Building communication optical cable fault image processing method

Abstract

According to the building communication optical cable fault image processing method provided by the invention, on the basis of a U-Net model, by introducing a multi-scale aggregation pooling module MASP and a feature channel attention mechanism FCM, efficient capture of optical cable fault subtle differences and effective fusion of multi-scale features are realized, and the detection precision and efficiency are remarkably improved.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to a method for processing building communication optical cable fault images. Background Art

[0002] With the surge in demand for data transmission in modern society, communication optical cables have become an indispensable information transmission infrastructure. However, physical damage or environmental impacts on communication optical cable lines often lead to a decline in signal transmission quality or even interruption, seriously affecting the reliability of communication services. Traditional fault detection methods such as manual inspections are time-consuming and labor-intensive and cannot meet the requirements of rapid response. Therefore, efficient and accurate automated detection technology is urgently needed to ensure the stable operation of the network.

[0003] Although deep learning technology has made remarkable achievements in the field of image recognition, it still faces many challenges in processing communication optical cable fault images. For example, communication optical cable fault images usually contain complex textures and tiny defects, which makes it difficult for conventional convolutional neural networks to effectively extract fault features. In addition, existing models generally have problems such as high training difficulty and high overfitting risk when dealing with such tasks. Therefore, exploring new network architectures and algorithm optimization strategies is of great significance to improving fault detection accuracy.

[0004] As a classic image segmentation model, U-Net performs well in fields such as medical imaging. However, due to the unique complexity and diversity of communication optical cable fault images, the standard U-Net architecture may not be able to achieve ideal detection effects. To this end, the present invention proposes an improved model based on U-Net. By designing a novel downsampling module and a channel attention mechanism, the model's ability to learn subtle fault features is enhanced, thereby solving the problems of low recognition rate and poor generalization ability that are common in the prior art, and providing more reliable technical support for communication optical cable fault detection. Summary of the invention

[0005] The present invention provides a method for processing building communication optical cable fault images, which aims to achieve efficient capture of subtle differences in optical cable faults and effective fusion of multi-scale features based on a U-Net model by introducing a multi-scale aggregation pooling module MASP and a feature channel attention mechanism FCM, thereby helping the model to significantly improve the detection accuracy and efficiency of building communication optical cable fault images.

[0006] The present invention aims to propose a network model for identifying and segmenting building communication optical cable fault images, and provides a method for processing building communication optical cable fault images, comprising the following steps: S1. The data set of building communication optical cable fault images is prepared. A high-definition camera is used to shoot faulty communication optical cables in the production environment and the obtained images are sorted to obtain communication optical cable fault images. The images of the three most common situations are annotated, including images of three types of faults: surface wear, cracks, and corrosion. The fault locations in the images are annotated. After the annotation is completed, a communication optical cable fault image data set is obtained; S2. Build a feature channel attention module FCM, which includes 1×1 convolution and Softmax calculation operations; S3, build a multi-scale aggregation pooling module MASP, including 3×3 convolution, maximum pooling and sigmoid calculation operations; S4. Build an encoder module, including 3×3 convolution, multi-scale aggregation pooling module and channel feature module; S5. Build a decoder module, including 3×3 convolution, upsampling and skip connection; S6, constructing a communication optical cable fault identification and segmentation network model, which is composed of input, encoder, decoder, activation function and output in sequence; S7, training a communication optical cable recognition and segmentation network model, using a communication optical cable fault image dataset to train a communication optical cable recognition and segmentation network model; S8, communication optical cable fault image processing, obtaining the communication optical cable fault image data that needs to be identified and segmented, and inputting it into the trained communication optical cable fault identification and segmentation network to obtain the communication optical cable fault image after identification and segmentation.

[0007] Preferably, in step S2, for the feature channel attention module FCM, the input image is first input with a feature map of ,in is the number of channels, is the height, For width, first feature Through a Convolution operation to compress spatial information: ,in represents the convolution kernel size, represents the step size, Indicates the padding size, after The feature map after convolution is ,in is the number of channels after convolution, is the height, is the width, and then the feature map Flatten to In the form of a two-dimensional matrix, the features are application Function gets the weight matrix , for the features Reshape to original shape , for the original feature input And the generated weight matrix Perform element-by-element multiplication to obtain the weighted feature map , through a Convolution operation to adjust the channel dimension to obtain features , the final output features are added to the original input Get the output of the FCM module .

