A method and system for automatically identifying defects of concrete components

Abstract

The present invention relates to the field of automatic identification technology, specifically to a method and system for automatic identification of concrete component defects, comprising the following steps: based on an initialization detection model, the color, illumination and resolution of an input image are standardized, and preliminary crack width and crack length detection is performed by adjusting brightness and contrast parameters. The present invention improves the recognition capability of tiny cracks and defects in concrete components through an attention mechanism and refined image processing, optimizes the color and illumination adjustment of the image, ensures that the best visual data can be obtained under various environmental conditions, and thus makes defect detection more accurate. Through conditional random field technology, the recognition accuracy of crack edges is further enhanced, allowing accurate segmentation of real defect areas in complex backgrounds. Combined with cosine annealing strategy and transfer learning, the model can adapt to different types of concrete components, and improves the application capability and accuracy of the model in new scenarios.

Description

Technical Field

[0001] The present invention relates to the field of automatic identification technology, and in particular to a method and system for automatically identifying defects in concrete components. Background Art

[0002] The field of automatic identification technology can automatically analyze input information from sensors, cameras or other data sources to achieve fast and efficient identification. In practical applications, automatic identification technology is widely used in manufacturing, security monitoring, biometrics, medical diagnosis and traffic management. Its core advantage lies in reducing labor costs, improving processing speed and accuracy, and reducing the possibility of misidentification. With the introduction of algorithms such as deep learning, the application scope and effectiveness of automatic identification technology are constantly expanding and improving.

[0003] Among them, the automatic identification method of concrete component defects focuses on using automatic identification technology to detect and analyze defects in concrete components, such as structural problems such as cracks, holes and falling off. It usually includes using high-resolution cameras or scanning equipment to capture images of the component surface, and then using specific image processing algorithms to analyze the image data and identify existing defects. Its main uses include improving the safety of buildings and infrastructure, and avoiding catastrophic structural failures by detecting potential defects early. In addition, it also helps to adjust maintenance and repair plans, thereby reducing long-term maintenance costs.

[0004] When dealing with cracks and defects in concrete components, existing recognition technologies are limited by the algorithm's generalization ability and adaptability to complex backgrounds. They perform unstable in highly variable environments, especially under different lighting conditions and with large visual interference, and are prone to misjudgment. Most existing methods lack effective real-time learning and adaptation mechanisms and are slow to respond to newly emerging or unrecorded defect types. This lack of flexibility requires retraining when faced with unknown or mutated data, increasing time and economic costs. The lack of an effective feedback adjustment mechanism makes it difficult for recognition methods to update their recognition strategies in real time, resulting in delayed identification of structural problems and increased potential safety risks. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a method and system for automatically identifying defects in concrete components.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a method for automatically identifying defects in concrete components, comprising the following steps:

[0007] S1: Based on the initialization detection model, the color, lighting, and resolution of the input image are standardized. By adjusting the brightness and contrast parameters, preliminary crack width and length detection is performed. The attention mechanism is used to identify cracks and defects in concrete components and generate a crack feature map.

[0008] S2: Based on the crack feature map, a conditional random field is applied to refine the crack edge, adjust the visibility of the boundary between the crack and the component, and integrate the differentiated hierarchical feature map to identify defects of multiple sizes and shapes in the image, generating a structural feature defect dataset;

[0009] S3: Based on the structural feature defect dataset, a cosine annealing strategy is used to dynamically adjust the learning rate parameters, optimize the weight update during the training process, and perform transfer learning on new data of the target concrete type to optimize the accuracy of defect recognition and generate an optimized defect recognition model;

[0010] S4: Based on the optimized defect recognition model, a recursive learning feedback mechanism is implemented to automatically adjust the model's loss function and optimizer parameters according to the comparison between the defect detection results and the current defect data, thereby optimizing the recognition capability of similar defects in the future and generating defect recognition logic results.

