Concrete dam hollowing hidden danger identification method based on small sample infrared feature learning

Abstract

The invention discloses a concrete dam hollowing hidden danger identification method based on small sample infrared feature learning, and the method comprises the following steps: firstly, calculating a simulated infrared image through a dam body temperature field, and constructing a pre-training data set of transfer learning; secondly, constructing a double-attention and multi-scale feature enhancement semantic segmentation model which comprises a space-channel double-attention encoder, a multi-scale feature enhancement module, a space-channel double-attention decoder and a prediction module based on a combination loss function; secondly, the constructed semantic segmentation model is deployed on a simulation data set for pre-training to generate a pre-training model, and then the pre-training model is called on an actually measured dam hollowing infrared image data set to complete transfer learning under a small sample. And finally, deploying the model to an edge end, and identifying concrete dam hollowing. The detection efficiency of concrete dam hollowing is improved, and an intelligent basis is provided for hidden danger investigation of the concrete dam.

Description

Technical Field

[0001] The present invention relates to the technical field of semantic segmentation, and in particular to a method for identifying hollowing hazards of concrete dams based on small sample infrared feature learning. Background Art

[0002] Infrared non-destructive hidden danger detection technology is a non-contact non-destructive testing technology widely used in aerospace, transportation, industrial manufacturing and other fields. When cracks or hollows appear in concrete buildings, internal insulation defects will occur, causing a temperature difference between the abnormal part of the concrete structure and the surrounding area, which appears as color spots in the infrared image. Infrared non-destructive testing technology is to observe color spots through instruments and check for hidden dangers. In the industrial field, this technology is widely used to detect fatigue loads, underground pipelines, and metal cracks. For example, in the construction industry, infrared instruments are used to detect walls and output infrared images. The staff found color spots in the image and preliminarily determined that there were insulation defects here, which facilitated further investigation of building hazards.

[0003] Traditional infrared non-destructive hidden danger detection technology is an important method for eliminating hidden dangers in the construction industry. This technology uses instruments and staff as carriers, and has active and passive types. The active type requires the arrangement of heat sources and is suitable for detection in key areas, but the heat source is not easy to arrange, which is not conducive to long-distance and large-area detection; the passive type does not require the arrangement of heat sources and performs well in long-distance and large-area detection tasks, but has high requirements for external temperature and time, and low resolution.

[0004] With the development of machine learning, deep learning is widely used in image classification, semantic segmentation, and feature extraction. Semantic segmentation technology is applied in the field of infrared detection. Infrared images are used as training sample sets to segment temperature anomaly images, which effectively improves the efficiency of hidden danger detection, reduces costs, and avoids errors and omissions caused by subjective factors in manual detection. Traditional deep learning requires a large number of training samples to produce effective models. The development of transfer learning also makes deep learning possible under small sample conditions. Transfer learning is a new machine learning method that uses existing knowledge to solve problems in different but related fields. In machine learning, transfer learning first uses a large data set with strong correlation for pre-training, extracts features when deployed to a specific data set, and then trains the target model. This technology effectively solves the problem caused by insufficient training samples under small sample conditions, and is highly suitable for deployment in the embedded field.

[0005] In the field of concrete dam hidden danger inspection, the use of infrared non-destructive hidden danger detection technology to detect dam hollowing requires staff to hold infrared instruments to inspect the dam without external heat sources, and pay attention to the color spots in the infrared image and the location of the color spots in the dam. The detection range of the instrument and the range of activities of the staff limit the detection area of ​​infrared detection technology. The characteristics of long dam inspection time, small range, and poor timeliness make it inefficient to use manual handheld instruments to detect hollowing with infrared. In addition, the thermal imaging of concrete dam hollowing has the characteristics of small number, high cost, and inconvenient collection. The training results of traditional deep learning under small sample conditions may not perform well. Therefore, it is necessary to study a new type of transfer learning semantic segmentation deep learning model under small sample conditions and deploy it on the edge to achieve efficient identification of concrete dam hollowing. Summary of the invention

[0006] Purpose of the invention: In order to improve the efficiency of infrared detection of hollow concrete dams, the present invention provides a low-cost, high-efficiency semantic segmentation deep learning model to provide technical support for the hidden danger inspection of concrete dams.

