Bridge underwater structure disease identification model evaluation method, system and equipment

Abstract

The invention relates to the technical field of bridge underwater structure disease detection, in particular to a bridge underwater structure disease recognition model evaluation method, system and equipment, and the method comprises the steps: receiving an apparent disease image of a bridge underwater structure, carrying out the disease pixel semantic segmentation and marking of the image containing the disease, obtaining a marked image, and carrying out the image segmentation and marking; forming a bridge underwater structure apparent disease preliminary data set; dividing the apparent disease preliminary data set into a training data set and a verification data set; according to the invention, from three key angles of disease detection rate, size accuracy and positioning precision, the performance of the disease identification model of the bridge underwater structure disease detection task is comprehensively evaluated, and a single model integration method is combined, so that the influence of data set quality and model training randomness is effectively reduced, and the accuracy of the model is improved. The problem that traditional general indexes are insufficient in adaptability in underwater detection tasks is solved, and an effective, stable and reliable method is provided for disease recognition model performance quantification in bridge underwater structure disease detection.

Description

Technical Field

[0001] The present invention relates to the field of disease identification of bridge underwater structures, and specifically to a method, system and equipment for evaluating a disease identification model of bridge underwater structures. Background Technique

[0002] As a core component of modern transportation infrastructure, the underwater structure of bridges is under the action of a complex water environment for a long time, and is easily affected by multiple factors such as water flow scouring, chemical corrosion, biological attachment and mechanical loads, resulting in diseases such as concrete spalling, steel bar corrosion and crack propagation. Traditional detection mainly relies on manual diving visual inspection or image acquisition by an underwater robot (ROV) equipped with a camera device, which has problems such as high operation risk, low efficiency, strong subjectivity and difficulty in achieving high-frequency monitoring. In recent years, the intelligent identification technology of underwater structure diseases based on computer vision and deep learning has developed rapidly, and the disease characteristics can be automatically identified through object detection and semantic segmentation models, significantly improving the detection efficiency.

[0003] However, the existing technologies mostly focus on the algorithm optimization of the model itself, lack a systematic performance evaluation system for underwater scene characteristics, the evaluation indicators are single, overly rely on general indicators such as accuracy and recall rate, and ignore the impact of special environments such as underwater image blur, low contrast and suspended matter interference on the robustness of the model; the lack of engineering applicability, not combined with the actual requirements of bridge underwater detection tasks, resulting in performance degradation after the model is deployed. Therefore, the present invention proposes a method, system and equipment for evaluating a disease identification model of bridge underwater structures. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system and equipment for evaluating a disease identification model of bridge underwater structures, to fill the blank of the current method for evaluating a disease identification model of bridge underwater structures, and to provide a selection criterion for disease identification models for actual engineering applications.

[0005] According to the first aspect of the present invention, to achieve the above object, the present invention provides the following technical solution: A method for evaluating a disease identification model of bridge underwater structures, the specific steps include:

[0006] Receiving the apparent disease images of the bridge underwater structure, performing disease pixel semantic segmentation annotation on the images containing diseases to obtain annotated images, and constituting a preliminary dataset of the apparent diseases of the bridge underwater structure;

[0007] Dividing the preliminary dataset of apparent diseases into a training dataset and a validation dataset, and expanding the images in the training dataset and the validation dataset by using a sliding window strategy to finally obtain a disease image dataset;

[0008] Constructing disease identification models with a variety of different segmentation models and backbone combinations, and setting the initial values of the hyperparameter combinations of the disease identification models;

[0009] Input the training dataset and the validation dataset into multiple established disease recognition models for training. Adopt the single-model integration strategy to conduct ten parallel trainings on the disease recognition models, and output the disease recognition information.

[0010] Analyze the disease recognition information output from the ten parallel trainings. Conduct performance evaluation based on the predicted diseases and the labeled diseases. Use the log-odds model to fit the disease detection rate model, use the linear model to fit the disease size accuracy, and use threshold transformation to represent the localization accuracy.

[0011] Integrate the results after the ten parallel trainings. Adopt the disease detection rate, size accuracy, and localization accuracy as evaluation indicators to obtain the performance evaluation results of multiple disease recognition models respectively.

