A lightweight distributed fiber optic sensing image target detection method and system

Abstract

The present invention relates to the technical field of distributed fiber optic perception image target detection, and discloses a lightweight distributed fiber optic perception image target detection method and system. The method comprises the following steps: collecting distributed fiber optic perception images and corresponding annotations and preprocessing them to obtain a training set; building a target detection network based on an improved YOLOv8 model: the distributed fiber optic perception image target detection network comprises a backbone network, a lightweight multi-scale fusion module and a detection head; training the distributed fiber optic perception image target detection network through the training set until the distributed fiber optic perception image target detection network converges or reaches a set number of training times. The distributed fiber optic perception image used for testing is input into the trained target detection network. The present invention realizes fast feature extraction of the backbone network and efficient feature fusion of the neck through a customized module.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a lightweight distributed optical fiber sensing image target detection method and system. Background Art

[0002] Distributed fiber optic sensing technology is a technology that uses optical fiber as a sensor to monitor and locate changes in physical quantities in real time. It is based on the optical properties of optical fiber and the principle of signal transmission. It achieves distributed sensing and positioning of physical quantities such as temperature, stress, pressure, and deformation by introducing certain changes in the optical fiber or using external physical fields to affect the optical fiber.

[0003] The basic principle of this technology is to use the total internal reflection characteristics of optical fiber and the propagation mode of optical signals. The optical signal injected into the optical fiber will continue to propagate in the optical fiber. When the optical signal encounters changes in medium parameters or physical quantities, such as temperature changes causing slight changes in the length of the optical fiber, stress or pressure causing changes in the refractive index of the optical fiber, etc., it will affect the propagation characteristics of the optical signal, thereby forming a specific signal change pattern in the optical fiber. By analyzing and processing these changes, real-time monitoring and positioning of the distribution state of physical quantities can be achieved. The optical signals transmitted back from different monitoring points and different monitoring times in the entire optical fiber are described as a time-space signal matrix, and then the time-space matrix information is mapped into a two-dimensional picture according to a certain rule, forming a distributed optical fiber perception image. Each frame of the distributed optical fiber perception image contains the optical signal conditions of each monitoring point at a certain moment. By subsequently identifying and processing such images, the target of interest can be found.

[0004] However, identifying targets from distributed fiber optic sensing images has the following difficulties:

[0005] 1. Large background noise: These noises may come from environmental interference, stray signals of the optical fiber itself, etc., which may cause the target signal to be confused with the background noise, making it difficult to accurately extract the target information; 2. Complex signal patterns: The collected signal images may have complex signal patterns, which contain information on the changes of multiple physical quantities, making the recognition and extraction of target signals more difficult; 3. High target detection accuracy requirements: Recognizing targets in signal images usually requires high-precision detection and positioning, because the target may be a small change or an abnormal signal, requiring high-resolution and high-sensitivity processing of the signal image; 4. Uncertain target morphology: The morphology of the target in the distributed optical fiber sensing image may be uncertain due to environmental changes or the characteristics of the target itself, which increases the difficulty of target recognition and requires certain target morphology adaptability and robustness; 5. Large data volume: The signal images collected by distributed optical fiber sensing technology often have a large amount of data. To perform target recognition, efficient algorithms and processing technologies are required to process large-scale data and improve recognition efficiency and accuracy;

[0006] Taking the above difficulties into consideration, combining distributed fiber optic sensing technology and target detection technology for target recognition requires overcoming problems such as the complexity of signal images, the interference of background noise, and the uncertainty of target morphology. At the same time, it is necessary to design and optimize target detection algorithms suitable for distributed fiber optic sensing images to improve the accuracy and robustness of target recognition.

[0007] Object detection is one of the basic problems in the field of computer vision. It mainly refers to finding all the targets (objects) of interest in the input image and selecting the targets with appropriate detection rectangles. Since each type of target has various personalized instances (such as different types of dogs and cats) and image information differences (such as lighting, angle, occlusion, etc.), object detection has always been one of the most challenging problems in the field of computer vision. In addition, object detection is also an important part of many other downstream tasks such as instance segmentation, video content detection, and target tracking.