[0008] Preferably, for the constructed feature channel attention module FCM, through a series of carefully designed operations, the convolution operation can effectively reduce the computational complexity and memory consumption of subsequent operations by reducing the number of channels, making the model more efficient. The function can normalize the channel response at each position into a probability distribution, emphasizing important channel information and suppressing unimportant channel information. The residual connection helps to stably transfer the gradient, alleviate the gradient disappearance or explosion problem in deep networks, retain the original input information, avoid information loss, and also incorporate the feature information enhanced by the attention mechanism, which helps to improve the performance of the model.

[0009] Preferably, in step S3, for the multi-scale aggregation pooling module MASP, the input feature First, through a basic convolution unit, the basic convolution unit includes a convolution, and a Activation function, get features ,in represents the activation function, represents the normalization operation, Represented as a convolution operation, then the feature After four consecutive max pooling Get Features , for each downsampling result , with the characteristics Splice to get the final feature ,feature After passing through the basic convolution unit again, the final output is obtained , .

[0010] Preferably, the multi-scale aggregation pooling module MASP uses convolutional basic units to achieve more effective feature extraction, improve gradient propagation and accelerate the training process, introduce nonlinearity to enhance the expressiveness of the model, improve the generalization and robustness of the model, and use the splicing of feature maps of different scales to capture details and contextual information at different levels, so that the model can simultaneously utilize local and global features to improve the richness and diversity of features. Feature maps of different scales provide complementary information, which helps the model to better understand and process input data. Splicing these features can enable the model to learn more diverse feature representations and enhance the expressiveness of the model. Through the fusion of multi-scale features, the weights of key features can be enhanced, noise and unimportant features can be suppressed, and the robustness of the model can be improved.

[0011] Preferably, in step S4, for the encoder module, the input feature First, a convolution operation is performed to obtain the features ,in represents the convolution kernel size, Expressed as the channel ratio, then, the characteristic The features are obtained through the multi-scale aggregation pooling module MSAP The whole process is repeated four times, and the size of the feature map is gradually reduced after each operation. , , , , the size of the final feature map becomes the original image , the number of channels becomes 16 times the original.

[0012] Preferably, the encoder adopts convolution and multi-scale aggregation pooling MSAP stacking, which effectively improves gradient propagation, reduces the problems of gradient disappearance and explosion, and accelerates the training process of the model. Multi-scale feature extraction enables the model to capture details and contextual information of different scales in the image. The expressive ability of the model is significantly enhanced, and more complex feature representations can be learned. The fusion of multi-scale features enhances the weight of key features, suppresses noise and unimportant features, and further improves the effect of image enhancement.

[0013] Preferably, in step S5, the decoder module performs channel attention weighting on the output feature of the encoder after the feature channel module performs channel attention weighting. To restore the features, first Perform one upsampling and two convolution operations to reduce the number of channels and restore to the original image. , get the features ,in Represented as a convolution operation, Represented as a downsampling operation, is the convolution kernel size, is the number of channels of the output feature, and the generated features And the feature map of the corresponding encoder Make a skip connection , after the same four decoding operations, we finally get , restore to its initial size, and then pass The activation function obtains the fault segmentation result.

[0014] Preferably, the low-level feature map of the encoder is spliced ​​with the high-level feature map of the decoder through jump links, which can retain more low-level detail information and help restore the fine structure and edges of the image. Secondly, the fusion of multi-scale features enables the model to capture information at different scales and improve the richness and diversity of features. In addition, the combination of convolutional layers and upsampling layers improves gradient propagation and reduces the gradient vanishing problem.

[0015] Preferably, in step S6, for the building communication optical cable fault identification and segmentation network model, the input feature First, it is processed through an encoder to obtain features with gradually reduced size. , , , , These features are then further processed by the feature channel attention module FCM to generate feature maps after feature channel weighted enhancement , the feature size remains unchanged after weighting; in the decoder stage, the generated features The features generated by the decoder are The features of the decoder are fused with the features of the encoder at the same layer through 3×3 convolutions through jump connections in the channel dimension. The features in the encoder are used to guide the restoration of the features in the decoder to generate , and obtained by upsampling 4 times , , , , , so that the features gradually return to their original size, and finally The activation function obtains the fault segmentation result.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: the invention proposes a building communication optical cable fault image recognition and segmentation model, which is inspired by the classic U-Net model, but has made a number of key improvements to improve detection accuracy and efficiency. First, the network adopts a unique downsampling module, which can maximize the retention of important feature information while reducing the image resolution, which is crucial for capturing subtle differences in optical cable faults. Secondly, the output of the encoder is processed by an innovative feature channel attention module FCM, which can intelligently identify and enhance feature channels useful for fault detection, thereby improving the sensitivity and specificity of the model. Finally, the channel attention weights adjusted by the convolution operation are applied to each jump connection to achieve effective fusion of multi-scale features, ensuring multi-level information transmission from coarse to fine, and ultimately improving the segmentation accuracy of the fault area. This series of technological innovations gives the invention a significant advantage in the field of building communication optical cable fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of building communication optical cable fault image processing provided by the present invention.