[0012] Furthermore, the apparent disease images of the underwater structure of the bridge are obtained by collecting the diseases of full-scale pier components through an industrial binocular camera.

[0013] Furthermore, the disease recognition information output by the disease recognition model is the result after the fusion of ten parallel trainings. The same hyperparameters and datasets are used for parallel training of the same disease recognition model.

[0014] Furthermore, the disease recognition information incorporates the single-model integration strategy. In terms of data sampling, the dependent dataset strategy is used, and all disease recognition models use the same dataset for training, validation, and testing.

[0015] Furthermore, the single-model integration strategy means that in training the baseline classifier, a parallel integration method is used. Multiple identical disease recognition models are trained at the same time, generating data independently, and finally data fusion is performed to obtain the output disease recognition information.

[0016] In the data fusion process, the average voting method is used to perform arithmetic averaging on the evaluation indicators output by multiple models.

[0017] Furthermore, the disease recognition models with various different segmentation models and backbones are respectively UNet-VGG, UNet-Resnet50, PSPNet-MobilenetV2, PSPNet-Resnet50, DeepLabV3Plus-MobilenetV2, HRNet-HRNetV2-W18, HRNet-HRNetV2-W32, HRNet-HRNetV2-W48, SegFormer-Mit-B0, and SegFormer-Mit-B1.

[0018] Furthermore, the disease recognition model training adopts the same initial values of hyperparameter combinations, and the specific settings are as follows:

[0019] The downsampling factor is 8;

[0020] The batch size is set to 16;

[0021] The initial learning rate is 0.0005;

[0022] The training optimizer is the Adam optimizer;

[0023] The momentum is set to 0.9.

[0024] Furthermore, the disease recognition information output from ten parallel trainings is analyzed to obtain the final disease detection rate curve. Performance evaluation is carried out based on the predicted diseases and labeled diseases. The disease detection rate model is fitted using the log-odds model, the disease size accuracy is fitted using a linear model, and the localization accuracy is represented using a threshold transformation. The specific method is as follows:

[0025] (81) Analyze the disease recognition information output from ten parallel trainings to obtain the final disease detection rate curve:

[0026] (81.1) If there is an overlap between the disease area in the predicted image and the disease area in the labeled image, it is determined as detected and 1 is output; if there is no overlap between the disease area in the predicted image and the disease area in the labeled image, it is determined as not detected and 0 is output;

[0027] (81.2) Fit according to the output result in (81.1) using the log-odds model:

[0028]

[0029] In the formula, b0 and b1 are regression coefficients obtained by fitting using the maximum likelihood method, and ln(a) is the natural logarithm of the disease size;

[0030] (81.3) Analyze the results obtained from ten parallel trainings simultaneously according to the analysis in steps (81.1) and (81.2), and take the average of the results obtained from ten fittings to obtain the final disease detection rate curve;

[0031] (82) Fit the disease size accuracy using a linear model, specifically as follows:

[0032] Based on the full detection of the disease recognition information in ten parallel trainings, analyze the accuracy of the predicted disease size. First, draw a scatter plot of the number of pixels in the specific disease area predicted in ten parallel trainings and the number of pixels in the specific disease area in the labeled image, then take the average of the results of ten trainings. Finally, perform a linear fit on the average value to obtain the curve of the predicted disease size accuracy;

[0033] (83) Represent the localization accuracy using a threshold transformation, specifically as follows:

[0034] (83.1) Calculate the centroids of the specific disease regions in the predicted image and the specific disease regions in the label image respectively, as shown in the following formula:

[0035]

[0036]

[0037] where \(x_0, y_0\) are the centroid coordinates, and \(f(x,y)\) is the pixel value of the binary image at the point \((x,y)\);

[0038] (83.2) Calculate the Euclidean distance between the centroid of the labeled disease and the centroid of the predicted disease using the following formula:

[0039]

[0040] where (x0,y0) are the centroid coordinates of the labeled disease and the predicted disease respectively, and AEL is the Euclidean distance between the centroid of the predicted disease and the centroid of the labeled disease;

[0041] (83.3) Use the Heaviside-step function to calculate the localization accuracy, defined as follows:

[0042]

[0043] where \(K\) is the total number of disease samples, and AEL j represents the localization error of the \(j\)th disease, and \(\epsilon\) is the specified threshold;

[0044] (83.4) Take the average of the results obtained from ten parallel trainings to obtain the final localization accuracy curve of the model.