[0007] Technical solution: To achieve the above invention purpose, the technical solution adopted by the present invention is: a method for identifying hollowing hazards of concrete dams based on small sample infrared feature learning, comprising the following steps:

[0008] (1) Construct a temperature field model of the concrete dam body and hollowing. Calculate the simulated temperature of the dam surface pixel by pixel at different time nodes based on the temperature field model to construct an infrared feature image of the dam body surface. Semantically annotate the hollowing area to form a pre-training dataset for transfer learning. Use the infrared images of concrete dam hollowing collected in the field as a small sample infrared feature learning dataset.

[0009] (2) Construct a semantic segmentation model with spatial-channel dual attention and multi-scale feature enhancement, including a spatial-channel dual attention encoder, a multi-scale enhanced feature generation module, a spatial-channel dual attention decoder, and a prediction module based on a weighted combination loss function;

[0010] (3) Construct a spatial-channel dual attention encoder, which includes a 4-layer feature extraction module, each of which contains 2 convolution modules and a spatial-channel dual attention mechanism. Each convolution module contains a 2D convolution layer, a batch normalization layer, and a rectified linear unit layer. The number of convolution kernels is 32, 64, 128, and 128 respectively;

[0011] (4) Construct a multi-scale enhanced feature generation module, which is located between the encoder and the decoder. It uses the shallow feature information of the encoder layers 1-3 and the context feature information of the 4th layer to enhance the network's adaptability to complex backgrounds and generate enhanced feature maps.

[0012] (5) Construct a spatial-channel dual attention decoder. This module has a three-layer network structure. Each layer of the network performs bilinear interpolation on the feature map for upsampling. The number of convolution kernels for upsampling in each layer is consistent with that of the encoder network at the same layer, and each layer embeds a spatial and channel dual attention mechanism.

[0013] (6) Construct a prediction module based on the combined loss function and construct a loss function consisting of the sum of BCE Loss and Dice Loss, denoted as BCEDiceLoss;

[0014] (7) The constructed semantic segmentation model is pre-trained using the constructed simulation data set, and the model parameters are transferred and used in the learning of the measured dam hollowing infrared features to complete the identification of concrete dam hollowing hazards based on the measured small sample image data set;

[0015] Furthermore, the simulation data set construction process in step (1) specifically includes:

[0016] (11) The dam temperature field is modeled as follows:

[0017]

[0018] Where T is T(x, 0), i.e., the surface temperature of the dam, x is the depth perpendicular to the dam surface, t is the time, and a is the thermal conductivity of the dam material.

[0019] The initial conditions and boundary conditions for constructing the dam temperature field to obtain the definite solution are:

[0020] T(x,0)=T0 (2)

[0021] T(x max ,t)=T0 (3)

[0022]

[0023] Where T0 is the initial temperature, x max is the maximum depth at which hidden dangers can be detected using infrared detection technology, h is the comprehensive heat transfer coefficient, is the comprehensive atmospheric temperature, and λ is the thermal conductivity of the concrete dam.

[0024] (12) By setting different time nodes and hollow depth parameters, the surface temperature of the normal area and the potential hazard area of ​​the concrete dam can be calculated based on the numerical solution of the partial differential equation, thereby obtaining the pixel-by-pixel temperature field of the dam surface at different time points, which is converted into a two-dimensional infrared image as a sample set for pre-training of the semantic segmentation model.

[0025] Furthermore, the construction of the space-channel dual attention mechanism in step (3) includes:

[0026] (31) In the channel attention module, a global pooling operation is first performed on the feature map. Then the module performs two 1×1 convolution operations on the feature map output by the pooling layer, and finally passes it to the activation function and maps it to generate a channel weight vector.

[0027] (32) After the spatial attention module performs a 1×1 convolution operation on the feature map, it is transmitted to the activation function to generate a spatial weight vector;

[0028] (33) The channel weight vector and the spatial weight vector are multiplied with the feature map respectively and then added together to obtain the weight-corrected channel-spatial dual attention fusion feature map.

[0029] Furthermore, the multi-scale enhancement feature module in step (4) includes:

[0030] (41) Feature generation consists of two parts: one is that the first three layers of the module are convolved and downsampled to obtain shallow feature information; the other is that the fourth layer generates multi-scale contextual enhanced features based on different hole convolutions;

[0031] (42) The context enhancement process of the feature map of the 4th layer includes compressing the feature map channel to 32 using 1×1 convolution, extracting multi-scale context features through three dilation convolutions with different dilation degrees, and splicing the channel dimension to obtain a 32-channel feature map again; downsampling the shallow feature maps from the first three layers of the encoder by 8 times, 4 times, and 2 times respectively, and using 1×1 convolution to synchronously adjust the number of channels to 32, splicing and fusion with the context-enhanced 4th layer feature map in the channel dimension, and applying the channel weight vector to the fused feature map to generate a multi-scale enhanced feature map.