[0018] Figure 2 It is a structural diagram of the feature channel attention module FCM provided by the present invention.

[0019] Figure 3 This is a structural diagram of the multi-scale aggregation pooling MSAP provided by the present invention.

[0020] Figure 4 It is a network structure diagram of building communication optical cable fault image processing provided by the present invention.

[0021] Figure 5 It is the overall structural diagram provided by the present invention.

[0022] Figure 6 This is a building communication cable fault segmentation diagram provided by the present invention. DETAILED DESCRIPTION

[0023] The present invention aims to propose a model for image recognition and segmentation of building communication optical cable faults. The model is based on the U-Net architecture, but is optimized in many aspects to improve the accuracy and efficiency of detection. Specifically, this model introduces an innovative multi-scale aggregation pooling module MASP, which not only effectively reduces the resolution of the image, but also preserves key feature information to the maximum extent, which is particularly important for identifying subtle changes in optical cable faults. In addition, the model utilizes a novel feature channel attention mechanism FCM, which can automatically identify and strengthen feature channels that are helpful for fault detection, thereby enhancing the detection ability and accuracy of the model. Finally, the feature channel attention weights adjusted by the convolutional layer are integrated into each jump connection, which promotes the effective combination of features of different scales and ensures that the information flow from macro to micro is unimpeded, thereby greatly improving the recognition accuracy of the fault area. The technical solution in the embodiment of the present invention will be clearly and completely described below, specifically including the following steps, such as Figure 1 shown.

[0024] S1. Use high-definition cameras in the production environment to collect image data of faulty building communication optical cables, including images of three types of faults: surface wear of the outer protective layer of the optical cable due to mechanical friction and improper handling, cracks and cracks caused by aging, ambient temperature changes and excessive bending, and corrosion of metal conductors and protective layers in humid environments. The number of various fault types in the collected images is 1,000, which are evenly distributed and sufficient in number. The images are clear and the characteristics of various types of faults are obvious. For the collected communication optical cable fault image data, the image annotation program Labelme is used to annotate the images, and the segmentation boundary box of the fault area is drawn for the faulty part in the image. The part within the segmentation boundary box is used as the target area for model recognition in the image, that is, the fault position and shape identified by segmentation.

[0025] S2. Construct a feature channel attention module FCM, which includes two 1×1 convolution operations and one Softmax calculation operation.

[0026] Furthermore, a feature channel attention module FCM is constructed, whose structure is as follows Figure 2 As shown, for the input feature map ,in is the number of channels, is the height, For width, first feature Through a Convolution operation to compress spatial information: , where the convolution kernel size is , the step length is , filled with ,go through The feature map after convolution is ,in is the number of channels after convolution, is the height, is the width, and then the feature map Flatten to In the form of a two-dimensional matrix, the features are application Function gets the weight matrix , for the features Reshape to original shape , for the original feature input And the generated weight matrix Perform element-by-element multiplication to obtain the weighted feature map , through a Convolution operation is used to adjust the channel dimension to obtain features , the final output feature Add to original input Get the output of the FCM module .

[0027] S3. Construct a multi-scale aggregation pooling module MASP, which includes two basic convolution unit operations, four maximum pooling operations and one aggregation operation. Each basic convolution unit contains a 3×3 convolution operation, a BatchNormal operation and a Sigmoid activation function.

[0028] Furthermore, a multi-scale aggregation pooling module MASP is constructed, and its structure is as follows: Figure 3 As shown, the input features First, through a basic convolution unit, the convolution unit includes a Convolution operation, a Operation and a Activation function, get features , then the feature After four consecutive max pooling Get Features , for each downsampling result , , , With features Concatenate in the channel dimension to get the final feature , feature After passing through the basic convolution unit again, the final output is obtained , .