[0045] According to the second aspect of the present invention, the present invention provides a task-oriented evaluation system for the disease recognition model of the underwater structure of a bridge, which is used to implement the method for evaluating the disease recognition model of the underwater structure of a bridge according to any one of claims 1 to 8, including:

[0046] A receiving module, configured to receive the apparent disease images of the underwater structure of the bridge, perform semantic segmentation annotation on the pixels of the disease-containing images to obtain the annotated images, and form a preliminary data set of the apparent diseases of the underwater structure of the bridge;

[0047] A data segmentation and augmentation module: configured to divide the preliminary data set of the apparent disease into a training data set and a validation data set, and use a sliding window strategy to augment the images in the training data set and the validation data set, and finally obtain a disease image data set;

[0048] Model building module: used to build disease recognition models with various different segmentation models and backbone configurations, and set the initial values of the hyperparameter combinations of the disease recognition models;

[0049] Training module: used to input the training dataset and validation dataset into the established multiple disease recognition models for training, perform ten parallel trainings on the disease recognition models using the single model integration strategy, and output disease recognition information;

[0050] Analysis module: used to analyze the disease recognition information output from the ten parallel trainings, perform performance evaluation based on the predicted diseases and labeled diseases, fit the disease detection rate model using the log-odds model, fit the disease size accuracy using the linear model, and represent the localization accuracy using threshold transformation;

[0051] Integration and output module: used to integrate the results after the ten parallel trainings, use the disease detection rate, size accuracy, and localization accuracy as evaluation indicators, and respectively obtain the performance evaluation results of multiple disease recognition models.

[0052] According to the third aspect of the present invention, a terminal device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The feature is that when the processor loads and executes the computer program, the above-mentioned evaluation method for the disease recognition model of the underwater structure of the bridge is adopted.

[0053] The present invention at least has the following beneficial effects:

[0054] 1. By designing a dedicated evaluation framework and multi-dimensional indicators for the disease recognition segmentation model of the underwater structure of the bridge, the present invention solves the problem of insufficient adaptability of traditional general indicators (such as accuracy and recall rate) in underwater detection tasks, and significantly improves the systematicness and engineering guiding value of model performance evaluation.

[0055] 2. The present invention introduces the detection probability theory into the task evaluation of the disease recognition model. Combining the visual analysis of the disease detection rate, it can intuitively quantify the recognition sensitivity of the model to diseases of different sizes, provide data support for optimizing model parameters, and reduce the risk of missed detection.

[0056] 3. Based on the linear fitting method, the present invention quantifies the pixel-level size accuracy of the recognition results, effectively evaluates the measurement error of key disease parameters such as the area of the spalling area by the model, and improves the credibility and practicality of the detection results in the bridge safety rating.

[0057] 4. The present invention adopts a threshold-based localization accuracy analysis method to reveal the stability and anti-interference ability of the model in disease localization in complex underwater images, and provides a reliable basis for formulating subsequent repair plans.

[0058] 5. The present invention integrates multi-dimensional evaluation indicators by constructing a single model set framework, reduces the influence of random factors on the evaluation results during the training process, and ensures the stability of performance evaluation.

[0059] Of course, any product implementing the present invention does not necessarily need to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a schematic flowchart of the evaluation method described in the present invention;

[0061] Figure 2 is a schematic framework principle diagram of the evaluation method described in the present invention;

[0062] Figure 3 is a schematic diagram of the single model integration strategy described in the present invention;

[0063] Figure 4 is a schematic diagram of the result of the disease detection rate described in the present invention;

[0064] Figure 5 is a schematic diagram of the result of the dimension accuracy described in the present invention;

[0065] Figure 6 is a schematic diagram of the result of the positioning accuracy described in the present invention;