[0032] Furthermore, the loss function BCEDiceLoss in step (6) is formed by adding the BCE Loss function and the Dice Loss function. The BCE Loss function is a binary cross entropy loss function, and the formula is:

[0033]

[0034] Where N is the number of pixels in the image, and t i represents the true label of the pixel (0 or 1), P i is the network's prediction for the pixel.

[0035] Dice Loss, also known as the Dice coefficient, is a metric function used to evaluate the similarity between two samples. The larger the value, the more similar the two samples are, indicating:

[0036]

[0037] Here, ε is the smoothing coefficient.

[0038] The combined loss function BCEDiceLoss formed by adding the two can be written as:

[0039]

[0040] Dice loss is more inclined to explore the foreground area. Using the sum of BCELoss and Dice Loss as the loss function of the model can make the model pay attention to the difference between the predicted value and the true label value while also focusing on the extraction of foreground information, which is conducive to alleviating the imbalance problem of positive and negative samples in the crack segmentation task where the foreground accounts for a small proportion.

[0041] Beneficial effects: The method for identifying the hidden dangers of hollowing of concrete dams based on small sample infrared feature learning constructed by the present invention can effectively solve the problems of low efficiency and high cost in the investigation of hidden dangers of hollowing of concrete dams. Since the number of measured infrared images of hollowing of concrete dams is small, the cost is high, and the acquisition is difficult, the transfer learning model based on small sample conditions has important use value; at the same time, transfer learning under small samples is also an important research direction of image recognition, so the invention has important theoretical value. Through pre-training on simulated data sets and retraining on small samples, unmanned intelligent detection of hollowing of concrete dams becomes possible. The model is built on the edge and carried on the drone to improve the detection efficiency and accuracy of hollowing of concrete dams. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of the method for identifying hollowing of concrete dams according to an embodiment of the present invention

[0043] Figure 2 This is a schematic diagram of the network structure of the dual attention and multi-scale feature enhanced semantic segmentation model of the embodiment of the present invention.

[0044] Figure 3 This is a graph of the loss function of the model of the embodiment of the present invention after multiple complete training of all the data in the training set.

[0045] Figure 4 This is the experimental effect diagram of the present invention on the measured concrete dam hollowing data set Specific implementation methods

[0047] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0048] Step 1: Construct a temperature field model of the concrete dam body and hollowing. Calculate the pixel-by-pixel simulated temperature of the dam surface at different time nodes based on the temperature field model to construct an infrared feature image of the dam body surface, and semantically annotate the hollowing area to form a pre-training dataset for transfer learning. Use the infrared images of concrete dam hollowing collected in the field as a small sample infrared feature learning dataset, and construct a simulation dataset of concrete dam hollowing.

[0049] It is difficult to collect infrared images of hollow concrete dams. The images are costly, small in number, and have a small conversion ratio between measured images and sample images. Therefore, we use transfer learning to create a simulated data set as a pre-training sample for transfer learning. Most dams are large in size and long in shape. The length and width are often much greater than the thickness. The temperature field of the dam is modeled using the formula:

[0050]

[0051] Where T is T(x,t), which is the surface temperature of the dam in °C, and x is the depth perpendicular to the dam surface in meters, ranging from 0 to x max , x max is the maximum depth that infrared can affect, that is, the maximum depth that can be explored by infrared nondestructive testing technology. After exceeding this depth, no specific effective position can be obtained. The present invention is 1.0m, t is time, the unit is s, and 7:00:00 in the morning is 0s, 19:00:00 in the afternoon is 43200s, and the value range is 0 to t max , t max is the maximum analysis time, which is 43200s, and a is the thermal conductivity of the dam material, in m 2 ·s -1 , the normal concrete dam value is 5.737×10 -7 , the value of air hollowing is 2.311×10 -5 .