[0008] With the increasing demand for real-time performance and the development of deep learning technology, target detection algorithms have gradually evolved into methods based on deep learning. In 2016, the YOLO (You Only Look Once) algorithm was proposed. The YOLO algorithm divides the image into S×S grids and directly predicts the category probability and regresses the position information based on the bounding box corresponding to each grid. It is a popular algorithm with high efficiency and real-time performance. It has shown excellent performance in various scenarios. The YOLO model can be roughly divided into the backbone network, the neck fusion module, and the detection head. The backbone network is responsible for extracting the features of the input image using a multi-layer deep learning network, which is usually extracted into feature pyramids of different field sizes; the neck feature fusion module is mainly responsible for fusing different feature maps. The neck in the YOLO series usually uses the feature pyramid network (FPN) and its variants to fuse multi-level features; the detection head is used to classify, detect and locate the feature maps of multiple scales transmitted from the neck, including two structures without anchor boxes and with anchor boxes.

[0009] A crucial module in the YOLO model is the feature fusion module. Shallow features contain low-dimensional texture details and the locations of smaller objects. In contrast, deep features contain high-dimensional information and the locations of larger objects. Some researchers have proposed different feature fusion modules.

[0010] For example: Feature Pyramid Networks (FPN), FPN provides an efficient architecture design for fusing multi-scale features through cross-scale connections and information exchange; Path Aggregation Network (PANet) based on FPN adopts a bottom-up path to make the information fusion between different levels more sufficient; and repeatable module (BiFPN), it improves the efficiency of information fusion between different levels.

[0011] The centralized feature pyramid (CFP) method explores single feature information; in addition, some people extend FPN and propose an asymptotic feature pyramid network (AFPN) structure to enable interaction between non-adjacent layers.

[0012] However, the current information fusion method has some possible defects when applied to distributed fiber-optic sensing images: when cross-layer contrast and integration of information is required (such as fusion of level 1 and level 3), the traditional FPN-like structure cannot transmit information efficiently and quickly, and there is no mechanism for contrasting and segmenting information at different levels, which hinders the YOLO model from better information fusion. In order to alleviate the above problems, the present invention proposes a lightweight distributed fiber-optic sensing image target detection method and system. Summary of the invention

[0013] In order to solve the above technical problems, the present invention provides a lightweight distributed fiber optic sensing image target detection method and system, which can solve the shortcomings of existing feature fusion methods that cannot quickly and efficiently transmit cross-layer information, and enhance the filtering of different layers of information during feature fusion. It can take into account the differences between feature maps of different scales. For each feature layer that needs to be detected, it can remove deep high-dimensional semantic information and shallow low-dimensional texture information, thereby achieving better target detection effects.

[0014] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0015] A lightweight distributed fiber-optic sensing image target detection method is provided, wherein a distributed fiber-optic sensing image for testing is input into a trained target detection network, and a target detection result is output; a target detection network construction and training method includes the following steps:

[0016] Step 1: collect distributed fiber sensing images x and corresponding annotations and perform preprocessing to obtain a training set;

[0017] Step 2: Build a target detection network based on the improved YOLOv8 model:

[0018] The target detection network includes a backbone network, a lightweight multi-scale fusion module and a detection head; the backbone network of the target detection network is obtained by replacing the convolutional layer of the backbone network in the YOLOv8 model with the PConv module and replacing the C2f module of the backbone network in the YOLOv8 model with the PC3 module; the detection head of the target detection network adopts the detection head of the YOLOv8 model; the backbone network of the target detection network is used to extract feature maps B3, B4, and B5 of three different sizes of the distributed optical fiber perception image x; the lightweight multi-scale fusion module is used to fuse the feature maps B3, B4, and B5, and output fused feature maps D3, D4, and D5; the three fused feature maps are respectively input into the three detection heads of the target detection network to obtain the target detection results;

[0019] The lightweight multi-scale fusion module consists of two stages. In the first stage, the feature maps B3, B4, and B5 are fused into three intermediate feature maps P3, P4, and P5 through the upsampling module, the concatenation operation, the PC3 module, and the PConv module. In the second stage, the intermediate feature maps P3, P4, and P5 are taken as input, one branch compares and fuses P3, P4, and P5 with each other, and the other branch retains the original information and jumps to the output fused feature maps D3, D4, and D5.