[0066] Figure 7 is the network structure diagram of the UNet-VGG model described in the present invention;

[0067] Figure 8 is the network structure diagram of the UNet-Resnet50 model described in the present invention;

[0068] Figure 9 is the network structure diagram of the PSPNet-MobilenetV2 model described in the present invention;

[0069] Figure 10 is the network structure diagram of the PSPNet-Resnet50 model described in the present invention;

[0070] Figure 11 is the network structure diagram of the DeepLabV3 Plus-MobilenetV2 model described in the present invention;

[0071] Figure 12 is the network structure diagram of the HRNet-HRNetV2-W18, HRNet-HRNetV2-W32, HRNet-HRNetV2-W48 models described in the present invention;

[0072] Figure 13The network structure diagrams of the SegFormer-Mit-B0 and SegFormer-Mit-B1 models according to the present invention. Detailed implementation manners

[0073] Next, the technical solutions in the embodiments of the present disclosure will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0074] Embodiment 1:

[0075] Please refer to Figures 1 - 2 , the present invention provides a technical solution: a framework for evaluating the performance of a disease identification model for underwater structures of bridges. The specific steps are as follows:

[0076] S1. Based on the underwater image acquisition platform of the column-supported pier, use a high-precision industrial binocular camera to collect the underwater apparent disease images of full-scale pier components, and form a preliminary dataset of the underwater apparent diseases of the bridge structure.

[0077] S2. Divide the diseases included in the preliminary dataset of the disease images obtained in S1 into one category of peeling, and the other category is the background. Use Labelme software to perform semantic segmentation annotation of the pixels of the diseases in the images containing diseases. After obtaining the annotation file in JSON format, convert it into an image annotation file in PNG format.

[0078] S3. Divide the preliminary dataset of the disease images obtained in S2 into a training dataset and a validation dataset according to a ratio of 7:3. Use the sliding window strategy to expand the images in the training dataset and the validation dataset, and finally form a disease image dataset.

[0079] Among them, the sliding window strategy uses a window of size 1024*1024 and a step size of 512 to expand each image in the training dataset and the validation dataset, and expands the number of images to 54 times the original.

[0080] S4. Establish ten different disease segmentation methods with different segmentation models and skeletons and train them to verify the effectiveness of the evaluation method. Please refer to Figures 7 - 13, Specifically, UNet-VGG uses VGG as the encoder and utilizes the skip connections of U-Net to fuse multi-scale features; UNet-ResNet50 uses ResNet-50 as the encoder and combines the U-Net structure to enhance the feature extraction ability; PSPNet-MobilenetV2 uses MobileNetV2 as the lightweight backbone network and combines the pyramid pooling module of PSPNet to improve the global context awareness; PSPNet-ResNet50 uses ResNet-50 as the backbone and combines the pyramid pooling module (PPM) to enhance the global feature extraction ability; DeepLabV3Plus-MobilenetV2 is based on MobileNetV2 as the lightweight backbone and combines the ASPP and decoder of DeepLabV3+ to enhance the segmentation details; HRNet-HRNetV2-W18: a high-resolution HRNetV2 structure with 18 channels, suitable for scenarios with limited computing resources; HRNet-HRNetV2-W32: an HRNetV2 structure with 32 channels, taking into account both computational efficiency and segmentation accuracy; HRNet-HRNetV2-W48: an HRNetV2 structure with 48 channels, with a large amount of computation but the strongest feature expression ability; SegFormer-Mit-B0: uses the lightweight MiT-B0 as the encoder, without position encoding, and combines Transformer for efficient feature extraction; SegFormer-Mit-B1: based on MiT-B1, with a larger model capacity, enhancing the Transformer encoding ability while maintaining efficient inference;

[0081] S5. Set the initial values of the hyperparameter combinations of the model, input the training dataset and validation dataset obtained in S3 into the established ten different segmentation methods for training. Adopt the single-model ensemble strategy, conduct ten parallel trainings on the same segmentation method, reduce the influence of random factors in the training process on the model evaluation, and output the disease recognition information;