[0052] The initial conditions and boundary conditions for constructing the dam temperature field to obtain the definite solution are:

[0053] T(x, 0) = T0 (2)

[0054] T(x max ,t)=T0 (3)

[0055]

[0056] Where T0 is the initial temperature, which is 20, x max is the maximum depth at which the infrared detection technology can be used to detect hidden dangers, h is the comprehensive heat transfer coefficient, and its value is 10. is the comprehensive atmospheric temperature, the specific value requires the measured temperature of the local meteorological station at the identification time, and λ is the thermal conductivity of the concrete dam.

[0057] Set different time nodes and hollow depth parameters. First, model a cube with a length of 4m, a width of 1m, and a height of 4m as a concrete dam. Dig a cylinder with a radius of 1m and a height of 0.2m in it as an air interlayer. Change the depth and thickness of the air interlayer to create four types of hollow hidden dangers, namely, depth 0.1m, thickness 0.20m, depth 0.10m, thickness 0.10m, depth 0.05m, thickness 0.05m, depth 0.05m, thickness 0.15m. Then, using specific software, the surface temperature of the normal area and the hidden danger area of ​​the concrete dam can be calculated based on the numerical solution of the partial micro equation, so as to obtain the pixel-by-pixel temperature field of the dam surface at different time points, convert it into a two-dimensional infrared image, and perform data enhancement operations such as zooming in, zooming out, and cropping on the generated image to generate the final simulation data set as a sample set for pre-training the semantic segmentation model.

[0058] Step 2: Build a deep learning neural network. The present invention proposes a dual attention and multi-scale feature enhancement semantic segmentation model, which includes a space-channel dual attention encoder, a multi-scale enhancement feature module, a space-channel dual attention decoder, and a loss function based on a combination function. The sample set is convolved and down-sampled by the encoder to generate a feature map, and then a feature enhancement feature map is generated by a multi-scale enhancement module feature block. Finally, the final feature map is output through convolution and up-sampling of the decoder. The space-channel dual attention mechanism is also introduced in the encoder and decoder. There are 4 layers in the encoder, each layer has 2 convolution modules and a space-channel dual attention module. Each convolution module contains a two-dimensional convolution layer, a batch normalization layer, and a rectified linear unit layer. The number of convolution kernels in the 4 layers is 32, 64, 128, and 128 respectively.

[0059] The spatial and channel dual attention mechanism cleverly combines the channel attention module and the spatial attention module. In the channel attention module, the feature map is first globally pooled, and its size is compressed from H×W×C to 1×1×C to achieve dimensional compression. Then the module performs two 1×1 convolution operations on the feature map output by the pooling layer to capture the deep feature information in the channel dimension. The size of the output feature map is reduced to H×W×1, which achieves the compression of the feature map channel. Finally, it is passed to the activation function, which accurately maps the values ​​in the feature map to the interval from 0 to 1 as the channel weight vector. After the spatial attention module performs a 1×1 convolution operation on the feature map, it is transmitted to the activation function to generate the spatial weight vector. The channel weight vector and the spatial weight vector are multiplied with the feature map respectively and added and fused to obtain the weight-corrected channel-spatial dual attention fusion feature map. This design based on the spatial and channel dual attention mechanism enables the model to more effectively enhance the feature information while weakening the influence of interference noise.

[0060] The multi-scale enhanced feature module will fuse the features of two parts. One part is the shallow feature information obtained by convolution and downsampling of the first three layers of the module, and the other part is the multi-scale contextual enhanced features generated by the fourth layer based on different hole convolutions. The shallow feature maps from the first three layers of the encoder are downsampled by 8 times, 4 times, and 2 times respectively, and the number of channels is synchronously adjusted to 32 using 1×1 convolution. The feature map context enhancement process of the fourth layer includes compressing the feature map channels to 32 using 1×1 convolution, extracting multi-scale contextual features through three hole convolutions with different expansion degrees, splicing the channel dimension to obtain a 32-channel feature map again, splicing and fusion with the first three layers of downsampled feature maps in the channel dimension, and applying the channel weight vector to the fused feature map to generate a multi-scale enhanced feature map.

[0061] Traditional convolutional neural networks increase the receptive field by stacking convolutional layers and performing pooling operations. However, this process is accompanied by a reduction in the resolution of feature maps, resulting in a loss of feature information. Dilated convolutions can expand the receptive field without performing pooling operations or reducing the resolution of feature maps by expanding the convolution sampling area.