[0011] The present invention is improved in that the steps of detecting the preliminary crack width and crack length are specifically as follows:

[0012] S111: Based on the initialization detection model, image processing is performed to convert the input image into grayscale and filter out noise. The local gradient of each pixel is calculated using the formula:

[0013]

[0014] Generate a gradient map where Representative pixel The gradient value of and Representative images in and Pixel difference in direction;

[0015] S112: Based on the gradient map, dynamic threshold segmentation is applied to identify the crack location using the formula:

[0016]

[0017] Get the binary image ,in, and represent the mean and standard deviation of the gradient map, respectively. Representative pixel The gradient value of

[0018] S113: Based on the binarized image, quantitative measurement of crack width and length is performed using the formula:

[0019]

[0020] and

[0021]

[0022] Calculate the width of the crack and length ,in, and Respectively represent and Pixel size in direction, It is a binary image.

[0023] The present invention is improved in that the steps of identifying cracks and defects are specifically as follows:

[0024] S121: Based on the attention mechanism, for potential cracks and defect areas, optimize the feature representation of the area using the formula:

[0025]

[0026] Generate a weighted feature map, where Indicates location The attention-weighted features of Is the image at position The original characteristics of 、 and is the weight parameter, is the bias term, and Respectively indicate the position Raw features of the right and bottom pixels;

[0027] S122: Based on the weighted feature map, key features are extracted, including crack width, length, and shape defect information, using the formula:

[0028]

[0029] Get crack and defect characterization ,in, is the feature data extracted from the image, is the attention weight of each point, Indicates the integration of image pixels.

[0030] The present invention is improved in that the step of thinning the crack edge is specifically as follows:

[0031] S211: Based on the crack feature map, a conditional random field model is applied to perform preliminary edge refinement, and the dependency relationship and edge probability between adjacent pixels are calculated using the formula:

[0032]

[0033] Generate preliminary crack edges ,in, represents the conditional random field model, Represents the crack characteristic map, Represents the parameter settings of the CRF model;

[0034] S212: Based on the preliminary crack edge, the image is integrated with the original image, and the visibility of the boundary between the crack and the component is optimized using the formula:

[0035]

[0036] Get the refined crack edge map ,in, represents the edge sharpening function, represents the initial crack edge, represents the original image, and represent the strength parameters of enhancement and sharpening respectively.

[0037] The present invention is improved in that the steps of acquiring the structural feature defect data set are specifically as follows:

[0038] S221: Integrate the differentiated hierarchical feature map, extract crack features of multiple sizes and shapes, and use linear transformation to standardize feature data using the formula:

[0039]

[0040] Generate feature set ,in, yes The mean of yes The standard deviation of is the total number of pixels in the image, is the refined crack edge map;

[0041] S222: Based on the feature set and in combination with the original image, the formula is used:

[0042]

[0043] Generate structural feature defect dataset ,in, is the feature set, is the original image, is the weight factor.

[0044] The present invention is improved in that the step of adjusting the learning rate parameter is specifically as follows:

[0045] S311: Based on the structural feature defect dataset, a basic learning rate for model training is set using the formula:

[0046]

[0047] Get the base learning rate ,in, is the adjustment coefficient, which is set according to the initial response sensitivity of the model. is the total number of training cycles;

[0048] S312: Based on the basic learning rate, a cosine annealing strategy is applied to dynamically adjust the learning rate using the formula:

[0049]

[0050] Generate an adjusted learning rate where Is the current cycle The learning rate, is the base learning rate, is the current training cycle, is the total number of training cycles, is the learning rate decay exponent;

[0051] S313: Apply the adjusted learning rate to update the weights. The formula is:

[0052]

[0053] Get the updated model weights ,in, is the historical weight, Is the current cycle The learning rate, is the gradient of the loss function.

[0054] The present invention has the following improvements: the steps for optimizing the accuracy of defect recognition are as follows:

[0055] S321: Based on the new data of the target concrete type, determine the basic parameters of the training model using the formula:

[0056]

[0057] Get the initial parameter settings of the model ,in, is the adaptability parameter of the model, is the historical performance index, is the characteristic coefficient of the new data, It is an indicator of data diversity;

[0058] S322: Based on the preliminary parameter settings of the model, the model weight and learning rate are adjusted to match the new concrete type using the formula:

[0059]

[0060] Get the adjusted model parameters ,in, are the model weights, is the learning rate, is the adaptability parameter of the model, is the historical performance index, It is an indicator of data diversity;

[0061] S323: Based on the adjusted model parameters, a validation test is performed on the new concrete data to evaluate the accuracy of the model using the formula:

[0062]

[0063] Get the accuracy evaluation results of the model ,in, are the adjusted model parameters, is the base of natural logarithms.

