An automatic recognition system for pipeline shallow section images based on deep learning

Abstract

The present invention discloses a system and method for automatic recognition of pipeline shallow section images based on deep learning, which comprises an introduction module, a model improvement module, a data acquisition module, a training module and a recognition module. The model improvement module integrates an attention feature pyramid network, a target detection dynamic head structure and an optimization loss function into a basic model to form an initial recognition model. The attention feature pyramid network dynamically adjusts the feature map extraction weights; the target detection dynamic head structure adjusts parameters according to image complexity; and the optimization loss function adjusts the penalty terms of angle, scale and aspect ratio. The data acquisition module collects data to form a training set, a test set and a validation set. The training module trains the initial model based on the training set data, and adjusts the hyperparameters through the validation set until the model performance meets the preset conditions. The recognition module uses the optimization model to recognize the test set image to obtain the final result. The present invention can effectively improve the accuracy and efficiency of automatic recognition of pipeline shallow section images.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a pipeline shallow section image automatic recognition system based on deep learning. Background Art

[0002] Submarine pipelines, as crucial infrastructure in marine engineering, are widely used in oil and gas transportation, submarine communications, resource development, and environmental monitoring. Due to the unique characteristics of the submarine environment, pipeline safety monitoring and maintenance are particularly crucial, involving regular inspections, fault detection, and environmental adaptability assessments. With the continuous advancement of submarine engineering technology, pipeline monitoring methods are gradually shifting from traditional manual inspections and physical probing to more efficient and accurate automated technologies, with image recognition technology playing a key role.

[0003] Traditional submarine pipeline inspection methods typically rely on image processing-based technologies. Image recognition technology has become a key component, particularly when using ground-penetrating radar (GPR) for submarine pipeline inspection. Currently, advanced image recognition models such as the YOLOv5 algorithm, the dual YOLOv8-pose model, and the YOLOv8n algorithm are widely used in this field. The YOLO family of algorithms, with their efficient target detection capabilities, has achieved excellent results, particularly in terrestrial environments, enabling rapid target detection and location. The YOLOv5 and YOLOv8 family of algorithms, in particular, utilize deep learning models to optimize feature extraction and target detection accuracy, playing a significant role in practical applications.

[0004] However, while algorithms such as YOLOv5 and YOLOv8 perform well in traditional terrestrial environments, their application in complex submarine environments faces numerous challenges. Shallow cross-section images used in submarine pipeline inspections often exhibit complex background noise due to the unique characteristics of the underwater environment, and pipeline targets are small and have diverse shapes. The complex underwater background and uneven illumination in deep waters make it difficult for existing target detection technologies to meet these challenges. The recognition accuracy and efficiency of submarine pipelines are insufficient to meet practical requirements, especially for small targets, and the processing of background noise remains a major bottleneck for current technologies.

[0005] To address this issue, existing technologies have gradually begun to adopt the YOLOv8 algorithm for improvement, in order to improve the accuracy and efficiency of pipeline identification in more complex submarine environments. The YOLOv8 algorithm has multiple technical advantages over previous models. First, through an improved feature fusion mechanism (FPN+PAN), YOLOv8 can effectively process multi-scale features, thereby more accurately detecting small targets in submarine environments. Second, the dynamic anchor frame mechanism used by the YOLOv8 algorithm further enhances its adaptability in complex backgrounds, enabling it to more effectively identify pipeline targets of different scales in complex submarine scenes. In addition, the YOLOv8 algorithm's advanced data augmentation technology and dynamic label allocation strategy further enhance the model's generalization ability and detection accuracy, significantly reducing false detections and missed detections.

[0006] Despite this, the application of the YOLOv8 algorithm in submarine pipeline detection still faces certain limitations. For example, in low-resolution images, the detailed information of pipeline targets is easily lost, which affects the detection effect. Moreover, factors such as complex backgrounds and uneven lighting in the submarine environment often lead to interference from similar background objects, increasing the risk of false detection. In addition, although YOLOv8 has improved in recognition accuracy, its high computational requirements may limit its application in scenarios with high real-time requirements or limited hardware resources. Finally, YOLOv8's ability to handle target occlusion is still insufficient, and its performance is more dependent on the diversity and quality of training data, which also makes the model's performance in complex environments uncertain.

[0007] Therefore, although the YOLOv8 algorithm has shown great potential in submarine pipeline detection, in order to further improve its performance in complex submarine environments, the current YOLOv8 algorithm still needs to improve its submarine pipeline detection accuracy and efficiency in complex submarine environments. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a pipeline shallow section image automatic recognition system based on deep learning, which is used to improve the accuracy and efficiency of submarine pipeline detection in complex submarine environments.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a pipeline shallow section image automatic recognition system based on deep learning, comprising:

[0010] Import module, used to introduce the YOLOv8 model as the base model;

[0011] A model improvement module, connected to the introduction module, is used to integrate the attention feature pyramid network, the target detection dynamic head structure and the optimization loss function into the basic model to form an initial recognition model;

[0012] The attention feature pyramid network dynamically adjusts the weights of feature extraction based on different feature maps. The target detection dynamic head structure automatically adjusts the parameter configuration of the detection head based on the feature complexity of the input image. The optimization loss function dynamically adjusts the penalty terms of angle, scale and aspect ratio to optimize the directional error of the detection frame.

