Long-distance pipeline exposure detection method, device, storage medium and electronic equipment

Abstract

The present disclosure discloses a long-distance pipeline exposure detection method, device, storage medium and electronic device, the method comprising the following steps: S100: collecting images of the area where the long-distance pipeline to be tested is located; S200: constructing a long-distance pipeline exposure detection model and training the model; S300: inputting the collected images of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model to detect whether the long-distance pipeline to be tested in the image is exposed. The present disclosure adopts the improved YOLOX‑S network, which can quickly and accurately identify whether the long-distance pipeline is exposed in its area, thereby improving the safety and reliability of the long-distance pipeline.

Description

Technical Field

[0001] The present invention relates to the field of long-distance pipeline detection, and in particular to a long-distance pipeline exposure detection method, device, storage medium and electronic equipment. Background Art

[0002] With the popularity of long-distance pipelines, the safety issues of long-distance pipelines have become more and more noticeable. Among them, pipeline exposure is one of the common safety hazards of long-distance pipelines. Once the pipeline is exposed, it is easy to be damaged by external objects and the environment, which seriously endangers the safe operation of the pipeline. During the installation and use of long-distance pipelines, pipeline surface defects may occur. If these surface defects cannot be checked in time, especially dangerous defects such as cracks, high-temperature corrosion pits, and deformations, they will bring great safety hazards and seriously affect the safety of long-distance pipelines. At present, the detection methods for photographing pipeline surface defects are mostly single-angle, and can only be shot locally. The existing method of detecting surface defects on the inner wall of long-distance pipelines through manual inspections has the disadvantages of inconvenient operation, low efficiency, and poor accuracy. Therefore, it is very necessary to improve the existing detection methods. Summary of the invention

[0003] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a long-distance pipeline exposure detection method, which can quickly and accurately identify whether a long-distance pipeline is exposed in its area, thereby improving the safety and reliability of the pipeline.

[0004] To achieve the above objectives, the present disclosure provides the following technical solutions:

[0005] A long-distance pipeline exposure detection method comprises the following steps:

[0006] S100: collecting images of the area where the long-distance pipeline to be tested is located;

[0007] S200: Build a long-distance pipeline exposure detection model and train the model;

[0008] The long-distance pipeline exposure detection model adopts an improved YOLOX-S network, which includes a backbone network, an improved feature extraction and fusion module, and an improved prediction and output module; the improved feature extraction and fusion module additionally introduces a Ghost module and an attention mechanism module, and the improved prediction and output module additionally introduces an AuxHead module;

[0009] S300: Inputting the collected image of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model to detect whether the long-distance pipeline to be tested in the image is exposed.

[0010] Preferably, in step S200, the long-distance pipeline exposure detection model is trained by the following steps:

[0011] S201: Construct an image dataset of long-distance pipelines with exposure;

[0012] S202: Screening and labeling the image data set;

[0013] S203: Divide the labeled image dataset into a training set and a test set;

[0014] S204: Setting training parameters, and training the model using the training set. When the maximum number of training rounds is reached, the model training is completed.

[0015] S205: Use the test set to test the trained model. During the test, when the average precision reaches 0.95 or above, the test passes and the trained model is obtained; otherwise, the training parameters are adjusted to retrain the model.

[0016] The present disclosure also provides a long-distance pipeline exposure detection device, the device comprising:

[0017] An acquisition module is used to acquire images of the area where the long-distance pipeline to be tested is located;

[0018] A model building and training module, used to build a long-distance pipeline exposure detection model and train the model; wherein the long-distance pipeline exposure detection model adopts an improved YOLOX-S network, and the improved YOLOX-S network includes a backbone network, an improved feature extraction and fusion module, and an improved prediction and output module; the improved feature extraction and fusion module additionally introduces a Ghost module and an attention mechanism module, and the improved prediction and output module additionally introduces an AuxHead module;

[0019] The detection module is used to input the collected image of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model to detect whether the long-distance pipeline to be tested in the image is exposed.