[0064] The present invention is improved in that the steps of obtaining the defect identification logic result are specifically as follows:

[0065] S411: Based on the optimized defect recognition model, combined with input data through the model, using the formula:

[0066]

[0067] Generate preliminary defect detection results ,in, is the input defect data, is the defect recognition model, is to adjust the zoom factor, is the bias term, is the base of natural logarithm;

[0068] S412: Based on the preliminary defect detection results, the current defect data is compared and the loss function and optimizer parameters of the model are automatically adjusted using the formula:

[0069]

[0070] Get the adjusted optimizer parameters ,in, is the initial parameter, is the preliminary defect detection result, is the current defect data, is the loss function, is the learning rate, is the gradient of the loss function with respect to the parameters;

[0071] S413: Retrain the defect recognition model using the adjusted optimizer parameters to optimize the ability to recognize similar defects in the future, using the formula:

[0072]

[0073] Get defect identification logic results ,in, is the defect recognition model, is the input defect data, are the tuned optimizer parameters.

[0074] A concrete component defect automatic identification system, the system comprising:

[0075] The initialization processing module normalizes the color, lighting, and resolution of the input image based on the initialization detection model, performs image processing by adjusting brightness and contrast parameters, and uses the attention mechanism to identify the preliminary features of cracks and defects to generate a crack feature map.

[0076] The crack feature refinement module uses a conditional random field to refine the crack edge based on the crack feature map, integrates multi-level differential feature maps, and identifies multi-size and multi-morphological defects in the map to generate a structural feature defect dataset;

[0077] The learning rate optimization module applies a cosine annealing strategy based on the structural feature defect dataset, dynamically adjusts the learning rate, updates the weights, and performs transfer learning on new data of the target concrete type to optimize recognition accuracy and generate an optimized defect recognition model;

[0078] The model adjustment module adjusts the loss function parameters of the model based on the optimized defect recognition model and the comparison between the defect detection results and the current defect data, optimizes the model performance, and generates adjusted model parameters;

[0079] The result optimization module optimizes the logic processing of defect identification based on the adjusted model parameters, and automatically adjusts the optimizer parameters to optimize the recognition capability of similar defects in the future and generate defect identification logic results.

[0080] Compared with the prior art, the advantages and positive effects of the present invention are:

[0081] In the present invention, the ability to recognize tiny cracks and defects in concrete components is improved through the attention mechanism and image refinement processing, and the color and lighting adjustment of the image are optimized to ensure that the best visual data can be obtained under various environmental conditions, thereby making defect detection more accurate. The conditional random field technology is used to further enhance the recognition accuracy of crack edges, allowing the accurate segmentation of real defect areas in complex backgrounds. Combined with the cosine annealing strategy and transfer learning, the model can adapt to different types of concrete components, improving the model's application ability and accuracy in new scenarios. The introduction of the recursive learning feedback mechanism allows the model to optimize parameters in real time according to detection feedback, improving the model's prediction ability for new defect types, thereby reducing errors and missed detections, and alleviating dependence on human intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A flowchart of a method for automatically identifying defects in concrete components is proposed in the present invention;

[0083] Figure 2 The present invention proposes a flow chart for detecting preliminary crack width and crack length in a method for automatically identifying defects in concrete components;

[0084] Figure 3 The present invention proposes a flow chart for identifying cracks and defects in a method for automatically identifying defects in concrete components;

[0085] Figure 4 The present invention proposes a flow chart for refining crack edges in a method for automatically identifying defects in concrete components;

[0086] Figure 5 The present invention proposes a flow chart for obtaining a structural feature defect data set in a method for automatically identifying defects in concrete components;