[0013] A data acquisition module is used to collect multiple submarine pipeline shallow section images, annotate them to form a pipeline shallow section image dataset, and then divide the pipeline shallow section image dataset into a training set, a test set, and a validation set;

[0014] a training module, connected to the data acquisition module and the model improvement module, respectively, for training the initial recognition model based on the data in the training set, and adjusting hyperparameters during the training of the initial recognition model based on the data in the validation set, until the performance parameters of the model meet the preset parameter conditions, and outputting the model as an optimized recognition model;

[0015] The recognition module is connected to the training module and the data acquisition module, and is used to recognize the image data in the test set according to the optimized recognition model to obtain the pipeline shallow section image recognition result.

[0016] Furthermore, the optimization loss function is a SIoU function, and the SIoU function is also used to dynamically adjust the penalty term of the center point distance. The penalty term calculation formula of the center point distance is configured as:

[0017] ,

[0018] in, A penalty term for the distance from the center point, and They are used to represent the central horizontal coordinate and central vertical coordinate of the prediction box of the target detection task, and They are used to represent the central horizontal coordinate and central vertical coordinate of the real box of the target detection task respectively.

[0019] Furthermore, the penalty calculation formula for the scale and aspect ratio is configured as follows:

[0020] ,

[0021] in, A penalty term for the scale and aspect ratio, and are used to represent the width and height of the prediction box respectively, and They are used to represent the width and height of the real frame respectively;

[0022] The penalty calculation formula for the angle is configured as:

[0023] ,

[0024] in, A penalty term representing the angle, and are respectively used to represent the angle between the predicted frame and the real frame;

[0025] The formula configuration of the SIoU function is:

[0026] ,

[0027] in, Used to represent the SIoU function, Used to represent the traditional IoU function, 、 and They are used to represent the preset first weight coefficient, second weight coefficient and third weight coefficient respectively.

[0028] Furthermore, the ratio of data in the training set, the test set and the validation set is 10:5:1.

[0029] Furthermore, the performance parameters include accuracy, recall, average precision, model parameter quantity and total floating-point operations.

[0030] Furthermore, it also includes an image processing module, which is connected to the data acquisition module and is used to perform background separation, image denoising and target enhancement on the original image data in the training set, the test set and the validation set in sequence to obtain optimized image data.

[0031] Furthermore, it also includes an environment detection module connected to the training module, the environment detection module is used to detect the light brightness and light uniformity of the seabed environment where the pipeline is located, and detect the relative distance from the pipeline;

[0032] The training module inputs the light illuminance, the light uniformity, and the relative distance into a preset learning rate optimization formula to calculate an optimized learning rate, and inputs the light illuminance, the light uniformity, and the relative distance into a preset weight optimization formula to calculate an optimized weight attenuation parameter;

[0033] The optimized recognition model improves the model structure according to the optimized learning rate and the optimized weight decay parameter and then outputs it.

[0034] Furthermore, the learning rate optimization formula is configured as:

[0035] ,

[0036] in, is used to represent the optimized learning rate, It is used to represent the initial learning rate of the optimized recognition model. It is used to express the adjustment coefficient of the light brightness to the learning rate. Used to indicate the brightness of the light, An exponential adjustment coefficient for expressing the effect of the light intensity on the learning rate. Used to express the relative distance influence coefficient, Used to indicate the uniformity of illumination, Used to express the relative distance, Used to represent the distance factor adjustment coefficient.

[0037] Furthermore, the weight optimization formula is configured as follows:

[0038] ,

[0039] in, Used to represent the optimized weight decay coefficient, It is used to represent the initial weight attenuation coefficient of the optimized recognition model, Used to represent the adjustment coefficient of the light brightness to the weight attenuation, It is used to represent the influence index of the light brightness on the attenuation parameter. Used to express the adjustment coefficient of lighting uniformity, It is used to represent the attenuation parameter adjustment coefficient related to the pipeline distance. Used to indicate the preset time attenuation coefficient, Used to represent time variables, An adjustment factor used to represent the effect of time on attenuation.

[0040] A pipeline shallow section image automatic recognition method based on deep learning, applied to the above-mentioned pipeline shallow section image automatic recognition system based on deep learning, comprising:

[0041] Step S1, the introduction module introduces the YOLOv8 model as the basic model;

[0042] In step S2, the model improvement module integrates the attention feature pyramid network, the target detection dynamic head structure, and the optimization loss function into the basic model to form an initial recognition model. The attention feature pyramid network dynamically adjusts the weight of feature extraction according to different feature maps. The target detection dynamic head structure automatically adjusts the parameter configuration of the detection head according to the feature complexity of the input image. The optimization loss function dynamically adjusts the penalty terms of angle, scale, and aspect ratio to optimize the directional error of the detection box.