[0082] It should be noted that the disease recognition information output by the disease recognition model is the result after the fusion of ten parallel trainings. For the same disease recognition model, parallel training is carried out using the same hyperparameters and dataset. The disease recognition information incorporates the single-model ensemble strategy. In terms of data sampling, the dependent dataset strategy is used, and all disease recognition models are trained, validated, and tested using the same dataset;

[0083] As Figure 3 shown, the single-model ensemble strategy (method) refers to using the parallel ensemble method in training the baseline classifier, training multiple identical disease recognition models at the same time, generating data independently, and finally performing data fusion to obtain the output disease recognition information;

[0084] In the data fusion process, the average voting method is used to perform arithmetic averaging on the evaluation indicators output by multiple models;

[0085] Among them, the hyperparameters designed according to experience include: the downsampling factor is 8, the batch size is set to 16, the initial learning rate is 0.0005, the training optimizer is the Adam optimizer, and the momentum is set to 0.9;

[0086] S6. Analyze the model disease recognition information, and perform performance evaluation according to the predicted diseases and labeled diseases. Use the log-odds model to fit the disease detection rate model, use the linear model to fit the disease size accuracy, and use the threshold transformation to represent the localization accuracy;

[0087] Among them, the specific calculation steps of the disease detection rate are as follows:

[0088] (S61) If there is an overlap between the disease area in the predicted image and the disease area in the labeled image, it is determined as detected and output 1; if there is no overlap between the disease area in the predicted image and the disease area in the labeled image, it is determined as not detected and output 0;

[0089] (S62) Fit the output result in (S61) using the log-odds model:

[0090]

[0091] Among them, b0 and b1 are regression coefficients obtained by maximum likelihood fitting, ln(a) is the natural logarithm of the disease size, and this model is used to describe the relationship between the disease size and the detection probability;

[0092] (S63) Perform the above analysis on the results obtained from ten parallel trainings at the same time, and take the average of the results obtained from ten fittings to obtain the final disease detection rate curve;

[0093] Among them, the specific calculation of the size accuracy is to perform disease prediction size accuracy analysis on the basis that all diseases are detected in ten parallel trainings. Specifically, first, draw a scatter plot of the number of pixels in the specific disease area predicted in ten parallel trainings and the number of pixels in the specific disease area in the labeled image, then take the average of the results of ten trainings, and finally, perform linear fitting on the average value to obtain the disease prediction size accuracy curve;

[0094] (S64) The specific calculation steps of the localization accuracy are as follows:

[0095] (S64.1) First, calculate the centroids of the specific disease area in the predicted image and the specific disease area in the labeled image respectively, as shown in the following formula:

[0096]

[0097] where \(x_0, y_0\) are the centroid coordinates, and \(f(x, y)\) is the pixel value of the binarized image at the point \((x, y)\);

[0098] (S64.2) Then, calculate the Euclidean distance between the centroid of the labeled disease and the predicted disease using the following formula:

[0099]

[0100] where (\(x_0, y_0\)) are the centroid coordinates of the labeled disease and the predicted disease respectively, and AEL is the Euclidean distance between the centroids of the predicted disease and the labeled disease. By calculating the positioning error of each sample, the positioning accuracy of the model is further evaluated;

[0101] (S64.3) Then, use the Heaviside-step function (step function) to calculate the positioning accuracy rate, which is defined as follows:

[0102]

[0103] where \(K\) is the total number of disease samples, AEL j represents the positioning error of the \(j\)-th disease, and \(\epsilon\) is a specified threshold used to determine whether the error is within an acceptable range;

[0104] (S64.4) Finally, take the average of the results obtained from ten parallel trainings to obtain the final positioning accuracy rate curve of the model;

[0105] S7. Integrate the results after analyzing ten parallel trainings, eliminate the influence of random factors in the training process on the model evaluation, and obtain the final performance of the model in terms of disease detection rate, size accuracy rate, and positioning accuracy rate.

[0106] The specific embodiments are as follows:

[0107] To verify the effectiveness of this method, use an underwater image acquisition platform based on a column-supported pier, and use a high-precision industrial binocular camera to collect full-scale underwater apparent disease images of pier components, constituting a preliminary dataset of underwater apparent diseases of bridge structures with a total of 390 images, including one type of disease, spalling. Label the images, and then divide them into an image training dataset and a validation dataset according to a ratio of 7:3; adopt a sliding window strategy to expand the original dataset, so that the training dataset contains 21,060 book data samples.