[0020] Preferably, the model building and training module includes:

[0021] The dataset construction submodule is used to construct a dataset of images of long-distance pipelines with exposure;

[0022] The screening and annotation submodule is used to screen and annotate the image dataset;

[0023] A partitioning submodule is used to divide the labeled image dataset into a training set and a test set;

[0024] The training submodule is used to train the model using the training set. When the maximum number of training rounds is reached, the model training is completed;

[0025] The test submodule is used to test the trained model using the test set.

[0026] The present disclosure also provides an electronic device, comprising:

[0027] A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein:

[0028] When the processor executes the program, any of the above methods is implemented.

[0029] The present disclosure also provides a computer storage medium storing computer executable instructions for executing any of the above methods.

[0030] Compared with the prior art, the beneficial effects brought by the present disclosure are: the present disclosure can quickly and accurately identify whether a long-distance pipeline is exposed in its area, thereby improving the safety and reliability of the pipeline, and is suitable for the inspection and maintenance of various long-distance pipelines. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of a long-distance pipeline exposure detection method based on a drone provided by an embodiment of the present disclosure;

[0032] Figure 2 is a schematic diagram of a network structure of a long-distance pipeline exposure detection model provided by another embodiment of the present disclosure;

[0033] Figure 3 Another embodiment of the present disclosure provides Figure 2 Schematic diagram of the structure of the Ghost module in the model shown;

[0034] Figure 4 Another embodiment of the present disclosure provides Figure 2 Schematic diagram of the structure of the attention module in the shown model;

[0035] Figure 5 It is a structural diagram of an AuxHead module provided in another embodiment of the present disclosure. DETAILED DESCRIPTION

[0036] The following will refer to the attached Figures 1 to 5 Specific embodiments of the present disclosure are described in detail. Although specific embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0037] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present disclosure, but the description is based on the general principles of the specification and is not used to limit the scope of the present disclosure. The scope of protection of the present disclosure shall be determined by the attached claims.

[0038] To facilitate understanding of the embodiments of the present disclosure, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and each of the accompanying drawings does not constitute a limitation on the embodiments of the present disclosure.

[0039] In one embodiment, if Figure 1 As shown, the present disclosure provides a long-distance pipeline exposure detection method, comprising the following steps:

[0040] S100: collecting images of the area where the long-distance pipeline to be tested is located;

[0041] S200: Constructing a long-distance pipeline exposure detection model;

[0042] S300: training the long-distance pipeline exposure detection model;

[0043] S400: Inputting the collected image of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model, and performing feature extraction on the image to detect whether the long-distance pipeline to be tested in the image is exposed.

[0044] The above embodiments constitute a complete technical solution of the present disclosure. This embodiment can quickly and accurately identify whether a long-distance pipeline is exposed in its area by constructing a long-distance pipeline exposure detection model, thereby improving the safety and reliability of the long-distance pipeline.

[0045] In another embodiment, the long-distance pipeline exposure detection model adopts an improved YOLOX-S network.

[0046] In this embodiment, the improved YOLOX-S network includes a backbone network (Backbone), an improved feature extraction and fusion module (Neck) and an improved prediction output module (Prediction).

[0047] The backbone network includes the following layers from top to bottom:

[0048] Inputs (640×640×3) (width×height×number of channels);

[0049] Focus (320×320×12);

[0050] Conv2D_BN_Silu(160×160×128);

[0051] CspLayer(160×160×128);

[0052] Conv2D_BN_Silu(80×80×256);

[0053] CspLayer (80×80×256);

[0054] Conv2D_BN_Silu(40×40×512);

[0055] CspLayer (40×40×512);

[0056] Conv2D_BN_Silu(20×20×1024);

[0057] SPPBottleneck(20×20×1024);

[0058] CspLayer (20×20×1024).