[0087] Figure 6 The present invention proposes a flow chart for adjusting the learning rate parameter in a method for automatically identifying defects in concrete components;

[0088] Figure 7 The present invention proposes a flow chart for optimizing the accuracy of defect recognition in a method for automatically identifying defects in concrete components;

[0089] Figure 8 The present invention proposes a flow chart for obtaining defect recognition logic results in a method for automatic recognition of concrete component defects. DETAILED DESCRIPTION

[0090] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0091] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the devices or elements referred to must have a specific direction, be constructed and operate in a specific direction, and therefore should not be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined. Example

[0092] See also Figure 1 The present invention provides a technical solution: a method for automatically identifying defects in concrete components, comprising the following steps:

[0093] S1: Based on the initialization detection model, the color, lighting, and resolution of the input image are standardized. By adjusting the brightness and contrast parameters, preliminary crack width and length detection is performed. The attention mechanism is used to identify cracks and defects in concrete components and generate a crack feature map. Defects include honeycomb, rough surface, missing edges and corners, rust and spalling, and mold expansion.

[0094] S2: Based on the crack feature map, a conditional random field is applied to refine the crack edge, adjust the visibility of the boundary between the crack and the component, and integrate the differentiated hierarchical feature map to identify defects of multiple sizes and shapes in the image, generating a structural feature defect dataset;

[0095] S3: Based on the structural feature defect dataset, a cosine annealing strategy is used to dynamically adjust the learning rate parameters to optimize weight updates during training. Transfer learning is also performed on new data of the target concrete type to optimize the accuracy of defect recognition and generate an optimized defect recognition model.

[0096] S4: Based on the optimized defect recognition model, a recursive learning feedback mechanism is implemented. According to the comparison between the defect detection results and the current defect data, the model's loss function and optimizer parameters are automatically adjusted to optimize the recognition ability of similar defects in the future and generate defect recognition logic results.

[0097] The crack feature map includes width features, length features, and position identification. The structural feature defect data set includes edge maps, defect shape maps, and defect depth maps. The defect identification logic results include logic adjustment results, parameter update records, and prediction capability improvement results.

[0098] See also Figure 2 , the specific steps for detecting the initial crack width and crack length are:

[0099] S111: Based on the initialization detection model, image processing is performed to convert the input image into grayscale and filter out noise. The local gradient of each pixel is calculated using the formula:

[0100]

[0101] Generate a gradient map where Representative pixel The gradient value of and Representative images in and Pixel difference in direction;

[0102] S112: Based on the gradient map, dynamic threshold segmentation is applied to identify the crack location using the formula:

[0103]

[0104] Get the binary image ,in, and represent the mean and standard deviation of the gradient map, respectively. Representative pixel The gradient value of

[0105] S113: Based on the binary image, quantitative measurement of crack width and length is performed using the formula:

[0106]

[0107] and

[0108]

[0109] Calculate the width of the crack and length ,in, and Respectively represent and Pixel size in direction, It is a binary image.

[0110] and It is obtained by applying the Sobel operator or other difference operators to the original image. The operator can highlight the rapidly changing areas of the image, that is, the edges.

[0111] Assume that the pixel difference in the x direction at a certain point (x, y) is , pixel difference in y direction ,but:

[0112]

[0113] The resulting gradient value indicates that the edge strength at that point is higher.

[0114] and It can be calculated by counting all pixel values ​​of the entire gradient map G(x, y);

[0115] Assume the average value of the gradient map , standard deviation ;

[0116] For the gradient value 18.03 above:

[0117]

[0118] Therefore, this point is considered to be part of the crack.

[0119] and Based on the camera configuration or image resolution setting, assuming a crack spans 10 pixels in width and 20 pixels in length, the actual size of each pixel is:

[0120]

[0121]

[0122]

[0123] It means the total width of the crack is 1 mm and the total length is 2 mm.