[0020] The PConv module is a fast convolution module with partial convolution as the core operator;

[0021] The PC3 module includes two tributaries. The feature map input to the PC3 module passes through a PConv module in one tributary, and is sequentially input into a PConv module and n Bottleneck modules in the other tributary. The feature maps output by the two tributaries are concatenated and then passed through a PConv module to obtain the output of the PC3 module.

[0022] Step three: train the target detection network using the training set until the target detection network converges or reaches a set number of training times.

[0023] Furthermore, the PConv module consists of a standard partial convolution layer, batch normalization and SiLU activation function; the feature map input to the PConv module passes through the standard partial convolution layer in sequence, and then passes through the batch normalization operation and the SiLU activation function to obtain the feature map output by the PConv module.

[0024] Furthermore, the first stage of the lightweight multi-scale fusion module specifically includes:

[0025] After upsampling, feature map B5 is concatenated with feature map B4 to obtain feature map A1; A1 is input into the PC3 module and the upsampling module, and concatenated with feature map B3 to obtain feature map A2; A2 is input into the PC3 module to obtain the intermediate feature map P3;

[0026] The result obtained by inputting feature map A2 into PC3 module and PConv module in sequence is concatenated with the result obtained by inputting feature map A1 into PC3 module to obtain feature map A3; A3 is input into PC3 module to obtain intermediate feature map P4;

[0027] The feature map A3 is input into the PC3 module and the PConv module in sequence, and the result is concatenated with the feature map B5 to obtain the feature map A4; A4 is input into the PC3 module to obtain the intermediate feature map P5;

[0028] The upsampling module is used to increase the resolution of the input feature map; the concatenation operation is to superimpose the input feature maps of the same size according to the channel dimension and output a composite feature map; the PConv module is used to reduce the resolution of the input feature map and set learnable parameters; the PC3 module performs a learnable internal feature fusion on the composite features generated after the concatenation operation.

[0029] Furthermore, the second stage of the lightweight multi-scale fusion module specifically includes:

[0030] The original intermediate feature map P3 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F3 (P3, P4, P5) to generate the fused feature map D3;

[0031] The original intermediate feature map P4 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F4 (P3, P4, P5) to generate the fused feature map D4;

[0032] The original intermediate feature map P3 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F5 (P3, P4, P5), and the generated fused feature map is D5;

[0033] F3(P3, P4, P5) = P3-PConv(Upsample(P4))-2*PConv(Upsample(P5)), which means P3 minus P4 after passing through an upsample module Upsample and a PConv module, and then minus P5 after passing through an upsample module and a PConv module twice;

[0034] F4(P3, P4, P5) = P4-PConv(Upsample(P5))-PConv(Downsample(P4)), which means P4 minus P5 after an upsampling module and a PConv module, and then minus P3 after a downsampling module Downsample and a PConv module;

[0035] F5(P3, P4, P5) = P5 - PConv(Downsample(P4)) - 2*PConv(Downsample(P3)), which means P5 minus P4 after one downsampling module and one PConv module, and then minus P3 after two downsampling modules and one PConv module.

[0036] The upsampling module is used to increase the resolution of the input feature map; the downsampling module is used to reduce the resolution of the input feature map; the PConv module is used to adjust the number of channels of the input feature map and set the learnable parameters; the PC3 module performs a learnable internal feature fusion on the composite feature map generated after the feature map addition operation.

[0037] A lightweight distributed fiber optic sensing image target detection system, comprising:

[0038] The data set construction module collects distributed fiber sensing images and corresponding annotations and performs preprocessing to obtain a training set;

[0039] A model building module is used to build a target detection network based on the improved YOLOv8 model: the target detection network includes a backbone network, a lightweight multi-scale fusion module and a detection head; the backbone network of the target detection network is obtained by replacing the convolutional layer of the backbone network in the YOLOv8 model with the PConv module and the C2f module of the backbone network in the YOLOv8 model with the PC3 module; the detection head of the target detection network adopts the detection head of the YOLOv8 model; the backbone network of the target detection network is used to extract feature maps B3, B4, and B5 of three different sizes of the distributed optical fiber perception image x; the lightweight multi-scale fusion module is used to fuse the feature maps B3, B4, and B5, and output fused feature maps D3, D4, and D5; the three fused feature maps are respectively input into the three detection heads of the target detection network to obtain the target detection results; the lightweight multi-scale fusion module includes Two stages; in the first stage, the feature maps B3, B4, and B5 are fused into three intermediate feature maps P3, P4, and P5 through the upsampling module, the splicing module, the PC3 module, and the PConv module; in the second stage, the intermediate feature maps P3, P4, and P5 are taken as input, one branch compares and fuses P3, P4, and P5 with each other, and the other branch retains the original information and jumps to the output fused feature maps D3, D4, and D5; the PConv module is a fast convolution module with partial convolution as the core operator; the PC3 module includes two tributaries, and the feature map input to the PC3 module passes through a PConv module in one tributary, and is sequentially input into a PConv module and n Bottleneck modules in the other tributary; the feature maps output by the two tributaries are spliced ​​and then passed through a PConv module to obtain the output of the PC3 module;