[0059] The existing feature extraction fusion module includes two branches, wherein the first branch includes the following layers connected in sequence from bottom to top:

[0060] Conv2D_BN_Silu layer;

[0061] UpSampling2D layer;

[0062] Concat+CSPLayer layer;

[0063] Conv2D_BN_Silu layer;

[0064] UpSampling2D layer;

[0065] Concat+CSPLayer layer.

[0066] The second branch includes the following connected from top to bottom:

[0067] CBS_2 layer;

[0068] Concat+CSPLayer layer;

[0069] CBS_2 layer;

[0070] Concat+CSPLayer layer.

[0071] As an improvement to the existing YOLOX-S network, the present disclosure introduces a Ghost module in the feature extraction fusion module to replace the CBS_2 module (composed of ordinary convolution (Conv2D), batch normalization (BatchNorm) and Silu activation function) in the feature extraction fusion module of the existing YOLOX-S network as shown above, that is, Figure 2 As shown in the figure, the improved feature extraction fusion module replaces the CBS_2 layer in front of the first Concat+CSPLayer layer and the CBS_2 layer in front of the second Concat+CSPLayer layer of the second branch of the existing feature extraction fusion module with the Ghost module. Compared with the CBS_2 layer, the Ghost module can expand the image features by using linear operations and increase the number of channels. It can not only generate more feature maps by using fewer parameters, but also the total number of parameters and floating point numbers required are lower than ordinary convolution without changing the size of the feature map. In addition, compared with the CBS_2 layer, the Ghost module can reduce the computational cost of the convolution layer while maintaining similar recognition performance, thereby improving the computing power of the model.

[0072] Figure 3 This is a schematic diagram of the structure of the Ghost module. Figure 3 In the example, the input features are first passed through a 1×1 convolution module consisting of a 1×1 convolution, a batch normalization layer BN, and a relu activation function (i.e. Figure 3 The first downward arrow in the figure) performs a convolution operation to reduce the number of channels of the input feature, and then uses a cheap operation ( It is a cheap operation that uses deep separable convolution (DW Convolution) to generate similar feature maps. Its purpose is to retain effective information in a more cost-effective way. Finally, the feature map (Intrinsic) obtained by the identity transformation (Identity) of the feature map output by the convolution operation of the 1×1 convolution module is concatenated with the feature map (Ghost) generated by the cheap operation to generate the final feature map.

[0073] In actual operation, let the size of the input feature map be , the size of the output feature map is , the size of the common convolution kernel is , the depth convolution kernel size is , then the number of parameters of ordinary convolution and the amount of calculation They are:

[0074]

[0075]

[0076] In the Ghost module, the input feature map is divided into s parts, so the parameter and the amount of calculation They are:

[0077]

[0078] F g = n s × ( h y × w y ) × [ c × k × k + ( s − 1 ) × d × d ]

[0079] The number of parameters of Ghost module and ordinary convolution operation R p and the amount of calculation R f The ratios are:

[0080]

[0081]

[0082] It can be seen from the above formula that the use of the Ghost module can reduce the amount of calculation and the number of parameters by s times compared to ordinary convolution.

[0083] In addition, since the use of the Ghost module to replace the CBS_2 module in the feature extraction fusion module will still affect the network's ability to extract key features to a certain extent, in order to reduce its negative impact, as a further improvement to the existing YOLOX-S network, the present disclosure also introduces an attention mechanism module CA (such as Figure 2 As shown in FIG. 1 , a total of three attention mechanism modules CA are inserted into the feature extraction fusion module, wherein the first CA is located between the first Concat+CSPLayer layer and the second Conv2D_BN_Silu layer from bottom to top in the first branch; the second CA is located behind the last Concat+CSPLayer layer in the first branch and is connected to the first Ghost module; the third CA is located between the first Concat+CSPLayer layer and the second Ghost module from top to bottom in the second branch; the fourth CA is located behind the second Concat+CSPLayer layer in the second branch). Since feature mixing is easily caused in the process of fusion of features at all levels, embedding the attention mechanism module CA into the corresponding positions described above will bring the following advantages: by adding the attention mechanism module CA at the above-mentioned positions, it is possible to more effectively enhance the attention to features at all levels, screen out the key areas of fusion features, and remove background noise therein.