[0108] Furthermore, deploy ten different disease segmentation methods with different segmentation models and skeletons, and input the images of the training dataset and the validation dataset into the deployed segmentation methods;

[0109] Furthermore, the single - model integration method is used to fuse the model segmentation results to obtain a more reliable model prediction result;

[0110] Furthermore, the method in S6 is used to analyze the performance of the model in three aspects: disease detection rate, size accuracy, and localization accuracy. The results are respectively as Figure 4 、 Figure 5 and Figure 6 shown.

[0111] The experimental results show that the evaluation method in this embodiment is conducive to starting from the actual engineering tasks of bridge underwater structure apparent disease detection, more reliably evaluating the disease recognition model from different angles, solving the problem of insufficient adaptability of traditional general indicators (such as accuracy and recall rate) in underwater detection tasks, and significantly improving the systematicness and engineering guiding value of model performance evaluation.

[0112] In summary, the technical solution of the present invention mainly optimizes the evaluation method of the bridge underwater structure disease recognition model. Guided by the actual bridge underwater detection task, starting from the detection rate, size accuracy, and localization accuracy, it shows the performance of different models in different tasks in the form of visual images, providing technical support for the intelligent management and maintenance of bridge underwater structures.

[0113] Embodiment 2:

[0114] This embodiment provides a task - oriented bridge underwater structure disease recognition model evaluation system for the bridge underwater structure disease recognition model evaluation method described in Embodiment 1, including:

[0115] A receiving module, used to receive the bridge underwater structure apparent disease images, perform disease pixel semantic segmentation annotation on the images containing diseases to obtain annotated images, and form a preliminary dataset of bridge underwater structure apparent diseases;

[0116] A data segmentation and augmentation module: used to divide the preliminary dataset of apparent diseases into a training dataset and a validation dataset, and use the sliding window strategy to augment the images in the training dataset and the validation dataset, and finally obtain a disease image dataset;

[0117] A model building module: used to construct disease recognition models with multiple different segmentation models and backbone combinations, and set the initial values of the hyperparameter combinations of the disease recognition models;

[0118] A training module: used to input the training dataset and the validation dataset into the established multiple disease recognition models for training, perform ten - time parallel training on the disease recognition models using the single - model integration strategy, and output disease recognition information;

[0119] Analysis module: It is used to analyze the disease recognition information output by ten parallel trainings, evaluate the performance according to the predicted diseases and labeled diseases, fit the disease detection rate model using the log-odds model, fit the disease size accuracy using the linear model, and represent the localization accuracy using threshold transformation;

[0120] Integration and output module: It is used to integrate the results after analyzing ten parallel trainings, and uses the disease detection rate, size accuracy, and localization accuracy as evaluation indicators to obtain the performance evaluation results of various disease recognition models respectively.

[0121] Specifically, the above receiving module, classification and annotation module, data segmentation and augmentation module, model construction module, training module, analysis module, and integration and output module can be embedded in a computer processing system. The computer calls the above modules according to the provided low peak-to-average ratio short synchronization sequence design method to complete the task of evaluating the performance of each disease recognition model; the above receiving module, classification and annotation module, data segmentation and augmentation module, model construction module, training module, analysis module, and integration and output module can perform operations according to the specific steps given by the bridge underwater structure disease recognition model evaluation method.

[0122] It should be noted that it should be understood that the division of each module of the above system is only a logical function division. In actual implementation, it can be fully or partially integrated into a physical entity, or physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also all be implemented in the form of hardware; it is also possible that some modules are implemented in the form of software called by processing elements, and some modules are implemented in the form of hardware. For example, the receiving module can be a separately established processing element, or can be integrated in a certain chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a certain processing element of the above device to perform the functions of the above signal processing module. The implementation of other modules is similar. In addition, all or part of these modules can be integrated together or independently implemented. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed by the integrated logic circuit in the processor element or the instruction in the form of software.