[0062] Construct the BCEDiceLoss loss function, which is a combination of the BCE Loss loss function and the Dice Loss function. The BCE Loss function is a loss function based on the cross entropy in Shannon information theory, which is used to measure the difference information between two probability distributions. The cross entropy uses a non-real distribution to represent the average encoding length from the real distribution, which can be expressed as follows:

[0063] H(P,Q)=-∑ i P(i)·log Q(i) (5)

[0064] in For variables The true probability distribution of For variables The non-real probability distribution of and When the distributions are exactly the same, the value of the cross entropy is 0.

[0065] In the binary classification task, the binary cross entropy loss function is used to measure the difference between the model's prediction results and the true label of the data, which can be expressed as follows:

[0066]

[0067] Where N is the number of pixels in the image, and t i represents the true label of the pixel (0 or 1), P i is the network's prediction for the pixel.

[0068] The Dice Loss function is a metric function used to evaluate the similarity between two samples. Its mathematical expression is:

[0069]

[0070] Among them, |X∩Y| represents the number of elements in the intersection of set X and set Y, and |X| and |Y| represent the number of elements in X and Y respectively. In the semantic segmentation task, Dice Loss can be transformed into the following formula:

[0071]

[0072] Here, ε is a smoothing coefficient used to avoid the denominator being zero.

[0073] The combined loss function BCEDiceLoss formed by adding the two can be written as:

[0074]

[0075] Among them, the smoothing coefficient ε is set as

[0076] Dice loss is more inclined to explore the foreground area. Using the sum of BCELoss and Dice Loss as the loss function of the model can make the model pay attention to the difference between the predicted value and the true label value while also focusing on the extraction of foreground information, which is conducive to alleviating the imbalance problem of positive and negative samples in the crack segmentation task where the foreground accounts for a small proportion.

[0077] Step 3: Propose a method for transfer learning based on small sample infrared features. Since the hollowing infrared images in the dam surface infrared feature simulation images calculated based on the temperature field model come from various sources and have sufficient sample size, this dataset will be used for pre-training. The shallow layer of the convolutional neural network usually extracts structural features such as edges and textures, while the deep layer extracts abstract semantic features. In transfer learning, the shallow features should be mainly transferred, and the model is trained on the concrete dam hollowing measured dataset by freezing the shallow network parameters of the encoder and updating other parameters in the model. The constructed dual attention and multi-scale enhanced feature semantic segmentation model is deployed on the concrete dam hollowing simulated infrared image dataset solved by finite element method, and the pre-training model is trained. The model is then run on the concrete dam hollowing measured dataset, in which some parameters in the encoder network of some models are frozen and do not participate in the gradient update in the network direction propagation algorithm; the unfrozen model parameters are re-trained on the concrete dam hollowing measured dataset to extract the concrete dam hollowing feature information.

[0078] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying hollowing hazards in concrete dams based on small sample infrared feature learning, characterized in that The steps include: (1) Construct a temperature field model of the concrete dam body and hollowing. Calculate the simulated temperature of the dam surface pixel by pixel at different time nodes based on the temperature field model to construct an infrared feature image of the dam body surface. Semantically annotate the hollowing area to form a pre-training dataset for transfer learning. Use the infrared images of concrete dam hollowing collected in the field as a small sample infrared feature learning dataset. (2) Construct a semantic segmentation model with spatial-channel dual attention and multi-scale feature enhancement, including a spatial-channel dual attention encoder, a multi-scale enhanced feature generation module, a spatial-channel dual attention decoder, and a prediction module based on a weighted combination loss function; (3) Construct a spatial-channel dual attention encoder, which includes a 4-layer feature extraction module, each of which contains 2 convolution modules and a spatial-channel dual attention mechanism. Each convolution module contains a 2D convolution layer, a batch normalization layer, and a rectified linear unit layer. The number of convolution kernels is 32, 64, 128, and 128 respectively; (4) Construct a multi-scale enhanced feature generation module, which is located between the encoder and the decoder. It uses the shallow feature information of the encoder layers 1-3 and the context feature information of the 4th layer to enhance the network's adaptability to complex backgrounds and generate enhanced feature maps. (5) Construct a spatial-channel dual attention decoder. This module has a three-layer network structure. Each layer of the network performs bilinear interpolation on the feature map for upsampling. The number of convolution kernels for upsampling in each layer is consistent with that of the encoder network at the same layer, and each layer embeds a spatial and channel dual attention mechanism. (6) Construct a prediction module based on the combined loss function and construct a loss function consisting of the sum of BCE Loss and Dice Loss, denoted as BCEDiceLoss; (7) The constructed semantic segmentation model is pre-trained using the constructed simulation data set, and the model parameters are transferred and used in the learning of the measured dam hollowing infrared features to complete the identification of concrete dam hollowing hazards based on the measured small sample image data set; (8) A method for identifying hollowing hazards in concrete dams based on small sample infrared feature learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded into the processor, implements the method for identifying hollowing hazards in concrete dams according to any one of claims 1 to 7.