[0040] A training module, which trains the target detection network through a training set until the target detection network converges or reaches a set number of training times;

[0041] The inference module inputs the distributed fiber optic sensing image used for testing into the trained target detection network and outputs the target detection result.

[0042] The system of the present invention corresponds to the method, and the specific technical solutions applicable to the method are also applicable to the system.

[0043] Compared with the prior art, the beneficial technical effects of the present invention are:

[0044] 1. The Lightweight Multi-Scale Fusion Module (LMFM) proposed in the present invention is a neural network structure, which is responsible for the efficient and lightweight fusion of multiple input feature maps of different scales, and outputs the fused feature maps of different scales; it adjusts and aligns the feature maps of different scales through upsampling and downsampling, and quickly adjusts the number of channels through a custom module, so that the shapes and sizes of different feature map tensors are aligned, and for the first time, feature fusion is achieved through the basic operations of adding and subtracting feature map tensors, and no longer relies too much on the use of concatenation operations (Concat) for fusion, making the network more efficient; 2. The custom PConv module and the custom PC3 module can effectively improve the parameter utilization efficiency and feature map fusion performance, and it is an important component for realizing network lightweighting; 3. The modules proposed above are used to realize the rapid feature extraction of the backbone and the efficient feature fusion of the neck of the entire target detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is the overall flow chart of the present invention;

[0046] Figure 2 is a schematic diagram of the PConv module in the present invention;

[0047] Figure 3 is a schematic diagram of the PC3 module in the present invention;

[0048] Figure 4 is a schematic diagram of a lightweight multi-scale fusion module in the present invention;

[0049] Figure 5 It is a schematic diagram of the entire distributed fiber optic sensing image target detection network in the present invention. DETAILED DESCRIPTION

[0050] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0051] This embodiment discloses a lightweight distributed optical fiber sensing image target detection method, comprising the following steps:

[0052] S1, collect distributed fiber sensing images and preprocess them to obtain a reliable training set.

[0053] S11, distributed fiber optic sensing image data collection, including but not limited to publicly available commonly used distributed fiber optic sensing image target detection datasets and self-made distributed fiber optic sensing image target detection datasets.

[0054] S12, data labeling, data preprocessing, and data partitioning: manually or automatically label the targets in the distributed fiber optic sensing images; convert the distributed fiber optic sensing images and annotations into a format that can be processed by the target detection network, usually by normalizing and resizing the distributed fiber optic sensing images, and converting the annotation information into the required input format of the target detection network, such as the coordinate offset, width and height ratio in the YOLO model; divide the data set into training set, validation set, and test set.

[0055] S2. Build a distributed fiber-optic perception image target detection network, including a custom PConv module, a custom PC3 module, and a custom lightweight multi-scale fusion module.

[0056] S21, including custom PConv module:

[0057] PConv module (partial convolutions module, PconvModule) represents a fast convolution module with partial convolution as the core operator. Figure 2 As shown in the figure, for a given feature map, it passes through a standard partial convolution layer PConv2d, batch normalization BatchNorm2d and SiLU activation function in sequence, the convolution kernel size is k×k, the step size is s, the padding is p, and the number of output channels is c.