[0123] For example, the above modules can be one or more integrated circuits configured to implement the above methods. For example: one or more Application Specific Integrated Circuits (ASICs), or one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs), etc. Again, when a certain module above is implemented in the form of a processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processors that can call program code. Again, these modules can be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0124] Embodiment 3:

[0125] According to another aspect of the present invention, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The computer program stored in the memory is capable of running on the processor. When the processor loads and executes the computer program, the bridge underwater structure disease identification model evaluation method described in Embodiment 1 is adopted.

[0126] Among them, the computer program can be stored in a computer-readable medium. The computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some middleware form, etc. The computer-readable medium includes any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above components.

[0127] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0128] For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. When an element is referred to as being "assembled on", "mounted on", "fixed to" or "disposed on" another element, it can be directly on the other element or there may also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "upper", "lower", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation.

[0129] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

[0130] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

Claims

1. Evaluation method for bridge underwater structure disease identification model, characterized in that The specific steps include: Receiving the apparent disease images of the underwater structure of the bridge, performing semantic segmentation annotation on the pixels of the images containing diseases to obtain the annotated images, and constituting the preliminary dataset of the apparent diseases of the underwater structure of the bridge; Dividing the preliminary dataset of the apparent diseases into a training dataset and a validation dataset, and using a sliding window strategy to augment the images in the training dataset and the validation dataset, finally obtaining the disease image dataset; Constructing disease recognition models with various different segmentation models and backbone combinations, and setting the initial values of the hyperparameter combinations of the disease recognition models; Inputting the training dataset and the validation dataset into the established multiple disease recognition models for training, using a single-model integration strategy to perform ten parallel trainings on the disease recognition models, and outputting the disease recognition information; Analyzing the disease recognition information output by the ten parallel trainings, performing performance evaluation according to the predicted diseases and the labeled diseases, using the log-odds model to fit the disease detection rate model, using the linear model to fit the disease size accuracy, and using the threshold transformation to represent the localization accuracy; Integrating the results after the ten parallel trainings, using the disease detection rate, the size accuracy, and the localization accuracy as evaluation indicators, and respectively obtaining the performance evaluation results of the multiple disease recognition models.

2. The method for evaluating a bridge underwater structure disease identification model according to claim 1, wherein The apparent disease images of the underwater structure of the bridge are obtained by collecting the diseases of the full-scale pier components through an industrial binocular camera.

3. The method for evaluating a bridge underwater structure disease identification model according to claim 2, wherein: The disease recognition information output by the disease recognition model is the result after the fusion of ten parallel trainings, and the same hyperparameters and dataset are used for parallel training of the same disease recognition model.

4. The method for evaluating a bridge underwater structure disease identification model according to claim 3, characterized in that, The disease recognition information integrates the single-model integration strategy. In terms of data sampling, the dependent dataset strategy is used, and all disease recognition models use the same dataset for training, validation, and testing.

5. The evaluation method for the bridge underwater structure disease identification model according to claim 4, wherein The single-model integration strategy refers to using a parallel integration method in training the baseline classifier, training multiple identical disease recognition models at the same time, generating data independently, and finally performing data fusion to obtain the output disease recognition information; The average voting method is used in the data fusion process to perform arithmetic averaging on the evaluation indicators output by multiple models.

6. The method for evaluating a task-oriented bridge underwater structure disease identification model according to claim 5, wherein The disease recognition models with various different segmentation models and backbone combinations are respectively UNet-VGG, UNet-Resnet50, PSPNet-MobilenetV2, PSPNet-Resnet50, DeepLabV3 Plus-MobilenetV2, HRNet-HRNetV2-W18, HRNet-HRNetV2-W32, HRNet-HRNetV2-W48, SegFormer-Mit-B0, and SegFormer-Mit-B1.

7. The method for evaluating the semantic segmentation model of the task-oriented underwater bridge structure according to claim 1, characterized in that, All disease recognition models are trained with the same initial values of the hyperparameter combinations, and the specific settings are as follows: The downsampling factor is 8; The batch size is set to 16; The initial learning rate is 0.0005; The training optimizer is the Adam optimizer; The momentum is set to 0.

