the underwater feature contour information as a feature image to be processed;

[0042] S2, the industrial computer carries out filtering: the adaptive contrast enhancement and the algorithm combining the limited contrast adaptive histogram equalization (CLAHE) are used for image enhancement processing on the collected feature image to be processed, and an enhanced feature image is obtained;

[0043] S3, denoising processing: the enhanced feature image is denoised by a denoising template to obtain a filter-optimized feature image;

[0044] S4: build a deep learning model and set the model training parameters, run the YOLOv8 network model to extract underwater weld feature information from the filter-optimized feature image;

[0045] S5: accurately map the underwater weld feature information to the coordinate system of the underwater machine, so as to calibrate the pose of the laser sensor;

[0046] S6: based on the pose calibration, a RRT (rapidly-exploring random tree) path planning algorithm is used to generate a feature tracking trajectory, a controller drives the welding robot to guide the movement of the welding gun according to the feature tracking trajectory, deviation warning feedback analysis is performed on the feature tracking trajectory of the welding robot movement, the obtained avoidance deviation index is judged and processed, and the obtained adjustment instruction is outputted and fed back.

[0047] In the embodiments of the present application: S1: scanning underwater features, capturing underwater feature contour information, and setting the underwater feature contour information as a feature image to be processed;

[0048] Specifically, the system scans underwater features with the help of a laser sensor installed at the end of the robot arm, that is, the laser sensor scans the weldment to collect the contour information of the weld groove. The main target of scanning is to capture the initial data of the underwater feature edge, which is the premise of realizing accurate underwater feature tracking;

[0049] Laser sensors use triangulation technology to accurately extract distance information of target objects and delineate the outline details of objects through detailed analysis of scan data.

[0050] During the scanning operation, the laser sensor is securely mounted at the end of the underwater robotic arm and kept parallel to the arm to ensure that the projected laser line can accurately hit the detection surface at a 90-degree angle. This design aims to fully cover and record the underwater features in detail. Data is continuously collected by manipulating the movement of the machine, and then stitched together to form a complete tracking area morphology map to ensure a complete underwater feature map.

[0051] S2: Industrial control computer performs filtering: The acquired feature image to be processed is enhanced using an algorithm combining adaptive contrast enhancement and contrast-limited adaptive histogram equalization (CLAHE) to obtain the enhanced feature image. The specific process is as follows:

[0052] The feature image to be processed is divided into high-frequency and low-frequency images through template filtering. The high-frequency image is then enhanced, and the enhanced high-frequency image is fused with the low-frequency image using an image enhancement algorithm to obtain the enhanced feature image. The mathematical expression of the image enhancement algorithm is as follows: g(x,y)=m d (x,y)+G(x,y)[f(x,y)-m d [x,y], where g(x,y) represents the gray value of the processed image at point (x,y), and f(x,y) represents the gray value of the unprocessed image at point (x,y); m d (x,y) represents the low-frequency image portion obtained after filtering, and G(x,y) is the dynamic gain function. The value of G(x,y) can change as the number of processed pixels changes.

[0053] Mean filtering is used as a method to obtain low-frequency images. The specific mathematical expression is as follows: If we use the ratio of the image's grayscale mean to its standard deviation within the neighborhood as the dynamic gain, then the following formula holds: The definition of σ(x,y) is as follows: The template radius n is set to 2; MJ represents the mean gray level of the image; the standard deviation σ in the neighborhood of the point (x,y) is the mean gray level; ε is the correction parameter; and G is the dynamic gain at the point (x,y).

[0054] S3: Denoising process: The enhanced feature image is denoised using a denoising template to obtain a filtered and optimized feature image;

[0055] The denoising template acquisition method: according to the feature that the morphological closing operation can complete small holes in the enhanced feature image, that is, the black spots composed of impurity noise information, the number of black spots in the enhanced feature image after the morphological closing operation will be significantly reduced, and the originally black pixel points will get new gray values; wherein the pixel value of the black spot is significantly smaller than the average gray value of the image or the median gray value of the overall image; thus the denoising template image can be obtained by the following formula: Wherein, M(x,y) is the image element value of the denoising template, C(x,y) is the gray value of the pixel point of the image after the closing operation, k is the noise judgment threshold, and k is a constant. The purpose of this operation is to compare the difference between the pixel gray values of the enhanced feature image and the closing denoising enhanced feature image, and mark the granular black spots in the enhanced feature image whose difference is greater than the preset threshold as potential noise points. The gray value of the potential noise point is replaced by the gray value of other points by using the median filtering method, and the original enhanced feature image is updated after each potential noise point is filtered, and finally a filtered and optimized feature image is obtained, which can obtain better filtering effect.