[0058] The standard partial convolution layer PConv2d represents a convolution calculation similar to the ordinary two-dimensional convolution for a given feature map. Partial convolution is a special convolution operation, which is often used in tasks such as image restoration and occlusion removal. This concept was first proposed by researchers in 2018; traditional convolution operations will consider the information of all pixels in the image at the same time, including occluded areas. However, in tasks such as image restoration, occluded areas often lead to information loss, so a convolution operation that can adaptively handle occluded areas is required. The core idea of ​​partial convolution is to introduce a mask in the convolution operation. This mask corresponds to the occluded area of ​​the input image. For pixels in the occluded area, the calculation of partial convolution only considers the unoccluded pixels around it. This can avoid the influence of the occluded area on the convolution operation and retain the valid information in the image. It is precisely because of these characteristics that the computational cost of partial convolution is lower than that of standard convolution. When used in other occasions, it can effectively reduce redundant information, thereby saving computing resources;

[0059] In the PConv module, the feature map flows through the partial convolution and then goes through batch normalization, SiLU activation function, and then gets the output.

[0060] S22, Custom PC3 Module:

[0061] PC3 module (partial convolutions c3 module, PC3Module) represents a fusion module with three partial convolutions as core operators. Figure 3 As shown in the figure, for a given feature map, after entering the PC3 module, it is divided into two tributaries: one is the left tributary, which gets the output after passing through a PConv module; the other is the right tributary, which also passes through a Pconv module first, then n Bottleneck modules (a standard module in the YOLO series), and finally gets the output; then the outputs of the two tributaries are concatenated (Concat), and the combined feature map is passed through a PConv module to obtain the final output of the PC3 module.

[0062] In the PC3 module, the PConv module parameters in the two tributaries are set as follows: convolution kernel size k = 1, step size s = 1, padding p = 0, and number of output channels c = 0.5*c_out; the last PConv module parameters are set as follows: convolution kernel size k = 1, step size s = 1, padding p = 0, and number of output channels c = 0.5*c_out; where c_out refers to the final required number of output channels in the PC3 module.

[0063] S23, lightweight multi-scale fusion module:

[0064] The Lightweight Multi-Scale Fusion Module (LMFM) is different from the PC3 module, which only performs internal feature fusion on one feature map. The input of the Lightweight Multi-Scale Fusion Module is three feature maps of different sizes, and its output is the three feature maps generated after fusion. Figure 4 As shown in the figure, the three feature maps of different sizes input in the first stage are denoted as B3, B4, and B5, which are output as three intermediate feature maps of the same size after passing through the fusion network of the first stage, denoted as P3, P4, and P5; in the second stage, P3, P4, and P5 are taken as input, one branch compares and fuses P3, P4, and P5 with each other, and the other branch retains the original information and makes a jump connection to the output D3, D4, and D5;

[0065] In the lightweight multi-scale fusion module embodiment, the feature fusion of the first stage is specifically as follows:

[0066] For B3 to P3, B3 is concatenated and superimposed with the feature maps from the fusion of B4 and B5 modules, and then fused and adjusted by a PC3 module to generate P3; for B4 to P4, B4 is fused and adjusted by the module from the fusion of B5 module by the PC3 module, and then used as the first component, and the feature map of P3 after a PConv module is used as the second component, the two components are concatenated and superimposed, and finally fused and adjusted by a PC3 module to generate P4; for B5 to P5, the feature map of the previously generated P4 after a PConv module is used as one component, and the original B5 is used as another component, the two components are concatenated and superimposed, and finally P5 is generated after fusion and adjustment by a PC3 module.

[0067] In this stage, the upsample module is used to adjust the resolution of the feature map, adjusting the low-resolution feature map to a high-resolution feature map; the splicing operation refers to the superposition of multiple feature maps of the same size according to the channel dimension to generate a composite tensor; the PConv module parameters are set to: convolution kernel size k = 3, step size s = 2, padding p = 1, which is used to reduce the resolution of the feature map while setting the parameters that can be learned; the PC3 module performs a learnable internal feature fusion on the composite tensor generated after the splicing operation.

[0068] In the lightweight multi-scale fusion module embodiment, the feature fusion of the second stage is specifically as follows:

[0069] For P3 to D3, the new feature map generated by adding the original P3 through a jump connection and the feature map of F3 (P3, P4, P5) after the internal feature fusion of the PC3 module is D3; for P4 to D4, the new feature map generated by adding the original P4 through a jump connection and the feature map of F4 (P3, P4, P5) after the internal feature fusion of the PC3 module is D4; for P5 to D5, the new feature map generated by adding the original P3 through a jump connection and the feature map of F5 (P3, P4, P5) after the internal feature fusion of the PC3 module is D5.