[0084] At present, the more mainstream channel attention mechanism (Squeeze and Excite, SE) only considers the internal channel information and ignores the importance of position information, which can only slightly improve the performance of the network. The attention mechanism module CA introduced in this embodiment can not only capture information across channels, but also capture direction perception and position perception information, which enables the model disclosed in this disclosure to more accurately locate and identify the target area of ​​the long-distance pipeline. The main idea of ​​the attention mechanism module CA is to encode channel relationships and long-term dependencies through precise position information, which is specifically divided into two steps: coordinate information embedding and coordinate attention generation:

[0085] First, in the coordinate information embedding step, the size of the input feature map is set to , where C represents the number of channels, H represents the height, and W represents the width. The input feature map is reused after the residual connection. and The pooling kernel performs average pooling X Avg Pool and Y Avg Pool along the horizontal coordinate direction and the vertical coordinate direction respectively to encode each channel, and then the two transformations are feature aggregated along the two spatial directions to obtain a pair of sizes and Perceptual attention feature map and .

[0086] Secondly, in the coordinate attention generation step, the Concat function is used to concatenate the two feature maps generated by the coordinate information embedding module. and , then use 1×1 convolution (Conv2D) to perform channel transformation on the two feature maps, as shown below:

[0087]

[0088] in, represents 1×1 convolution, BN represents batch normalization (BatchNorm), δ represents non-linear activation function (Non-linear), z represents the input feature map, and the final result is is the intermediate feature map, H represents the height of the intermediate feature map, W represents the width of the intermediate feature map, C represents the number of channels of the intermediate feature map, which contains spatial information in the horizontal and vertical directions, and r represents the downsampling ratio. The intermediate feature map f is normalized (BN) and nonlinearly processed, and then split along the spatial dimension to generate a tensor and , and then use 1×1 convolution to transform the feature map and They are transformed to the same number of channels as the input feature map, and then activated using the sigmoid function to obtain the horizontal and vertical directions used for the input feature map respectively. and Two attention weights, finally and The output is multiplied by the input feature map connected by the residual connection to obtain the final output feature map (Re-weight), which is expressed as:

[0089]

[0090] in, is a feature of the input feature map with a height of i and a width of j in channel c. It is the output result obtained by the attention module.

[0091] The improved prediction output module is as follows Figure 5As shown, it includes the Main-Head module and the additionally introduced AuxHead module. The Main-Head module has the same structure as the AuxHead module, and both include the Yolo-head (Large) layer, Yolo-head (Medium) layer, and Yolo-head (Small) layer connected in sequence. During the training phase, feature maps of different sizes will pass through the Yolo-Head (Large) layer, Yolo-Head (Medium) layer, and Yolo-Head (Small) layer of the MainHead and AuxHead respectively. Large, Medium, and Small are only used to distinguish the feature maps of different sizes processed by Yolo-head. The internal structure of each Yolo-head is the same, and both include a Conv2D_BN_Silu layer. After the Conv2D_BN_Silu layer, two superimposed Ghost modules are connected in parallel. After each superimposed Ghost module, Cls, Reg, and Obj are connected respectively. The feature information obtained by the three Yolo-heads is connected through Concat, and the dimension of the feature information is changed to (5+class_num (number of categories), 8400) through the Tranpose function; the feature information output by MainHead and AuxHead is calculated after SimOTA (label assignment) to calculate their respective loss_iou (iou loss of the predicted box and the real box), loss_obj (confidence loss of the predicted box) and loss_cls (classification loss), and finally the losses of MainHead and AuxHead are added to obtain the total loss loss_total and the superimposed loss_iou, loss_obj and loss_cls. In the inference stage, the feature map does not pass through AuxHead, and the output result is directly obtained by MainHead.

[0092] The AuxHead module is a technique used in deep learning models to provide additional auxiliary training signals and feature extraction. The AuxHead module usually adds an auxiliary classification head in the middle layer of the network to assist in the training and optimization of the network.