[0056] Example two:

[0057] S4: build a deep learning model and set the model training parameters, run the YOLOv8 network model to extract underwater weld feature information from the filtered and optimized feature image;

[0058] Specifically, the model training parameters are set, and the model is run to start training whole-process monitoring. The main modification is the setting value of training rounds, batch size and learning rate. The changes of boundary box loss, target loss and classification loss in the training process are monitored.

[0059] Further, based on the model training parameters, a data set is constructed, and the data set is divided into a training set and a validation set. The training set is used for learning and adaptation of the model, and the validation set is used for monitoring and checking the training process.

[0060] The training set is used to teach the neural network model, and the validation set is responsible for verifying the effectiveness of the learning results of the model, helping to select the optimal YOLOv8 network model, and thus avoiding the problems of overfitting or underfitting;

[0061] The optimal YOLOv8 network model is adopted, and the main structure is to divide the grid first, and each grid is responsible for the prediction of the target in the corresponding region, and finally the non-maximum suppression is performed to output.

[0062] In this embodiment, for underwater feature detection, the number of detection categories is 1. The network model needs to configure the size of the prior bounding boxes, the backbone, and the head. The relevant parameters include which layer the input comes from, the number of layers, the module type, and other parameters. When building the network model, the depth coefficient depth_multiplied by the number of layers is used to adjust the depth, and the width coefficient width_multiplied by the number of convolutional kernels is used to obtain the actual number of convolutional kernels. The width of the network model is adjusted in this way.

[0063] Finally, the optimal YOLOv8 target detection network is used to extract weld feature information;

[0064] S5: Accurately map the underwater weld feature information into the coordinate system of the underwater machine in order to calibrate the attitude of the laser sensor;

[0065] Specifically, two coordinate systems are defined: {S} represents the sensor coordinate system, and {B} represents the reference coordinate system. The coordinates of point P in these two coordinate systems are respectively expressed as: B P (in the reference coordinate system) and S In sensor coordinate system P (where P is the sensor coordinate system), the purpose of sensor attitude calibration is to determine the transformation matrix of the sensor coordinate system {S} relative to the end coordinate system {E} as accurately as possible. For any known point in its coordinate system S Q, the formula for transforming it to the machine base coordinate system is:

[0066] Wherein, the transformation matrix between the sensor coordinate system {S} and the end coordinate system {E} is... Calculate the transformation matrix. The specific steps are as follows:

[0067] SS1: Determine a specific point P on the target and record the position of point P in the reference coordinate system {B}. B P=(x B ,y B ,z B ,1) T ;

[0068] SS2: Adjust the machine's position to ensure that the laser line emitted by its onboard sensor passes through point P. Also, record the position of point P in the sensor's coordinate system {S}. S P=(x S ,0,z S ,1) T ;

[0069] SS3: switch the current tool coordinate system of the machine to the end coordinate system {E}, record the pose data in the end coordinate system {E} at this time, and calculate the rotation matrix by using Euler rotation theorem Further, further obtain

[0070] SS4: according to the transformation relationship of point P in space

[0071] For any known point in its coordinate system S Q, the transformation formula to the base coordinate system of the machine is: In the formula, B Q is the position coordinates of point Q in the reference coordinate system {B}, S Q is the position coordinates of point Q in the sensor coordinate system {S}, is the calculation result of formula is the calculation result of formula

[0072] S6: based on the pose calibration, a RRT (rapidly-exploring random tree) path planning algorithm is used to generate a feature tracking trajectory, and a controller drives the welding robot to move according to the feature tracking trajectory, while a deviation warning feedback analysis is performed on the feature tracking trajectory of the welding robot, and the obtained avoidance deviation index is processed to output the adjustment instruction feedback.

[0073] The real-time movement trajectory of the welding robot is obtained, and the real-time movement trajectory is compared with the generated feature tracking trajectory to obtain the coincidence degree between the real-time movement trajectory and the generated feature tracking trajectory, and the coincidence degree between the real-time movement trajectory and the generated feature tracking trajectory is set as a movement deviation warning coefficient, and the movement deviation warning coefficient is processed:

[0074] If the movement deviation warning coefficient is less than the preset movement deviation warning coefficient threshold, a feedback instruction is generated.