[0070] F3(P3, P4, P5) = P3 - PConv(Upsample(P4)) - 2*PConv(Upsample(P5)); it means P3 minus P4 which has passed through an upsampling module and a PConv module, and then minus P5 which has passed through an upsampling module and a PConv module twice.

[0071] F4(P3, P4, P5) = P4-PConv(Upsample(P5))-PConv(Downsample(P4)); it means P4 minus P5 after an upsampling module and a PConv module, and then minus P3 after a downsampling module and a PConv module.

[0072] F5(P3, P4, P5) = P5 - PConv(Downsample(P4)) - 2*PConv(Downsample(P3)); it means that P5 is subtracted from P4 which has passed through a downsampling module and a PConv module, and then from P3 which has passed through a downsampling module and a PConv module twice.

[0073] In this stage, the upsample module is used to adjust the resolution of the feature map, adjusting the low resolution to a high resolution feature map; the downsample module is used to adjust the high resolution to a low resolution feature map; the PConv module parameters are set to: convolution kernel size k = 1, step size s = 1, padding p = 0, which is used to adjust the number of channels of the feature map and set the parameters available for learning; the PC3 module performs a learnable internal feature fusion on the composite tensor generated after the tensor addition operation.

[0074] The convolution step size of the PConv module in the first stage is set to 2, and the PConv module plays a role of downsampling; the convolution step size of the PConv module in the second stage is set to 1, and it does not play a role of downsampling, but uses the output channel setting to adjust the channel, adapt to tensor calculation, control the network scale and calculation amount. In addition, the presence of the PConv module brings learnable parameters.

[0075] S3. Build an overall distributed fiber optic perception image target detection network model and train it.

[0076] The overall distributed fiber optic sensing image target detection network model structure is as follows: Figure 5 As shown in the figure, it is based on the YOLO architecture and can be roughly divided into the backbone network part on the left, the neck fusion part (lightweight multi-scale fusion module), and the detection head part. The input distributed fiber perception image is uniformly adjusted to a 640*640*3 format, and the main features are extracted through the backbone network. The three main feature maps B3, B4, and B5 are input into the lightweight multi-scale fusion module, and then the output feature maps D3, D4, and D5 are obtained; the last step is to input D3, D4, and D5 into the standard detection head (Detect Module) in the YOLOv8 algorithm to complete the detection.

[0077] S4. Model application.

[0078] The training set, validation set, and test set obtained in S1 are used for the model trained in S3 to obtain the model reasoning result graph actually needed for distributed fiber optic sensing image target detection; the detection results are then fed back to the system user.

[0079] The following is a supplementary explanation of some of the English words used in the drawings of the present invention.

[0080] Input / output: input / output feature map;

[0081] BatchNorm2d: Batch normalization, used to normalize two-dimensional input data;

[0082] Concat: concatenates feature maps or feature vectors according to the specified dimension to form a larger feature set;

[0083] “-” sign: feature map subtraction operation in Pytorch;

[0084] “+” sign: feature map addition operation in Pytorch;

[0085] Figure 4 The dashed arrows in : represent skip connections;

[0086] SiLU: SiLU activation function proposed by Google in 2017;

[0087] Bottleneck: A common standard module used to reduce the number of model parameters and the amount of computation while increasing the depth and width of the network.

[0088] The system of the present invention corresponds to the method, and the specific technical solutions applicable to the method are also applicable to the system.