[0043] Step S3: The data acquisition module acquires multiple submarine pipeline shallow section images, annotates them to form a pipeline shallow section image dataset, and then divides the pipeline shallow section image dataset into a training set, a test set, and a validation set;

[0044] Step S4: The training module trains the initial recognition model based on the data in the training set, and adjusts hyperparameters during the training of the initial recognition model based on the data in the validation set, until the performance parameters of the model meet the preset parameter conditions, and outputs the model as the optimized recognition model;

[0045] Step S5: The recognition module recognizes the image data in the test set according to the optimized recognition model to obtain pipeline shallow section image recognition results.

[0046] Beneficial effects of the present invention:

[0047] This paper improves the YOLOv8 model by integrating optimization techniques such as an attention feature pyramid network, a dynamic head structure for target detection, and an optimized loss function. Combined with high-quality data acquisition and training mechanisms, this method yields an optimized recognition model. This significantly improves the accuracy, stability, and efficiency of target detection in shallow submarine pipeline profiles. Compared to traditional detection methods, this method achieves more efficient and accurate target detection in complex submarine environments, with excellent real-time performance and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a structural diagram of the pipeline shallow section image automatic recognition system of the present invention;

[0049] Figure 2 It is a model structure diagram of the optimized recognition model in the present invention;

[0050] Figure 3 is the image to be detected in Example 1 of the present invention;

[0051] Figure 4 is the image detected by the YOLOv8 original model in Example 1 of the present invention;

[0052] Figure 5 is an image detected by the optimized recognition model in Example 1 of the present invention;

[0053] Figure 6 It is a flow chart of the steps of the automatic recognition method of pipeline shallow section images in the present invention.

[0054] Figure numerals: 1. Introduction module; 2. Model improvement module; 3. Data acquisition module; 4. Training module; 5. Recognition module; 6. Image processing module; 7. Environment detection module. DETAILED DESCRIPTION

[0055] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom," "top," "inner," and "outer" refer to directions toward or away from the geometric center of a particular component, respectively.

[0056] Example 1, with reference to Figure 1 , which is the first embodiment of the present invention, provides a deep learning-based automatic recognition system for pipeline shallow section images, which can improve the accuracy and efficiency of submarine pipeline detection in complex submarine environments, including:

[0057] Module 1 is introduced to introduce the YOLOv8 model as the base model;

[0058] Model improvement module 2, connected to introduction module 1, is used to integrate the attention feature pyramid network, target detection dynamic head structure and optimized loss function into the basic model to form an initial recognition model;

[0059] The attention feature pyramid network dynamically adjusts the weights of feature extraction based on different feature maps. The dynamic head structure for target detection automatically adjusts the parameter configuration of the detection head based on the feature complexity of the input image. The optimization loss function dynamically adjusts the penalty terms of angle, scale, and aspect ratio to optimize the directional error of the detection frame.

[0060] Data acquisition module 3 is used to collect multiple submarine pipeline shallow section images, annotate them to form a pipeline shallow section image dataset, and then divide the pipeline shallow section image dataset into a training set, a test set, and a validation set;

[0061] The training module 4 is connected to the data acquisition module 3 and the model improvement module 2, and is used to train the initial recognition model according to the data in the training set, and adjust the hyperparameters during the initial recognition model training process according to the data in the validation set until the performance parameters of the model meet the preset parameter conditions and output it as the optimized recognition model, wherein the performance parameters include accuracy, recall rate, average precision, number of model parameters and total floating-point operations;

[0062] The recognition module 5 is connected to the training module 4 and the data acquisition module 3, and is used to recognize the image data in the test set according to the optimized recognition model to obtain the pipeline shallow section image recognition result.

[0063] Working principle of embodiment 1:

[0064] Module 1 introduces YOLOv8 as the foundational model for object detection, providing structural support for subsequent improvements. By loading the pre-trained YOLOv8 model weights, the model is equipped with preliminary object detection capabilities.

[0065] Based on the YOLOv8 model, the model improvement module 2 integrates the attention feature pyramid network, the target detection dynamic head structure and the optimization loss function to form an initial recognition model.

[0066] The Attention Feature Pyramid Network (AFPN) introduces an adaptive feature fusion mechanism to optimize feature extraction at different scales, enhancing the detection of small objects (such as pipeline fractures and foreign objects). As an adaptive feature fusion architecture, it has significantly improved the performance of YOLOv8. Building on the Feature Pyramid Network, AFPN introduces an adaptive weighting mechanism that dynamically adjusts weights based on feature maps at different levels, enabling the network to more accurately extract object features at different scales. Compared to the traditional FPN (Feature Pyramid Network), AFPN more effectively addresses the issues of feature redundancy and insufficiency in multi-scale object detection, particularly when detecting small objects. Furthermore, AFPN's cross-scale information flow design further enhances feature complementarity and information transfer efficiency. This not only improves the network's object detection capabilities in complex scenes, but also reduces false and missed detections, particularly in challenging scenarios such as occlusion and complex backgrounds. Experiments demonstrating that integrating AFPN into the YOLOv8 model improves both mean average precision (mAP) and detection speed. Especially in the evaluation of pipeline shallow section image dataset, the YOLOv8 model enhanced by AFPN showed higher accuracy and better generalization ability compared with the traditional FPN structure.