[0075] If the movement deviation warning coefficient is greater than or equal to the preset movement deviation warning coefficient threshold, an alarm instruction is generated, and the feedback instruction or the alarm instruction is outputted, and the preset warning operation corresponding to the feedback instruction or the alarm instruction is immediately performed, so as to timely adjust the real-time movement trajectory of the welding robot, and improve the movement accuracy of the welding robot.

[0076] When the feedback instruction is generated, the generated feature tracking trajectory is divided into an avoidance tracking trajectory and a non-avoidance tracking trajectory, the real-time avoidance trajectory of the welding robot is obtained, the minimum avoidance distance of the avoidance tracking trajectory of the welding robot is obtained, the minimum avoidance distance represents the minimum straight-line distance between the obstacle and the welding robot, the actual minimum avoidance distance of the real-time avoidance trajectory of the welding robot is obtained, the avoidance deviation index is set as the value obtained by subtracting the minimum avoidance distance from the actual minimum avoidance distance, and the avoidance deviation index is subjected to discrimination processing:

[0077] If the avoidance deviation index is greater than or equal to the preset avoidance deviation index threshold, no signal is generated.

[0078] If the avoidance deviation index is less than the preset avoidance deviation index threshold, an adjustment instruction is generated, the adjustment instruction is outputted feedback, and the preset warning operation corresponding to the adjustment instruction is immediately made, so as to timely manage and adjust the feature tracking trajectory generation scheme, thereby reducing the obstacle avoidance failure risk of the welding robot.

[0079] To sum up, the application analyzes from the perspective of image information processing to improve image processing quality, adopts the algorithm combining adaptive contrast enhancement and limited contrast adaptive histogram equalization to improve image contrast and highlight feature edge information, and performs denoising processing on the enhanced feature image through the denoising template method to effectively suppress noise interference in the underwater image, so as to provide data support for subsequent feature recognition, extract underwater weld feature information from the filtered and optimized feature image based on the YOLOv8 network model, and then improve feature recognition accuracy and accuracy, and calibrate the laser sensor posture through the feature information mapping method, so as to provide technical support for subsequent feature tracking, improve feature tracking accuracy, and perform actual tracking self-checking on the generated feature tracking trajectory through the information feedback method, so as to timely adjust the real-time movement trajectory of the welding robot, improve the movement accuracy of the welding robot, and further manage and adjust the feature tracking trajectory generation scheme to reduce the obstacle avoidance failure risk of the welding robot.

[0080] The size of the threshold is set to facilitate comparison, and the size of the threshold depends on the number of sample data and the base number set by the person skilled in the art for each group of sample data.

[0081] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent replacement or change according to the technical solution and inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A deep learning-based underwater feature recognition and tracking research, characterized in that, The method comprises the following steps: S1, scanning the underwater feature, capturing the underwater feature contour information, and setting the underwater feature contour information as a feature image to be processed; S2, the industrial computer performs filtering: the collected feature image to be processed is subjected to image enhancement processing by using an algorithm combining adaptive contrast enhancement and limited contrast adaptive histogram equalization (CLAHE), to obtain an enhanced feature image; S3, denoising processing: the enhanced feature image is subjected to denoising processing by using a denoising template, to obtain a filter-optimized feature image; S4: a deep learning model is built and model training parameters are set, and a YOLOv8 network model is run to extract underwater weld feature information from the filter-optimized feature image; S5: accurately map the underwater weld feature information to the coordinate system of the underwater machine, so as to calibrate the pose of the laser sensor; S6: based on the pose calibration, a RRT (rapidly-exploring random tree) path planning algorithm is used to generate a feature tracking trajectory, and a controller drives the welding robot to guide the movement of the welding torch according to the feature tracking trajectory, a deviation warning feedback analysis is performed on the feature tracking trajectory of the welding robot movement, a deviation avoidance index obtained is discriminated, and an adjustment instruction obtained is outputted and fed back; The YOLOv8 network model construction process is as follows: The model training parameters are set, a data set is constructed based on the model training parameters, the data set is divided into a training set and a validation set, the training set is used to teach the neural network model, and the validation set is responsible for verifying the effectiveness of the model learning results; And run the model to start training whole-process monitoring, mainly need to modify the training round, batch size and learning rate setting value, monitor the change of the training set boundary box loss, target loss and classification loss in the training process, and finally help to screen out the YOLOv8 network model with the best performance.