[0029] S4. Construct an encoder module consisting of eight 3×3 convolution operations, four multi-scale aggregation pooling modules MASP, and one feature channel attention module FCM.

[0030] Further, in step S4, for the encoder module, its structure is as follows Figure 4 As shown, the input features First, a convolution operation is performed to obtain the features , convolution kernel size for , channel ratio for , then, the features The features are obtained through the multi-scale aggregation pooling module MSAP The whole process is repeated four times, and the size of the feature map is gradually reduced after each operation. , , , , the size of the final feature map becomes the original image , the number of channels becomes 16 times the original.

[0031] S5. Construct a decoder module consisting of seven 3×3 convolution operations, four upsampling operations, and three skip connections.

[0032] Further, in step S5, for the decoder module, its structure is as follows Figure 4 As shown in the figure, the decoder module is the output feature of the encoder after the feature channel module performs channel attention weighting To restore the features, first Perform one upsampling and two convolution operations to reduce the number of channels and restore to the original image. , get the features , where is the convolution kernel size , is the number of channels of the output feature, , the generated features And the feature map of the corresponding encoder Make a skip connection , , after two convolutions, we get , after the same four decoding operations, we finally get .

[0033] S6. Construct a building communication optical cable fault identification and segmentation network model, which is composed of input, encoder, decoder, output and a Sigmoid activation function in sequence.

[0034] Further, in step S6, for the building communication optical cable fault identification and segmentation network model, its structure is as follows: Figure 5 As shown, the input features First, it is processed through an encoder to obtain features with gradually reduced size. , , , , These features are then further processed by the feature channel attention module FCM to generate feature maps after feature channel weighted enhancement , the feature size remains unchanged after weighting; in the decoder stage, the generated features The features generated by the decoder are The features of the decoder are fused with the features of the encoder at the same layer through 3×3 convolutions through jump connections in the channel dimension. The features in the encoder are used to guide the restoration of the features in the decoder to generate , and obtained by upsampling 4 times , , , , , so that the features gradually return to their original size, and finally The activation function obtains the fault segmentation result.

[0035] S7. Train a communication optical cable fault identification and segmentation network model. Use a communication optical cable fault image dataset to train a communication optical cable fault identification and segmentation network model.

[0036] Furthermore, in step S7, the method uses Python3.8 language, PyTorch framework in the CUDA11 environment, and is trained on the NVIDIA RTX 3090 GPU. There are 1,000 images for each class in the dataset, and the training set and test set are divided into 4:1. Stochastic gradient descent SGD is selected as the optimizer during training. The weight decay is set to 0.0001, the learning rate is 0.001, the momentum is set to 0.9, the training rounds are 64, and the learning rate decay strategy is adopted. After every 10 training rounds, the learning rate is reduced to the original 0.1.

[0037] S8, communication optical cable fault image processing, obtaining the communication optical cable fault image data that needs to be identified and segmented, and inputting it into the trained communication optical cable fault identification and segmentation network to obtain the communication optical cable fault image after identification and segmentation.

[0038] Further, in step S8, a segmented communication optical cable fault image is generated, and its structure is as follows: Figure 6As shown, for obvious damaged areas on the surface, the damaged parts are outlined by lines, showing the specific location of irregular wear or peeling, which helps staff to further inspect and repair to ensure its normal function and prevent potential safety problems.

[0039] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for processing building communication optical cable fault images, characterized in that: The following steps are involved: S1. The data set of building communication optical cable fault images is prepared. A high-definition camera is used to shoot faulty communication optical cables in the production environment and the obtained images are sorted to obtain communication optical cable fault images. The images of the three most common situations are annotated, including images of three types of faults: surface wear, cracks, and corrosion. The fault locations in the images are annotated. After the annotation is completed, a communication optical cable fault image data set is obtained; S2. Build a feature channel attention module FCM, which includes 1×1 convolution and Softmax calculation operations; S3, build a multi-scale aggregation pooling module MASP, including 3×3 convolution, maximum pooling and sigmoid calculation operations; S4. Build an encoder module, including 3×3 convolution, multi-scale aggregation pooling module and channel feature module; S5. Build a decoder module, including 3×3 convolution, upsampling and skip connection; S6, constructing a communication optical cable fault identification and segmentation network model, which is composed of input, encoder, decoder, activation function and output in sequence; S7, training a communication optical cable recognition and segmentation network model, using a communication optical cable fault image dataset to train a communication optical cable recognition and segmentation network model; S8, communication optical cable fault image processing, obtaining the communication optical cable fault image data that needs to be identified and segmented, and inputting it into the trained communication optical cable fault identification and segmentation network to obtain the communication optical cable fault image after identification and segmentation.