[0067] Object Detection Dynamic Head (LADH): This model utilizes a lightweight adaptive detection head that dynamically adjusts its parameters based on the complexity of the input image, reducing computational resource consumption and improving detection speed. LADH (Lightweight Adaptive Dynamic Head) introduces a more flexible and efficient object detection head architecture, significantly improving model performance. LADH's core advantage lies in its lightweight design and adaptive nature, enabling the network to significantly improve detection speed and resource efficiency while maintaining high accuracy. First, LADH employs a dynamic adjustment strategy that automatically adjusts the detection head's parameters based on the feature complexity of the input image. This adaptive mechanism enables the network to dynamically adjust the allocation of computational resources to suit different input scenarios, enabling more efficient extraction and utilization of key features, especially when dealing with complex scenes or extreme object distributions. This dynamic architecture significantly reduces computational overhead and improves inference efficiency. Second, LADH's lightweight design reduces the redundant computations in traditional detection heads while maintaining sufficient expressive power. Its modularity and weight sharing strategies reduce model parameters while ensuring high sensitivity and accuracy to target features. Experiments show that LADH significantly reduces the number of model parameters and computational complexity of YOLOv8, and demonstrates superior performance in edge devices and real-time applications.

[0068] By integrating LADH, the YOLOv8 model demonstrates excellent performance on multiple standard datasets. In an evaluation of a pipeline shallow section image dataset, the LADH architecture significantly reduces the model's inference time while maintaining a high mean average precision (mAP). Compared to a static head architecture, LADH significantly improves the YOLOv8 model's detection capabilities in multi-target and multi-scale scenarios, and significantly enhances the detection accuracy of small objects.

[0069] SIoU (Significant Loss Function) optimizes the matching of target frames, reduces the error of detection frames, and improves the robustness of target detection by dynamically adjusting the penalty terms of angle, scale, and aspect ratio.

[0070] Data acquisition module 3 collects images of measured shallow profile data of pipelines in a certain area of the seabed. These images contain 163 shallow profile images of pipelines at different angles and sizes, and use the image annotation tool LabelImg to form a pipeline shallow profile image dataset. The dataset is divided into a training set, a test set, and a validation set for model training, testing, and tuning. The data ratio of the training set, test set, and validation set is 10:5:1. Table 1 shows the data ratio of the training set, test set, and validation set.

[0071] Table 1

[0072]

[0073] In Table 1, the number of image data in the training set, test set, and validation set are 100, 50, and 10, respectively.

[0074] Training module 4 uses the data from the training set to train the initial recognition model. During the training process, hyperparameters (such as learning rate, batch size, optimizer type, and weight decay parameters) are adjusted using data from the validation set to improve model performance. The Adam optimizer is used for gradient updates during training, combined with a cosine annealing learning rate schedule to gradually optimize model parameters.

[0075] Among them, Table 2 is the experimental environment parameter table, and the experimental environment parameters are shown in Table 2:

[0076] Table 2

[0077]

[0078] In Example 1, the initial learning rate of training is set to 0.01, the initial weight decay parameter is 0.0005, the batch sample size is 64, the number of training rounds is 200, and the input image size is 640×640×3.

[0079] When the mean average precision (mAP) on the validation set reaches a preset threshold (e.g., above 85%), the optimized recognition model is output.

[0080] Recognition module 5 evaluates the optimized recognition model using a test set, calculating metrics such as precision, recall, and mean average precision (mAP) to verify model performance. The recognition results are output as object detection box coordinates, category labels, and confidence scores for subsequent applications such as submarine pipeline health monitoring and defect identification.

[0081] The model structure diagram of the improved optimization recognition model is as follows Figure 2As shown in the figure, Conv represents the convolution layer for feature extraction; C2f represents the feature fusion module based on CSPNet, which is used for more efficient feature transfer; ASFF_2 represents a two-layer adaptive feature fusion layer, which adaptively fuses multi-scale features by learning weights to solve the inconsistency problem between features of different scales; ASFF_3 represents a three-layer adaptive feature fusion layer, which adaptively fuses multi-scale features by learning weights to solve the inconsistency problem between features of different scales; Basic Block represents the basic building block, which is a module that combines convolution, normalization, activation function and CSP structure; Detect represents the detection head, which is responsible for the final target bounding box and category prediction. SPPF is the abbreviation of Spatial Pyramid Pooling Fast, which is an improved pooling module mainly used to process multi-scale feature fusion and optimize computational efficiency.