[0089] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0090] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A lightweight distributed fiber-optic sensing image target detection method, characterized in that: Input the distributed fiber optic sensing image for testing into the trained target detection network, and output the target detection result; the construction and training method of the target detection network includes the following steps: Step 1: collect distributed fiber sensing images x and corresponding annotations and perform preprocessing to obtain a training set; Step 2: Build a target detection network based on the improved YOLOv8 model: The target detection network includes a backbone network, a lightweight multi-scale fusion module and a detection head; the backbone network of the target detection network is obtained by replacing the convolutional layer of the backbone network in the YOLOv8 model with the PConv module and replacing the C2f module of the backbone network in the YOLOv8 model with the PC3 module; the detection head of the target detection network adopts the detection head of the YOLOv8 model; the backbone network of the target detection network is used to extract feature maps B3, B4, and B5 of three different sizes of the distributed optical fiber perception image x; the lightweight multi-scale fusion module is used to fuse the feature maps B3, B4, and B5, and output fused feature maps D3, D4, and D5; the three fused feature maps are respectively input into the three detection heads of the target detection network to obtain the target detection results; The lightweight multi-scale fusion module consists of two stages. In the first stage, the feature maps B3, B4, and B5 are fused into three intermediate feature maps P3, P4, and P5 through the upsampling module, the concatenation operation, the PC3 module, and the PConv module. In the second stage, the intermediate feature maps P3, P4, and P5 are taken as input, one branch compares and fuses P3, P4, and P5 with each other, and the other branch retains the original information and jumps to the output fused feature maps D3, D4, and D5. The PConv module is a fast convolution module with partial convolution as the core operator; The PC3 module includes two tributaries. The feature map input to the PC3 module passes through a PConv module in one tributary, and is sequentially input into a PConv module and n Bottleneck modules in the other tributary. The feature maps output by the two tributaries are concatenated and then passed through a PConv module to obtain the output of the PC3 module. Step 3: training the target detection network using the training set until the target detection network converges or reaches a set number of training times; The PConv module is composed of a standard partial convolution layer, batch normalization and SiLU activation function; the feature map input to the PConv module passes through the standard partial convolution layer in sequence, and then passes through the batch normalization operation and SiLU activation function to obtain the feature map output by the PConv module; The first stage of the lightweight multi-scale fusion module includes: After upsampling, feature map B5 is concatenated with feature map B4 to obtain feature map A1; A1 is input into the PC3 module and the upsampling module, and concatenated with feature map B3 to obtain feature map A2; A2 is input into the PC3 module to obtain the intermediate feature map P3; The result obtained by inputting feature map A2 into PC3 module and PConv module in sequence is concatenated with the result obtained by inputting feature map A1 into PC3 module to obtain feature map A3; A3 is input into PC3 module to obtain intermediate feature map P4; The feature map A3 is input into the PC3 module and the PConv module in sequence, and the result is concatenated with the feature map B5 to obtain the feature map A4; A4 is input into the PC3 module to obtain the intermediate feature map P5; The upsampling module is used to increase the resolution of the input feature map; the concatenation operation is to superimpose the input feature maps of the same size according to the channel dimension and output a composite feature map; the PConv module is used to reduce the resolution of the input feature map and set learnable parameters; the PC3 module performs a learnable internal feature fusion on the composite features generated after the concatenation operation; The second stage of the lightweight multi-scale fusion module includes: The original intermediate feature map P3 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F3 (P3, P4, P5) to generate the fused feature map D3; The original intermediate feature map P4 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F4 (P3, P4, P5) to generate the fused feature map D4; The original intermediate feature map P3 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F5 (P3, P4, P5), and the generated fused feature map is D5; F3(P3, P4, P5) = P3-PConv(Upsample(P4))-2*PConv(Upsample(P5)), which means P3 minus P4 after passing through an upsample module Upsample and a PConv module, and then minus P5 after passing through an upsample module and a PConv module twice; F4(P3, P4, P5) = P4-PConv(Upsample(P5))-PConv(Downsample(P4)), which means P4 minus P5 after an upsampling module and a PConv module, and then minus P3 after a downsampling module Downsample and a PConv module; F5(P3, P4, P5) = P5 - PConv(Downsample(P4)) - 2*PConv(Downsample(P3)), which means P5 minus P4 after one downsampling module and one PConv module, and then minus P3 after two downsampling modules and one PConv module. The upsampling module is used to increase the resolution of the input feature map; the downsampling module is used to reduce the resolution of the input feature map; the PConv module is used to adjust the number of channels of the input feature map and set the learnable parameters; the PC3 module performs a learnable internal feature fusion on the composite feature map generated after the feature map addition operation.