[0093] The AuxHead module creates one or more additional branches in the middle layers of a deep neural network, each with a classification head (usually a fully connected layer or a convolutional layer). These additional classification heads work in parallel with the main classification head (usually in the last layer of the network). Auxiliary heads are able to provide additional supervision signals to help the network learn better during training. Auxiliary heads usually classify the features of the middle layers, so that the network can obtain more comprehensive supervision signals at different levels. This can promote better propagation of gradients, alleviate the problem of gradient vanishing, and help the network learn feature representations better.

[0094] During training, the loss function of the auxiliary head is combined with the loss function of the main classification head as the overall loss function of the network. In this way, the network can be trained by both the main classification task and the auxiliary classification task at the same time, thereby improving the overall performance.

[0095] By introducing the AuxHead module into the prediction output module, the model can solve problems such as gradient vanishing, overfitting, and instability in model training. Without affecting the inference speed, the model can improve the detection accuracy of long-distance pipeline exposure.

[0096] In summary, by introducing the Ghost module and the attention mechanism module CA into the feature extraction module, and introducing the AuxHead module into the prediction output module, we can overcome the problems of limited image feature extraction ability and high computational cost caused by the use of ordinary convolution modules in the Neck module and Prediction module in the traditional YOLOX-S network.

[0097] Below, the present disclosure specifically describes the detection effect of the improved YOLOX-S network on the image of the area where the long-distance pipeline is located in combination with Table 1. Table 1 is as follows:

[0098] Table 1

[0099]

[0100] In Table 1, Ghost means replacing the CBS_2 module in the Neck and Prediction modules with the Ghost module, CA means adding the CA attention mechanism in the Neck module, and AuxHead means referencing the AuxHead module in the Prediction module. As shown in Table 1, compared with the original YOLOX-S network, the mAP@0.5 and mAP@0.5:0.95 of the improved YOLOX-S network are increased by 4.6% and 2.2% respectively, while the number of parameters and model size are reduced by 2.3M and 8.7MB respectively. In addition, for the YOLOX-S+Ghost model, if only the CA attention mechanism is added, mAP@0.5 and mAP@0.5:0.95 will increase by 0.9% and 1.8% respectively; if only the AuxHead module is referenced, the mAP@0.5 and mAP@0.5:0.95 will increase by 3.1% and 4.3% respectively; if both the CA attention mechanism and the AuxHead module are used, the mAP@0.5 and mAP@0.5:0.95 will increase by 5.9% and 4.6% respectively. These two improvements have no significant effect on the number of model parameters and the size of the model. The above results show that the last improvement scheme is the most effective.

[0101] Furthermore, the present disclosure compares the detection results of the improved YOLOX-S network with those of other mainstream algorithms, as shown in Table 2:

[0102] Table 2

[0103]

[0104] As shown in Table 2, compared with the mainstream lightweight algorithms, the improved YOLOX-S network is lighter and has better accuracy. Compared with YOLOv5-S, the mAP@0.5 of the improved YOLOX-S network is improved by 1.5%, and the number of parameters and model size are reduced by 7.8% and 7.6%, respectively. In addition, compared with YOLOv4-Tiny and YOLOX-Tiny, the mAP@0.5 of the improved YOLOX-Tiny is 9.7% and 2.5% higher than them, respectively. At the same time, the number of parameters, the amount of calculation and the model size are only 3.75M, 4.01Gflops and 14.5MB, respectively, which are much lower than YOLOv4-Tiny and YOLOX-Tiny. Therefore, based on various indicators, the improved YOLOX does have better performance, occupies less storage resources and computing resources, and is better than the current mainstream lightweight algorithms.