[0082] In this embodiment, the following evaluation metrics are used to effectively evaluate the performance of the improved model: precision, recall, mean average precision (mAP), model parameter count (params), and total floating-point operations (FLOPs). Precision measures the proportion of instances predicted as positive that are actually positive. Recall evaluates the model's ability to successfully predict positive examples. Mean average precision (mAP) represents the average of the AP values across categories and is a comprehensive metric for evaluating the algorithm's overall performance across different categories. Model parameter count refers to the total number of trainable parameters in a neural network, including weights and biases. These parameters are updated during training and used for model optimization and prediction. A larger number of parameters generally indicates a more complex model with stronger expressive power. Total floating-point operations is one measure of a model's computational complexity; a larger number of FLOPs indicates the greater the computing resources required for model inference and training. Table 3 is a comparison result table of the YOLOv8 original model and the YOLOv8 optimized recognition model in this embodiment. The comparison results of the YOLOv8 original model and the YOLOv8 optimized recognition model in this embodiment are shown in Table 3:

[0083] Table 3

[0084]

[0085] Comparing the original YOLOv8 model with the improved optimized recognition model clearly demonstrates significant advantages across all metrics. The optimized recognition model achieved a 39.3% increase in accuracy, a 7.6% increase in recall, and an 11.5% increase in mean average precision (mAP). These results demonstrate that the improved optimized recognition model significantly improves both prediction accuracy and overall detection performance in object detection tasks. Furthermore, the number of model parameters was reduced by 78.9%, and the total floating-point operations decreased by 82.3%, significantly reducing the model's complexity and computational cost. These improvements not only effectively enhance the model's performance but also enhance its potential for application on resource-constrained devices, such as mobile devices and embedded systems. Overall, the improved optimized recognition model achieves lightweight and efficient performance while maintaining high accuracy, offering broader practical application prospects.

[0086] In this embodiment, Figure 3 is the image to be detected, Figure 4 This is the image detected by the original YOLOv8 model. Figure 5 To optimize the image after recognition model detection. Figures 3 to 5 It can be found that the detection accuracy of the original YOLOv8 model in noisy and small target environments is relatively weaker than that of the improved YOLOv8 model, and the detection accuracy of the image detected by the optimized recognition model is higher.

[0087] Preferably, the optimization loss function is the SIoU function, which is also used to dynamically adjust the penalty term of the center point distance. The penalty term calculation formula of the center point distance is configured as:

[0088] ,

[0089] in, The penalty term used to represent the distance from the center point, and They are used to represent the central horizontal coordinate and central vertical coordinate of the prediction box of the target detection task, and They are used to represent the central horizontal coordinate and central vertical coordinate of the real box of the target detection task respectively.

[0090] Preferably, the penalty calculation formula for scale and aspect ratio is configured as follows:

[0091] ,

[0092] in, A penalty term for scale and aspect ratio, and They are used to represent the width and height of the prediction box respectively. and They are used to represent the width and height of the real frame respectively;

[0093] The penalty calculation formula for the angle is configured as:

[0094] ,

[0095] in, A penalty term for angles, and They are used to represent the angles between the predicted box and the real box;

[0096] The formula configuration of the SIoU function is:

[0097] ,

[0098] in, Used to represent the SIoU function, Used to represent the traditional IoU function, 、 and They are used to represent the preset first weight coefficient, second weight coefficient and third weight coefficient respectively.

[0099] Specifically, in this embodiment, by incorporating the SIoU function, the YOLOv8 model can more accurately fit the target boundary when processing irregular or non-horizontal targets. Secondly, the SIoU function solves the problem of the traditional IoU function being insufficiently sensitive when detecting small targets by dynamically adjusting the target box scale. Experiments show that the SIoU function improves the regression accuracy of the YOLOv8 model for small targets and targets with large aspect ratio differences, while reducing the mismatch between the box and the target size, thereby effectively reducing the missed detection rate and false detection rate.

[0100] In summary, the introduction of the SIoU function makes the YOLOv8 model's target positioning more accurate, especially in complex scenes. The experimental results in Table 1 show that after integrating the SIoU function, the YOLOv8 model further improves the mean average precision (mAP) on the pipeline shallow section image dataset and has better generalization capabilities when handling objects of different scales and shapes.

[0101] Example 2 is the second embodiment of the present invention. Different from the previous embodiment, this embodiment provides an image processing module 6, which can further improve the detection accuracy and robustness of the model. It includes an image processing module 6 connected to a data acquisition module 3, which is used to perform background separation, image denoising and target enhancement on the original image data in the training set, test set and validation set in turn to obtain optimized image data.

[0102] Working principle of embodiment 2:

[0103] The specific processing steps of the image processing module 6 are as follows:

[0104] Background separation:

[0105] Due to the complex seabed environment, background interference (such as marine sediments, seaweed, rocks, etc.) may affect the accuracy of pipeline detection, so background separation is required to highlight the target area.

[0106] Method: Gaussian mixture model is used for background modeling to separate pipeline targets and background.

[0107] Image denoising:

[0108] Since seabed images are affected by water media and are often accompanied by noise (such as scattering noise, uneven illumination noise, etc.), denoising is required to improve image clarity.

[0109] Method: Adaptive wavelet denoising: decompose the image at multiple scales, remove high-frequency noise and improve image quality.