[0124] See also Figure 3 , the specific steps for identifying cracks and defects are:

[0125] S121: Based on the attention mechanism, the feature representation of the potential cracks and defect areas is optimized using the formula:

[0126]

[0127] Generate a weighted feature map, where Indicates location The attention-weighted features of Is the image at position The original characteristics of 、 and is the weight parameter, is the bias term, and Respectively indicate the position Raw features of the right and bottom pixels;

[0128] S122: Based on the weighted feature map, key features are extracted, including crack width, length, and shape defect information, using the formula:

[0129]

[0130] Get crack and defect characterization ,in, is the feature data extracted from the image, is the attention weight of each point, Indicates the integration of image pixels, which is usually approximated by summation in actual implementation.

[0131] Assume that in position Pixel grayscale value ;

[0132] Adjacent right pixel ;

[0133] Pixels below ;

[0134] ;

[0135] ;

[0136] ;

[0137] .

[0138] Calculation process:

[0139]

[0140]

[0141]

[0142]

[0143]

[0144] It shows that the attention weight of this point is 45.17, which means that the pixel at this position is given higher importance after feature enhancement.

[0145] Assume the following:

[0146] Assume that ;

[0147] Use the above :

[0148]

[0149]

[0150]

[0151]

[0152] result It represents the comprehensive strength of the feature vector after weighting by the attention mechanism. A higher value indicates that the area has more significant crack or defect characteristics.

[0153] See also Figure 4 , the specific steps of crack edge refinement are:

[0154] S211: Based on the crack feature map, the conditional random field model is applied to perform preliminary edge refinement, and the dependency relationship and edge probability between adjacent pixels are calculated using the formula:

[0155]

[0156] Generate preliminary crack edges ,in, represents the conditional random field model, Represents the crack characteristic map, Represents the parameter settings of the CRF model;

[0157] S212: Based on the preliminary crack edge, it is integrated with the original image and the visibility of the boundary between the crack and the component is optimized using the formula:

[0158]

[0159] Get the refined crack edge map ,in, represents the edge sharpening function, represents the initial crack edge, represents the original image, and represent the strength parameters of enhancement and sharpening respectively.

[0160] Assumptions Obtained by the image edge detection algorithm, such as the Sobel operator;

[0161] Assumptions The values ​​in a small area are [0.1, 0.2, 0.1; 0.2, 0.8, 0.2; 0.1, 0.2, 0.1];

[0162] Assume that there are two weights , and a bias ;

[0163] The CRF model uses the following calculation method:

[0164]

[0165] in, It is a sigmoid function, which is used to map the result to [0, 1] and is suitable for probability expression.

[0166] Calculation example:

[0167]

[0168]

[0169]

[0170] (Use sigmoid function conversion)

[0171] The probability of the crack edge at the center point of the area is 73.3%.

[0172] Assume the following data:

[0173] is 0.733;

[0174] is 150;

[0175] ;

[0176] .

[0177] The enhancement function is defined as:

[0178]

[0179] calculate:

[0180]

[0181]

[0182]

[0183] The results show that after the enhancement process, the pixel value of the area increases, making the crack edge more obvious.

[0184] See also Figure 5 ,The steps for acquiring the structural feature defect dataset are as follows:

[0185] S221: Integrate differentiated hierarchical feature maps to extract crack features of multiple sizes and shapes, and use linear transformation to standardize feature data using the formula:

[0186]

[0187] Generate feature set ,in, yes The mean of yes The standard deviation of is the total number of pixels in the image, is the refined crack edge map;

[0188] S222: Based on the feature set and combined with the original image, the formula is used:

[0189]

[0190] Generate structural feature defect dataset ,in, is the feature set, is the original image, is the weight factor.

[0191] Assumptions The values ​​in a simplified 3x3 area are as follows:

[0192]

[0193] calculate and :

[0194]

[0195] calculate ( ):

[0196]

[0197]

[0198]

[0199]

[0200] calculate :

[0201]

[0202]

[0203]

[0204] result Indicates the uniformity of crack characteristics and the significance of abnormal points after processing.

[0205] Assume the following data:

[0206] Assume it is 100;

[0207] Assume it is 0.5.