[0105] In another embodiment, the training process of the long-distance pipeline exposure detection model includes the following steps:

[0106] S201: Construct an image dataset of long-distance pipelines with exposure;

[0107] S202: screening the image data set, that is, removing images where the pipeline is not exposed, retaining only images where the pipeline is exposed, and marking the exposed portion of the pipeline, specifically, manually marking features such as scratches, dents, corrosion, and cracks on the exposed portion, and the marked images form a new image data set;

[0108] S203: Divide the labeled image dataset into a training set and a test set;

[0109] S204: the number of training rounds is set to 300, the batch size is set to 8, the momentum is set to 0.9, the initial learning rate is set to 1.5e-4, and the model is trained using the training set. When the maximum number of training rounds is reached (for example, the maximum number of training rounds is set to 500), the model training is completed;

[0110] S205: Use the test set to test the trained model. During the test, when the average precision reaches 0.95 or above (the average precision mAP is an important indicator for evaluating the performance of the target detection algorithm, and the higher the mAP, the better the detection effect), the test passes and the trained model is obtained; otherwise, the training parameters are adjusted and retrained (for example, increase the number of data set samples or increase the number of training rounds or modify the learning rate).

[0111] In this embodiment, under normal circumstances, the pipeline is generally buried underground or in water. If the pipeline is exposed to the air, the surface of the pipeline may be damaged due to external factors, such as scratches, dents, corrosion, cracks and other defects. The pipeline image including these defects is used as input to train the model, so that the model can learn that once these defects are detected from the image to be tested containing the pipeline, it can be determined that the pipeline has been exposed, thereby realizing the detection of whether the pipeline is exposed.

[0112] In another embodiment, the present disclosure further provides a long-distance pipeline exposure detection device, the device comprising:

[0113] An acquisition module is used to acquire images of the area where the long-distance pipeline to be tested is located;

[0114] A model building and training module, used to build a long-distance pipeline exposure detection model and train the model; wherein the long-distance pipeline exposure detection model adopts an improved YOLOX-S network, and the improved YOLOX-S network includes a backbone network, an improved feature extraction and fusion module, and an improved prediction and output module; the improved feature extraction and fusion module additionally introduces a Ghost module and an attention mechanism module, and the improved prediction and output module additionally introduces an AuxHead module;

[0115] The detection module is used to input the collected image of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model to detect whether the long-distance pipeline to be tested in the image is exposed.

[0116] In another embodiment, the model building and training module includes:

[0117] The dataset construction submodule is used to construct a dataset of images of long-distance pipelines with exposure;

[0118] The screening and annotation submodule is used to screen and annotate the image dataset;

[0119] A partitioning submodule is used to divide the labeled image dataset into a training set and a test set;

[0120] The training submodule is used to train the model using the training set. When the maximum number of training rounds is reached, the model training is completed;

[0121] The test submodule is used to test the trained model using the test set.

[0122] In another embodiment, the present disclosure further provides an electronic device, including:

[0123] A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein:

[0124] When the processor executes the program, any of the above methods is implemented.

[0125] In another embodiment, the present disclosure further provides a computer storage medium storing computer executable instructions for executing any of the above methods.

[0126] Although the embodiments of the present invention are described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields, wherein the verification object is not limited to a specific sensor arrangement angle or a split blade disk structure, and the above-mentioned specific embodiments are only illustrative and instructive, and not restrictive. Under the guidance of this specification and without departing from the scope of protection of the claims of the present invention, a person of ordinary skill in the art can also make many forms, which all belong to the scope of protection of the present invention.