[0110] Target Enhancement:

[0111] Since targets in seabed images may have problems such as low contrast and blurred edges, target enhancement is required to make key targets such as pipelines clearer.

[0112] Method: Adaptive histogram equalization: avoids the problem of excessive noise enhancement in ordinary histogram equalization and makes the target details clearer.

[0113] Example 3 is the third embodiment of the present invention. Unlike the previous embodiment, this embodiment provides an environment detection module 7, which can improve the adaptability of the target detection model in complex submarine environments, while improving detection accuracy and reducing false detection rates. The environment detection module 7 includes an environment detection module 7 connected to the training module 4. The environment detection module 7 is used to detect the light brightness and uniformity of the submarine environment where the pipeline is located, as well as the relative distance from the pipeline.

[0114] Training module 4 inputs the light illuminance, light uniformity, and relative distance into a preset learning rate optimization formula to calculate an optimized learning rate, and inputs the light illuminance, light uniformity, and relative distance into a preset weight optimization formula to calculate an optimized weight attenuation parameter;

[0115] The optimized recognition model outputs the model structure after improving it by optimizing the learning rate and optimizing the weight decay parameters.

[0116] Working principle of embodiment 3:

[0117] Light intensity is detected by calculating the average grayscale value of pixels in the image, and illumination uniformity is detected by calculating the difference in local window mean values. The smaller the difference in local window mean values, the higher the illumination uniformity. Relative distance is detected by a depth camera. The optimized learning rate and weight decay parameters are calculated based on light intensity, illumination uniformity, and relative distance to further improve the model structure of the optimized recognition model. This eliminates the effects of complex illumination changes and uneven target distances on model recognition accuracy, allowing the improved optimized recognition model to maintain high accuracy even in the presence of complex illumination changes and uneven target distances, thereby improving the adaptability of the target detection model in complex seabed environments.

[0118] Preferably, the learning rate optimization formula is configured as:

[0119] ,

[0120] in, Used to represent the optimized learning rate, Used to represent the initial learning rate of the optimized recognition model, Used to express the adjustment coefficient of light intensity on learning rate, reflecting the impact of light on learning rate optimization;

[0121] Used to indicate light brightness. An exponential adjustment coefficient used to represent the effect of light intensity on the learning rate, controlling the intensity of the effect of light on the learning rate; Used to represent the relative distance influence coefficient, adjusting the relationship between illumination uniformity and pipeline relative distance to influence the learning rate; Used to indicate the uniformity of illumination. Used to express relative distance. Used to represent the distance factor adjustment coefficient, which controls the influence of relative distance on the learning rate.

[0122] Specifically, the key parameters in this embodiment are adjusted as follows:

[0123] Light intensity: In low-light environments (such as deep underwater), the light intensity is small, which may lead to low learning efficiency during training. Therefore, the optimization formula and Adjustments need to be made so that the learning rate increases in low-light environments to speed up the convergence of the model.

[0124] Lighting uniformity: Lighting uniformity directly affects the clarity of pipeline images. If the lighting distribution is uneven ( ≈0), the recognition effect of the model will be affected, so the relative distance influence coefficient The learning rate is optimized to enhance the model's learning ability under uneven lighting conditions.

[0125] Relative distance from the pipeline: When the pipeline is far from the camera, the clarity and details of the image may be incomplete, affecting the recognition effect. In this case, a larger relative distance value will reduce the learning rate, while a smaller relative distance value (closer pipeline) will increase the learning rate to better capture image details.

[0126] The parameters of one of the images in a set of submarine pipeline shallow section image data are as follows:

[0127] Light brightness: =0.4

[0128] Light uniformity: =0.8

[0129] Relative distance from pipeline: =5.0

[0130] Substituting these values into the formula, we set the initial learning rate to =0.001, adjustment coefficient of light brightness =0.5, light brightness index influence coefficient =1.2, lighting uniformity impact factor =0.3, relative distance adjustment coefficient =0.7;

[0131] By calculation, the optimized learning rate is obtained , this learning rate will be used for the next update in the training process.

[0132] Implementation effect:

[0133] Improved training results: By dynamically adjusting the learning rate, the model can better adapt to different environmental lighting conditions and pipeline locations, improving the accuracy of the YOLOv8 model in submarine pipeline identification.

[0134] Convergence speed: The optimized learning rate can speed up the convergence of the model, especially in low-light or uneven lighting environments, so the model can be trained more stably.

[0135] Preferably, the weight optimization formula is configured as:

[0136] ,

[0137] in, Used to represent the optimized weight decay coefficient, Used to represent the initial weight attenuation coefficient of the optimized recognition model and set the initial value of the attenuation parameter; Used to express the adjustment coefficient of light brightness on weight attenuation, and adjust the intensity of the influence of light on the attenuation parameter; It is used to represent the influence index of light brightness on attenuation parameters and adjust the intensity of the influence of light on weight attenuation; Used to represent the adjustment coefficient of lighting uniformity, indicating the influence of lighting uniformity on weight attenuation; It is used to express the attenuation parameter adjustment coefficient related to pipeline distance, and controls the influence of relative position on attenuation. Used to represent the preset time decay coefficient and control the speed of the exponential decay function; Used to represent time variables, An adjustment factor used to represent the effect of time on attenuation.