9.

8. The method for evaluating the bridge underwater structure disease identification model according to claim 7, wherein, Analyze the disease recognition information output by ten parallel trainings to obtain the final disease detection rate curve. Conduct performance evaluation based on the predicted diseases and labeled diseases. Fit the disease detection rate model using the log-odds model, fit the disease size accuracy using a linear model, and represent the localization accuracy using threshold transformation. The specific methods are as follows: (81) Analyze the disease recognition information output by ten parallel trainings to obtain the final disease detection rate curve: (81.1) If there is an overlap between the disease area in the predicted image and the disease area in the labeled image, it is determined as detected and output 1; if there is no overlap between the disease area in the predicted image and the disease area in the labeled image, it is determined as not detected and output 0. (81.2) Fit according to the output results in (81.1) using the log-odds model: In the formula, b0 and b1 are regression coefficients obtained by maximum likelihood fitting, and ln(a) is the natural logarithm of the disease size; (81.3) Simultaneously perform the analysis in steps (81.1) and (81.2) on the results obtained from ten parallel trainings, and take the average of the results obtained from ten fittings to obtain the final disease detection rate curve; (82) Fit the disease size accuracy using a linear model, specifically as follows: Based on the complete detection of the disease recognition information in ten parallel trainings, conduct an analysis of the disease prediction size accuracy. First, plot a scatter diagram of the number of pixels in the specific disease area predicted in ten parallel trainings and the number of pixels in the specific disease area in the labeled image, then take the average of the results of ten trainings. Finally, perform linear fitting on the average value to obtain the disease prediction size accuracy curve; (83) Represent the localization accuracy using threshold transformation, specifically as follows: (83.1) Calculate the centroids of the specific disease area in the predicted image and the specific disease area in the labeled image respectively, as shown in the following formula: Where x0, y0 are the centroid coordinates, and f(x, y) is the pixel value of the binary image at the point (x, y); (83.2) Calculate the Euclidean distance between the centroids of the labeled disease and the predicted disease using the following formula: Among them (x0, y0) are the centroid coordinates of the labeled disease and the predicted disease respectively, and AEL is the Euclidean distance between the centroids of the predicted disease and the labeled disease; (83.3) Calculate the localization accuracy using the Heaviside-step function, defined as follows: where K is the total number of disease samples, AEL j represents the positioning error of the j-th disease, and ∈ is a specified threshold; (83.4) Take the average of the results obtained from ten parallel trainings to obtain the final localization accuracy curve of the model.

9. A task-oriented evaluation system for the disease identification model of bridge underwater structures, which is used to implement the evaluation method for the disease identification model of bridge underwater structures described in any one of claims 1 to 8, and is characterized in that: Including: A receiving module, used to receive the apparent disease images of the underwater structure of the bridge, perform semantic segmentation annotation on the pixels of the images containing diseases to obtain the annotated images, and form a preliminary dataset of the apparent diseases of the underwater structure of the bridge; A data segmentation and augmentation module: used to divide the preliminary dataset of apparent diseases into a training dataset and a validation dataset, and use a sliding window strategy to augment the images in the training dataset and the validation dataset to finally obtain a disease image dataset; A model building module: used to construct disease recognition models with various different segmentation models and backbone combinations, and set the initial values of the hyperparameter combinations of the disease recognition models; Training module: It is used to input the training data set and the validation data set into multiple established disease recognition models, perform ten parallel trainings on the disease recognition models using the single model integration strategy, and output disease recognition information; Analysis module: It is used to analyze the disease recognition information output by the ten parallel trainings, conduct performance evaluation based on the predicted diseases and the labeled diseases, fit the disease detection rate model using the log-odds model, fit the disease size accuracy using the linear model, and represent the localization accuracy using the threshold transformation; Integration and output module: It is used to integrate the results after the ten parallel trainings, use the disease detection rate, size accuracy, and localization accuracy as evaluation indicators, and respectively obtain the performance evaluation results of multiple disease recognition models.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor loads and executes the computer program, it adopts the evaluation method for the underwater structure disease recognition model of the bridge described in any one of claims 1 to 8.

Images