Claims

1. A long-distance pipeline exposure detection method, characterized in that: The method comprises the following steps: S100: collecting images of the area where the long-distance pipeline to be tested is located; S200: Build a long-distance pipeline exposure detection model and train the model; The long-distance pipeline exposure detection model adopts an improved YOLOX-S network, which includes a backbone network, an improved feature extraction and fusion module, and an improved prediction and output module; the improved feature extraction and fusion module additionally introduces a Ghost module and an attention mechanism module, and the improved prediction and output module additionally introduces an AuxHead module; The backbone network includes, from top to bottom: Inputs 640×640×3; Focus 320×320×12; Conv2D_BN_Silu 160×160×128; CspLayer 160×160×128; Conv2D_BN_Silu 80×80×256; CspLayer 80×80×256; Conv2D_BN_Silu 40×40×512; CspLayer 40×40×512; Conv2D_BN_Silu 20×20×1024; SPBottleneck 20×20×1024; CspLayer20×20×1024; The improved feature extraction fusion module includes two branches, where the first branch includes the following layers connected from bottom to top: Conv2D_BN_Silu layer; UpSampling2D layer; Concat+CSPLayer layer; Attention mechanism module CA; Conv2D_BN_Silu layer; UpSampling2D layer; Concat+CSPLayer layer; Attention mechanism module CA; The second branch includes the following connected from top to bottom: Ghost module; Concat+CSPLayer layer; Attention mechanism module CA; Ghost module; Concat+CSPLayer layer; Attention mechanism module CA; The training process of the long-distance pipeline exposure detection model includes the following steps: S201: Construct an image dataset of long-distance pipelines with exposure; S202: screening the image data set, that is, removing images where the pipeline is not exposed, retaining only images where the pipeline is exposed, and marking the exposed portion of the pipeline, specifically, manually marking features such as scratches, dents, corrosion, and cracks on the exposed portion, and the marked images form a new image data set; S203: Divide the labeled image dataset into a training set and a test set; S204: The number of training rounds is set to 300, the batch size is set to 8, the momentum is set to 0.9, the initial learning rate is set to 1.5e-4, and the model is trained using the training set. When the maximum number of training rounds is reached, the model training is completed; S205: Testing the trained model using the test set. During the test, if the average precision reaches 0.95 or above, the test is passed and the trained model is obtained; otherwise, the training parameters are adjusted and retrained. S300: Inputting the collected image of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model to detect whether the long-distance pipeline to be tested in the image is exposed.

2. A long-distance pipeline exposure detection device for implementing the method as claimed in claim 1, characterized in that: The device comprises: An acquisition module is used to acquire images of the area where the long-distance pipeline to be tested is located; A model building and training module is used to build a long-distance pipeline exposure detection model and train the model; wherein the long-distance pipeline exposure detection model adopts an improved YOLOX-S network, and the improved YOLOX-S network includes a backbone network, an improved feature extraction and fusion module, and an improved prediction and output module; the improved feature extraction and fusion module additionally introduces a Ghost module and an attention mechanism module, and the improved prediction and output module additionally introduces an AuxHead module; the backbone network includes, from top to bottom: Inputs; Focus; Conv2D_BN_Silu; CspLayer; Conv2D_BN_Silu; CspLayer; Conv2D_BN_Silu; CspLayer; Conv2D_BN_Silu; SPPBottleneck; CspLayer; The improved feature extraction fusion module includes two branches, where the first branch includes the following layers connected from bottom to top: Conv2D_BN_Silu layer; UpSampling2D layer; Concat+CSPLayer layer; Attention mechanism module CA; Conv2D_BN_Silu layer; UpSampling2D layer; Concat+CSPLayer layer; Attention mechanism module CA; The second branch includes the following connected from top to bottom: Ghost module; Concat+CSPLayer layer; Attention mechanism module CA; Ghost module; Concat+CSPLayer layer; Attention mechanism module CA; A detection module is used to input the collected image of the area where the long-distance pipeline to be tested is located into the trained long-distance pipeline exposure detection model to detect whether the long-distance pipeline to be tested in the image is exposed; The model building and training module includes: The dataset construction submodule is used to construct a dataset of images of long-distance pipelines with exposure; The screening and annotation submodule is used to screen and annotate the image dataset; A partitioning submodule is used to divide the labeled image dataset into a training set and a test set; The training submodule is used to train the model using the training set. When the maximum number of training rounds is reached, the model training is completed; The test submodule is used to test the trained model using the test set.

3. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to claim 1 is implemented.

4. A computer storage medium, characterized in that: The computer storage medium stores computer executable instructions for executing the method of claim 1 .

Images