2. According to claim 1, a method for identifying hollowing hazards of concrete dams based on small sample infrared feature learning is characterized in that: The simulation data set construction process in step (1) includes: (11) Modeling the dam temperature field: Where T is T(x,0) i.e. the surface temperature of the dam, x is the depth perpendicular to the dam surface, t is the time, and a is the thermal conductivity of the dam material; The initial conditions and boundary conditions for constructing the dam temperature field to obtain the definite solution are: T(x,0)=T0 (2) T(x max ,t)=T0 (3) Where T0 is the initial temperature, x max is the maximum depth at which hidden dangers can be detected using infrared detection technology, h is the comprehensive heat transfer coefficient, is the integrated atmospheric temperature, λ is the thermal conductivity of the concrete dam; (22) By setting different time nodes and hollow depth parameters, the surface temperature of the normal area and the potential hazard area of ​​the concrete dam can be calculated based on the numerical solution of the partial differential equation, thereby obtaining the pixel-by-pixel temperature field of the dam surface at different time points, which is converted into a two-dimensional infrared image as a sample set for pre-training of the semantic segmentation model.

3. According to claim 1, a method for identifying hollowing hazards of concrete dams based on small sample infrared feature learning is characterized in that: The construction of the space-channel dual attention mechanism in step (3) includes: (31) In the channel attention module, a global pooling operation and two 1×1 convolution operations are performed on the feature map, and the feature map is passed to the activation function and mapped to generate a channel weight vector; (32) After the spatial attention module performs a 1×1 convolution operation on the feature map, it is transmitted to the activation function to generate a spatial weight vector; (33) The channel weight vector and the spatial weight vector are multiplied with the feature map respectively and then added together to obtain the weight-corrected channel-spatial dual attention fusion feature map.

4. According to claim 1, a method for identifying hollowing hazards of concrete dams based on small sample infrared feature learning is characterized in that: The multi-scale enhanced feature generation module in step (4) includes: (41) Feature generation consists of two parts: one is that the first three layers of the module are convolved and downsampled to obtain shallow feature information; the other is that the fourth layer generates multi-scale contextual enhanced features based on different hole convolutions; (42) The context enhancement process of the feature map of the 4th layer includes compressing the feature map channel to 32 using 1×1 convolution, extracting multi-scale context features through three dilation convolutions with different dilation degrees, and splicing the channel dimension to obtain a 32-channel feature map again; downsampling the shallow feature maps from the first three layers of the encoder by 8 times, 4 times, and 2 times respectively, and using 1×1 convolution to synchronously adjust the number of channels to 32, splicing and fusion with the context-enhanced 4th layer feature map in the channel dimension, and applying the channel weight vector to the fused feature map to generate a multi-scale enhanced feature map.

5. According to claim 1, a method for identifying hollowing hazards of concrete dams based on small sample infrared feature learning is characterized in that: The loss function BCEDiceLoss in step (6) is formed by adding the BCE Loss function and the Dice Loss function, where the BCE Loss function is a binary cross entropy loss function, and the formula is: Where N is the number of pixels in the image, and t i represents the true label of the pixel (0 or 1), p i is the network's predicted value for the pixel; Dice Loss, also known as the Dice coefficient, is a metric function used to evaluate the similarity between two samples. The larger the value, the more similar the two samples are. It is expressed as: Among them, ε is the smoothing coefficient, which is used to avoid the situation where the denominator is 0; The combined loss function BCEDiceLoss formed by adding the two is: Dice loss tends to explore the foreground area. Using the sum of BCELoss and Dice Loss as the loss function allows the model to focus on the difference between the predicted value and the true label value while also focusing on the extraction of foreground information, which is beneficial to the imbalance problem of positive and negative samples with a small foreground ratio in the hollow drum recognition task.

Images