[0208] Calculation process:

[0209]

[0210]

[0211]

[0212]

[0213] result Pixel values ​​of a structural feature defect dataset are shown, combining the intensity of the original image and the crack features obtained through analysis to improve the accuracy and visibility of crack detection.

[0214] See also Figure 6 , the specific steps for adjusting the learning rate parameters are:

[0215] S311: Based on the structural feature defect dataset, set the basic learning rate for model training using the formula:

[0216]

[0217] Get the base learning rate ,in, is the adjustment coefficient, which is set according to the initial response sensitivity of the model. is the total number of training cycles;

[0218] S312: Based on the basic learning rate, the cosine annealing strategy is applied to dynamically adjust the learning rate using the formula:

[0219]

[0220] Generate an adjusted learning rate where Is the current cycle The learning rate, is the base learning rate, is the current training cycle, is the total number of training cycles, is the learning rate decay exponent;

[0221] S313: Apply the adjusted learning rate to update the weights. The formula is:

[0222]

[0223] Get the updated model weights ,in, is the historical weight, Is the current cycle The learning rate, is the gradient of the loss function.

[0224] Assumptions ;

[0225] .

[0226] calculate :

[0227]

[0228] This means that the base learning rate is 0.005, which is the step size of each weight update at the beginning of training.

[0229] Assume the following data:

[0230] ;

[0231] ;

[0232] ;

[0233] .

[0234] calculate :

[0235]

[0236] The current learning rate is 0.0028125, which means that at the 100th iteration, the learning rate is reduced to avoid overfitting and optimize the learning process of the model.

[0237] Assumptions ;

[0238] ;

[0239] Current learning rate .

[0240] calculate :

[0241]

[0242] The updated weights show small adjustments, reflecting the effect of fine-grained control of the learning rate during training.

[0243] See also Figure 7 ,The optimization steps for defect recognition accuracy are as follows:

[0244] S321: Based on the new data of the target concrete type, determine the basic parameters of the training model using the formula:

[0245]

[0246] Get the initial parameter settings of the model ,in, is the adaptability parameter of the model, is the historical performance index, is the characteristic coefficient of the new data, It is an indicator of data diversity;

[0247] S322: Based on the preliminary parameter settings of the model, adjust the model weight and learning rate to match the new concrete type using the formula:

[0248]

[0249] Get the adjusted model parameters ,in, are the model weights, is the learning rate, is the adaptability parameter of the model, is the historical performance index, It is an indicator of data diversity;

[0250] S323: Based on the adjusted model parameters, validation tests are performed on new concrete data to evaluate the accuracy of the model using the formula:

[0251]

[0252] Get the accuracy evaluation results of the model ,in, are the adjusted model parameters, is the base of natural logarithms.

[0253] Assume the data is as follows:

[0254] ;

[0255] ;

[0256] ;

[0257] .

[0258] calculate :

[0259]

[0260] result Represents the basic parameter settings of the model, specifically representing the overall adaptability of the model to new data types.

[0261] Assumptions ;

[0262] ;

[0263] The remaining parameters are set as above, and the calculation :

[0264]

[0265] result Represents the adjusted model parameters, reflecting the learning efficiency of the model after the new learning rate and weight adjustment.

[0266] Use the above ;

[0267] calculate :

[0268]

[0269] result Indicates the model's prediction accuracy on new data, with higher values ​​indicating better performance.

[0270] See also Figure 8 ,The specific steps for obtaining the defect identification logic results are:

[0271] S411: Based on the optimized defect recognition model, combined with the input data through the model, the formula is used:

[0272]

[0273] Generate preliminary defect detection results ,in, is the input defect data, is the defect recognition model, is to adjust the zoom factor, is the bias term, is the base of natural logarithm;

[0274] S412: Based on the preliminary defect detection results, compare them with the current defect data and automatically adjust the model's loss function and optimizer parameters using the formula:

[0275]

[0276] Get the adjusted optimizer parameters ,in, is the initial parameter, is the preliminary defect detection result, is the current defect data, is the loss function, is the learning rate, is the gradient of the loss function with respect to the parameters;

[0277] S413: Use the adjusted optimizer parameters to retrain the defect recognition model to optimize the ability to recognize similar defects in the future. The formula is:

[0278]

[0279] Get defect identification logic results ,in, is the defect recognition model, is the input defect data, are the tuned optimizer parameters.