2. A method for processing building communication optical cable fault images according to claim 1, characterized in that: In step S2, for the feature channel attention module FCM, the input image is first input with a feature map ,in is the number of channels, is the height, For width, first feature Through a Convolution operation to compress spatial information: ,in represents the convolution kernel size, represents the step size, Indicates the padding size, after The feature map after convolution is ,in is the number of channels after convolution, is the height, is the width, and then the feature map Flatten to In the form of a two-dimensional matrix, the features are application Function gets the weight matrix , for the features Reshape to original shape , for the original feature input And the generated weight matrix Perform element-by-element multiplication to obtain the weighted feature map , through a Convolution operation is used to adjust the channel dimension to obtain features , the final output features are added to the original input Get the output of the FCM module .

3. A method for processing building communication optical cable fault images according to claim 2, characterized in that: In step S3, for the multi-scale aggregation pooling module MASP, the input feature First, through a basic convolution unit, the basic convolution unit includes a convolution, and a Activation function, get features ,in represents the activation function, represents the normalization operation, Represented as a convolution operation, then the feature After four consecutive max pooling Get Features , for each downsampling result , with the characteristics Splice to get the final feature ,feature After passing through the basic convolution unit again, the final output is obtained , .

4. A method for processing building communication optical cable fault images according to claim 3, characterized in that: In step S4, for the encoder module, input feature First, a convolution operation is performed to obtain the features ,in represents the convolution kernel size, Expressed as the channel ratio, then, the characteristic The features are obtained through the multi-scale aggregation pooling module MSAP The whole process is repeated four times, and the size of the feature map is gradually reduced after each operation. , , , , the size of the final feature map becomes the original image , the number of channels becomes 16 times the original.

5. A method for processing building communication optical cable fault images according to claim 4, characterized in that: In step S5, the decoder module performs channel attention weighting on the output features of the encoder after the feature channel module performs channel attention weighting. To restore the features, first Perform one upsampling and two convolution operations to reduce the number of channels and restore to the original image. , get the features ,in Represented as a convolution operation, Represented as a downsampling operation, is the convolution kernel size, is the number of channels of the output feature, and the generated features And the feature map of the corresponding encoder Make a skip connection , after the same four decoding operations, we finally get , restore to its initial size, and then pass The activation function obtains the fault segmentation result.

6. A method for processing building communication optical cable fault images according to claim 5, characterized in that: In step S6, for the building communication optical cable fault identification and segmentation network model, input feature First, it is processed through an encoder to obtain features with gradually reduced size. , , , , These features are then further processed by the feature channel attention module FCM to generate feature maps after feature channel weighted enhancement , the size remains unchanged after feature weighting; In the decoder stage, the generated features The features generated by the decoder are The features of the decoder are fused with the features of the encoder at the same layer through 3×3 convolutions through jump connections in the channel dimension. The features in the encoder are used to guide the restoration of the features in the decoder to generate , and obtained by upsampling 4 times , , , , , so that the features gradually return to their original size, and finally The activation function obtains the fault segmentation result.

7. A method for processing building communication optical cable fault images according to claim 6, characterized in that: In the step S7, a high-definition camera in a production environment is used to collect image data of faulty building communication optical cables, including images of three types of faults: surface wear of the outer protective layer of the optical cable due to mechanical friction and improper handling, cracks and cracks caused by aging, ambient temperature changes and excessive bending, and corrosion of the metal conductor and protective layer in a humid environment. The number of various fault types in the collected images is 1,000, which are evenly distributed and sufficient in number. The images are clear and the characteristics of various types of faults are obvious. For the collected communication optical cable fault image data, the image annotation program Labelme is used to annotate the image, and a segmentation boundary box of the fault area is drawn for the faulty part in the image. The part within the segmentation boundary box is used as the target area for model recognition in the image, that is, the segmented and recognized fault position and shape.

8. A method for processing building communication optical cable fault images according to claim 7, characterized in that: In step S8, the communication optical cable fault image is processed to obtain a communication optical cable fault image after identification and segmentation. For the fault area with obvious surface, the damaged part is outlined by lines to show the specific location of the fault.

Images