2. A lightweight distributed fiber-optic sensing image target detection system, characterized in that: include: The data set construction module collects distributed fiber sensing images and corresponding annotations and performs preprocessing to obtain a training set; A model building module is used to build a target detection network based on the improved YOLOv8 model: the target detection network includes a backbone network, a lightweight multi-scale fusion module and a detection head; the backbone network of the target detection network is obtained by replacing the convolutional layer of the backbone network in the YOLOv8 model with a PConv module and replacing the C2f module of the backbone network in the YOLOv8 model with a PC3 module; the detection head of the target detection network adopts the detection head of the YOLOv8 model; the backbone network of the target detection network is used to extract feature maps B3, B4, and B5 of three different sizes of the distributed optical fiber perception image x; the lightweight multi-scale fusion module is used to fuse the feature maps B3, B4, and B5, and output fused feature maps D3, D4, and D5; The three fused feature maps are respectively input into the three detection heads of the target detection network to obtain the target detection results; the lightweight multi-scale fusion module includes two stages; the first stage fuses the feature maps B3, B4, and B5 into three intermediate feature maps P3, P4, and P5 through the upsampling module, the splicing module, the PC3 module, and the PConv module; the second stage takes the intermediate feature maps P3, P4, and P5 as input, one branch compares and fuses P3, P4, and P5 with each other, and the other branch retains the original information and jumps to the output fused feature maps D3, D4, and D5; the PConv module is a fast convolution module with partial convolution as the core operator; the PC3 module includes two tributaries, the feature map input into the PC3 module passes through a PConv module in one tributary, and is sequentially input into a PConv module and n Bottleneck modules in the other tributary; the feature maps output by the two tributaries are spliced ​​and then pass through a PConv module to obtain the output of the PC3 module; A training module, which trains the target detection network through a training set until the target detection network converges or reaches a set number of training times; The inference module inputs the distributed fiber perception image used for testing into the trained target detection network and outputs the target detection result; The PConv module is composed of a standard partial convolution layer, batch normalization and SiLU activation function; the feature map input to the PConv module passes through the standard partial convolution layer in sequence, and then passes through the batch normalization operation and SiLU activation function to obtain the feature map output by the PConv module; The first stage of the lightweight multi-scale fusion module includes: After upsampling, feature map B5 is concatenated with feature map B4 to obtain feature map A1; A1 is input into the PC3 module and the upsampling module, and concatenated with feature map B3 to obtain feature map A2; A2 is input into the PC3 module to obtain the intermediate feature map P3; The result obtained by inputting feature map A2 into PC3 module and PConv module in sequence is concatenated with the result obtained by inputting feature map A1 into PC3 module to obtain feature map A3; A3 is input into PC3 module to obtain intermediate feature map P4; The feature map A3 is input into the PC3 module and the PConv module in sequence, and the result is concatenated with the feature map B5 to obtain the feature map A4; A4 is input into the PC3 module to obtain the intermediate feature map P5; The upsampling module is used to increase the resolution of the input feature map; the concatenation operation is to superimpose the input feature maps of the same size according to the channel dimension and output a composite feature map; the PConv module is used to reduce the resolution of the input feature map and set learnable parameters; the PC3 module performs a learnable internal feature fusion on the composite features generated after the concatenation operation; The second stage of the lightweight multi-scale fusion module includes: The original intermediate feature map P3 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F3 (P3, P4, P5) to generate the fused feature map D3; The original intermediate feature map P4 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F4 (P3, P4, P5) to generate the fused feature map D4; The original intermediate feature map P3 is added to the feature map after the internal feature fusion of the PC3 module through the jump connection operation F5 (P3, P4, P5), and the generated fused feature map is D5; F3(P3, P4, P5) = P3-PConv(Upsample(P4))-2*PConv(Upsample(P5)), which means P3 minus P4 after passing through an upsample module Upsample and a PConv module, and then minus P5 after passing through an upsample module and a PConv module twice; F4(P3, P4, P5) = P4-PConv(Upsample(P5))-PConv(Downsample(P4)), which means P4 minus P5 after an upsampling module and a PConv module, and then minus P3 after a downsampling module Downsample and a PConv module; F5(P3, P4, P5) = P5 - PConv(Downsample(P4)) - 2*PConv(Downsample(P3)), which means P5 minus P4 after one downsampling module and one PConv module, and then minus P3 after two downsampling modules and one PConv module. The upsampling module is used to increase the resolution of the input feature map; the downsampling module is used to reduce the resolution of the input feature map; the PConv module is used to adjust the number of channels of the input feature map and set the learnable parameters; the PC3 module performs a learnable internal feature fusion on the composite feature map generated after the feature map addition operation.

Images