2. The underwater feature recognition and tracking research based on deep learning according to claim 1, characterized in that, The acquisition process of the enhanced feature image is as follows: The feature image to be processed is divided into a high-frequency image and a low-frequency image through a template filtering operation, the high-frequency image is further processed to be enhanced, and the enhanced high-frequency image and the low-frequency image are fused through an image enhancement algorithm to obtain an enhanced feature image after processing, and the image enhancement algorithm is mathematically expressed as follows: Wherein, represents the gray value of the image after processing at point , represents the gray value of the image before processing at point ; is the low-frequency image part obtained after filtering, is a dynamic gain function, and the value changes with the change of the pixel point to be processed. The mean filter method is used as a means to obtain the low-frequency image, and the specific mathematical expression is shown as follows: The ratio of the image gray mean value and the standard deviation in the neighborhood is used as the dynamic gain, and the following formula is established: wherein, The definition of is shown as follows: The template radius size n is artificially set to 2.

3. The underwater feature recognition and tracking research based on deep learning according to claim 2, characterized in that, The denoising template acquisition method is as follows: The denoising template image is obtained by the following formula: wherein, is the image element value of the denoising template, is the gray value of the image pixel point after the closing operation, is the noise judgment threshold, and is a constant, the purpose of the operation is to compare the difference between the enhanced feature image and the pixel gray value of the closing operation denoising enhanced feature image, and mark the granular black spots in the enhanced feature image whose difference between the enhanced feature image and the pixel gray value of the closing operation denoising enhanced feature image is greater than the preset threshold as potential noise points; the gray value of the potential noise points is replaced by the gray value of other points by the median filtering method, and the original enhanced feature image is updated after each potential noise point is filtered, and finally the filtered and optimized feature image is obtained. 4.The underwater feature recognition and tracking research based on deep learning according to claim 1, characterized in that, The underwater weld feature information is accurately mapped to the coordinate system of the underwater machine, so as to calibrate the pose of the laser sensor, and the specific process is as follows: Two coordinate systems are defined: Let S denote the sensor coordinate system, Let R denote the reference coordinate system, the coordinates of point P in these two coordinate systems are denoted as and respectively; determining the sensor coordinate system relative to the end coordinate system the transformation matrix for any known point in its coordinate system the formula for transforming it to the machine base coordinate system is: .

5. The underwater feature recognition and tracking research based on deep learning according to claim 4, characterized in that, Sensor coordinate system with respect to the end coordinate system the transformation matrix The specific steps for calculating the transformation matrix are as follows: SS1 : determine a certain point P on the target, record the position of the point P in the reference coordinate system ;​ SS2: Adjust the position of the machine, ensuring that the laser line emitted by the sensor it carries can cross point P, also recording the position of point P in the sensor coordinate system ;​ SS3: switch the current tool coordinate system of the machine to the end coordinate system , record the pose data in the end coordinate system at this time, use Euler rotation theorem to calculate the rotation matrix , and further obtain ; SS4: According to the transformation relationship of P point in space has: ; For any known point in its coordinate system The formula for transforming it to the machine base coordinate system is: In the formula, Point Q in the reference coordinate system Location coordinates, Let point Q be in the sensor coordinate system The location coordinates.

6. The underwater feature recognition and tracking research based on deep learning according to claim 5, characterized in that, The deviation warning feedback analysis process is as follows: The real-time movement trajectory of the welding robot is obtained, and the real-time movement trajectory is compared with the generated feature tracking trajectory for coincidence degree, so as to obtain the coincidence degree between the real-time movement trajectory and the generated feature tracking trajectory, set the coincidence degree between the real-time movement trajectory and the generated feature tracking trajectory as a movement deviation warning coefficient, and discriminate the movement deviation warning coefficient, to generate a feedback instruction or an alarm instruction; When the feedback instruction is generated, the generated feature tracking trajectory is divided into an avoidance tracking trajectory and a non-avoidance tracking trajectory, the real-time avoidance trajectory of the welding robot is obtained, the minimum avoidance distance of the avoidance tracking trajectory of the welding robot is obtained, the minimum avoidance distance represents the minimum straight-line distance between the obstacle and the welding robot, the actual minimum avoidance distance of the real-time avoidance trajectory of the welding robot is obtained, the avoidance deviation index is set as the value obtained by subtracting the minimum avoidance distance from the actual minimum avoidance distance, and the avoidance deviation index is subjected to a discrimination process: if the avoidance deviation index is greater than or equal to a preset avoidance deviation index threshold, no signal is generated; if the avoidance deviation index is less than the preset avoidance deviation index threshold, an adjustment instruction is generated.

Images