[0138] Specifically, in this embodiment, light intensity affects image quality and indirectly impacts the stability and effectiveness of model training. When light intensity is high, the optimized weight decay coefficient increases, thereby improving the model's regularization strength and preventing overfitting. When light intensity is low, the decay coefficient decreases appropriately, allowing the model to more carefully learn the characteristics of darker environments.

[0139] The impact of lighting uniformity is reflected in image clarity. In uniform lighting environments, image quality is better, and the optimized weight decay coefficient is smaller, which helps the model learn features more precisely. In uneven lighting, weight decay is increased to reduce the possibility of overfitting the model.

[0140] The relative position of the pipeline and the camera affects the model's recognition difficulty. When the pipeline is far away, the pipeline features in the image become blurred, and the model requires stronger regularization to prevent overfitting. Therefore, the optimized attenuation coefficient is increased to enhance the model's regularization effect.

[0141] Time decay : As training progresses, the exponential decay function controls the gradual decrease of weight decay. In the early stages of training, a larger weight decay can prevent overfitting; in the later stages of training, the decay gradually decreases, allowing the model to fit the data more flexibly. To further enhance the control of time-related attenuation and make training more efficient.

[0142] The value range of is [0,∞), representing the optimized weight decay value, which affects the strength of model regularization. Larger optimized weight decay coefficient values indicate stronger model regularization, preventing overfitting. Smaller optimized weight decay coefficient values help improve the model's fitting ability, but may introduce the risk of overfitting.

[0143] Practical Application:

[0144] Optimization of lighting conditions: In brighter environments, the effect of lighting on weight decay increases, so the decay parameter is moderately increased to enhance regularization and prevent overfitting. In weaker environments, the model requires a higher decay parameter to prevent overfitting in low-light conditions.

[0145] The impact of illumination uniformity: Under uniform illumination conditions, the model can stably identify features, so the attenuation value is small. However, under uneven illumination conditions, the attenuation value needs to be increased to enhance the model's generalization ability.

[0146] Influence of pipeline position: As the pipeline distance increases, the optimized attenuation value will increase to prevent the long-distance pipeline features from overfitting the training results.

[0147] Control of time decay: As the training process progresses, the time decay factor enables the optimization process to gradually reduce weight decay, thereby adapting to more detailed feature learning and preventing overfitting.

[0148] A pipeline shallow section image automatic recognition method based on deep learning is applied to the above pipeline shallow section image automatic recognition system based on deep learning, such as Figure 6 As shown, including:

[0149] Step S1, introducing module 1 introduces the YOLOv8 model as the basic model;

[0150] In step S2, the model improvement module 2 integrates the attention feature pyramid network, the target detection dynamic head structure, and the optimization loss function into the basic model to form an initial recognition model. The attention feature pyramid network dynamically adjusts the weights of feature extraction according to different feature maps. The target detection dynamic head structure automatically adjusts the parameter configuration of the detection head according to the feature complexity of the input image. The optimization loss function dynamically adjusts the penalty terms of angle, scale, and aspect ratio to optimize the directional error of the detection frame.

[0151] Step S3: the data acquisition module 3 acquires a plurality of submarine pipeline shallow section images, annotates them to form a pipeline shallow section image dataset, and then divides the pipeline shallow section image dataset into a training set, a test set, and a validation set;

[0152] Step S4: The training module 4 trains the initial recognition model based on the data in the training set, and adjusts the hyperparameters during the initial recognition model training process based on the data in the validation set until the performance parameters of the model meet the preset parameter conditions and outputs it as the optimized recognition model;

[0153] In step S5 , the recognition module 5 recognizes the image data in the test set according to the optimized recognition model to obtain pipeline shallow section image recognition results.

[0154] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that improvements and modifications that do not depart from the principles of the present invention are within the scope of protection of the present invention.

Claims

1. A pipeline shallow section image automatic recognition system based on deep learning, characterized by: include: Introduction module (1) is used to introduce the YOLOv8 model as the basic model; A model improvement module (2), connected to the introduction module (1), is used to integrate the attention feature pyramid network, the target detection dynamic head structure and the optimization loss function into the basic model to form an initial recognition model; The attention feature pyramid network dynamically adjusts the weights of feature extraction based on different feature maps. The target detection dynamic head structure automatically adjusts the parameter configuration of the detection head based on the feature complexity of the input image. The optimization loss function dynamically adjusts the penalty terms of angle, scale and aspect ratio to optimize the directional error of the detection frame. A data acquisition module (3) is used to collect multiple submarine pipeline shallow section images, annotate them to form a pipeline shallow section image data set, and then divide the pipeline shallow section image data set into a training set, a test set and a validation set; A training module (4), connected to the data acquisition module (3) and the model improvement module (2), respectively, for training the initial recognition model according to the data in the training set, and adjusting hyperparameters during the training of the initial recognition model according to the data in the validation set, until the performance parameters of the model meet the preset parameter conditions and outputting the model as an optimized recognition model; A recognition module (5) is connected to the training module (4) and the data acquisition module (3), and is used to recognize the image data in the test set according to the optimized recognition model to obtain a pipeline shallow section image recognition result; It also includes an environment detection module (7) connected to the training module (4), and the environment detection module (7) is used to detect the light brightness and light uniformity of the seabed environment where the pipeline is located, and detect the relative distance from the pipeline; The training module (4) inputs the light luminance, the light uniformity, and the relative distance into a preset learning rate optimization formula to calculate an optimized learning rate, and inputs the light luminance, the light uniformity, and the relative distance into a preset weight optimization formula to calculate an optimized weight attenuation parameter; The optimized recognition model improves the model structure according to the optimized learning rate and the optimized weight decay parameter and outputs the result; The learning rate optimization formula is configured as: ； in, is used to represent the optimized learning rate, It is used to represent the initial learning rate of the optimized recognition model. It is used to express the adjustment coefficient of the light brightness to the learning rate. Used to indicate the brightness of the light, An exponential adjustment coefficient for expressing the effect of the light intensity on the learning rate. Used to express the relative distance influence coefficient, Used to indicate the uniformity of illumination, Used to express the relative distance, Used to indicate the distance factor adjustment coefficient; The weight optimization formula is configured as follows: ； in, Used to represent the optimized weight decay coefficient, It is used to represent the initial weight attenuation coefficient of the optimized recognition model, Used to represent the adjustment coefficient of the light brightness to the weight attenuation, It is used to represent the influence index of the light brightness on the attenuation parameter. Used to express the adjustment coefficient of lighting uniformity, It is used to represent the attenuation parameter adjustment coefficient related to the pipeline distance. Used to indicate the preset time attenuation coefficient, Used to represent time variables, An adjustment factor used to represent the effect of time on attenuation.

2. The deep learning-based pipeline shallow section image automatic recognition system according to claim 1 is characterized by: The optimization loss function is a SIoU function, which is also used to dynamically adjust the penalty term of the center point distance. The penalty term calculation formula of the center point distance is configured as: ； in, A penalty term for the distance from the center point, and They are used to represent the central horizontal coordinate and central vertical coordinate of the prediction box of the target detection task, and They are used to represent the central horizontal coordinate and central vertical coordinate of the real box of the target detection task respectively.

3. The deep learning-based pipeline shallow section image automatic recognition system according to claim 2, characterized in that: The penalty calculation formula for the scale and aspect ratio is configured as follows: ； in, A penalty term for the scale and aspect ratio, and are used to represent the width and height of the prediction box respectively, and They are used to represent the width and height of the real frame respectively; The penalty calculation formula for the angle is configured as: ， in, A penalty term representing the angle, and are respectively used to represent the angle between the predicted frame and the real frame; The formula configuration of the SIoU function is: ； in, Used to represent the SIoU function, Used to represent the traditional IoU function, 、 and They are used to represent the preset first weight coefficient, second weight coefficient and third weight coefficient respectively.

4. The deep learning-based pipeline shallow section image automatic recognition system according to claim 1, characterized in that: The ratio of data in the training set, the test set, and the validation set is 10:5:

1.

5. The deep learning-based pipeline shallow section image automatic recognition system according to claim 1 is characterized by: The performance parameters include accuracy, recall, average precision, model parameters and total floating-point operations.

6. The pipeline shallow section image automatic recognition system based on deep learning according to claim 1 is characterized in that: It also includes an image processing module (6) connected to the data acquisition module (3) for sequentially performing background separation, image denoising and target enhancement on the original image data in the training set, the test set and the validation set to obtain optimized image data.

7. A pipeline shallow section image automatic recognition method based on deep learning, applied to the pipeline shallow section image automatic recognition system based on deep learning according to any one of claims 1 to 6, characterized in that: include: Step S1, introducing module (1) to introduce the YOLOv8 model as the basic model; Step S2, the model improvement module (2) integrates the attention feature pyramid network, the target detection dynamic head structure and the optimization loss function into the basic model to form an initial recognition model, the attention feature pyramid network dynamically adjusts the weight of feature extraction according to different feature maps, the target detection dynamic head structure automatically adjusts the parameter configuration of the detection head according to the feature complexity of the input image, and the optimization loss function dynamically adjusts the penalty terms of angle, scale and aspect ratio to optimize the direction error of the detection frame; Step S3, the data acquisition module (3) acquires a plurality of submarine pipeline shallow section images, and annotates them to form a pipeline shallow section image dataset, and then divides the pipeline shallow section image dataset into a training set, a test set, and a validation set; Step S4, the training module (4) trains the initial recognition model according to the data in the training set, and adjusts the hyperparameters in the initial recognition model training process according to the data in the validation set, until the performance parameters of the model meet the preset parameter conditions and outputs it as the optimized recognition model; Step S5, the recognition module (5) recognizes the image data in the test set according to the optimized recognition model to obtain pipeline shallow section image recognition results.

Images