[0280] Assume the data is as follows:

[0281] ;

[0282] .

[0283] Model For a defective data The result of the processing is 2, and the calculation :

[0284]

[0285] result Indicates the perfect response of the model to the current input data.

[0286] Used to calculate the loss function ;

[0287] For evaluation and The error between

[0288] is the loss function About parameters gradient.

[0289] Assumptions:

[0290]

[0291]

[0292]

[0293] calculate :

[0294]

[0295] result Indicates the new value of the parameter after fine-tuning.

[0296] Assume the following data:

[0297] ;

[0298] .

[0299] calculate :

[0300]

[0301] result , indicating that the model’s ability to identify similar defects in the future has been significantly improved.

[0302] A concrete component defect automatic identification system, the system comprising:

[0303] The initialization processing module normalizes the color, lighting, and resolution of the input image based on the initialization detection model, performs image processing by adjusting brightness and contrast parameters, and uses the attention mechanism to identify the preliminary features of cracks and defects to generate a crack feature map.

[0304] The crack feature refinement module uses conditional random fields to refine the crack edges based on the crack feature map, integrates multi-level differential feature maps, and identifies multi-size and multi-morphological defects in the map to generate a structural feature defect dataset.

[0305] The learning rate optimization module applies a cosine annealing strategy based on the structural feature defect dataset, dynamically adjusts the learning rate, updates the weights, and performs transfer learning on new data of the target concrete type to optimize recognition accuracy and generate an optimized defect recognition model.

[0306] The model adjustment module is based on the optimized defect recognition model. According to the comparison between the defect detection results and the current defect data, it adjusts the loss function parameters of the model, optimizes the model performance, and generates the adjusted model parameters.

[0307] The result optimization module optimizes the logical processing of defect identification based on the adjusted model parameters, and automatically adjusts the optimizer parameters to optimize the recognition ability of similar defects in the future and generate defect identification logic results.

[0308] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A method for automatically identifying defects in concrete components, characterized in that: The following steps are involved: Based on the initialization detection model, the color, illumination and resolution of the input image are standardized, and preliminary crack width and crack length detection is performed by adjusting the brightness and contrast parameters. The attention mechanism is used to identify cracks and defects in concrete components and generate crack feature maps. Based on the crack feature map, conditional random fields are applied to refine the crack edge, adjust the boundary visibility between the crack and the component, and integrate the differentiated hierarchical feature map to identify multi-size and morphological defects in the image to generate a structural feature defect dataset; Based on the structural characteristic defect data set, a cosine annealing strategy is adopted to dynamically adjust the learning rate parameters, optimize the weight update during the training process, and perform transfer learning on new data of the target concrete type to optimize the accuracy of defect recognition and generate an optimized defect recognition model; Based on the optimized defect recognition model, a recursive learning feedback mechanism is implemented to automatically adjust the loss function and optimizer parameters of the model according to the comparison between the defect detection results and the current defect data, optimize the recognition ability of similar defects in the future, and generate defect recognition logic results; The steps for detecting the preliminary crack width and crack length are specifically as follows: Based on the initialization detection model, image processing is performed to convert the input image into grayscale and filter out noise. The local gradient of each pixel is calculated using the formula: ； Generate a gradient map where Represents pixel The gradient value of and Representing the image and Pixel difference in direction; Based on the gradient map, dynamic threshold segmentation is applied to identify the crack location using the formula: ； Get the binary image ,in, and Represent the mean and standard deviation of the gradient map, Represents pixel The gradient value of Based on the binary image, the crack width and length are quantitatively measured using the formula: ； and ； Calculate the width of the crack and length ,in, and Respectively represent and The pixel size in the direction, is a binary image; The steps for identifying cracks and defects are specifically as follows: Based on the attention mechanism, the feature representation of the region is optimized for potential cracks and defect regions using the formula: ； Generate a weighted feature map, where Indicates location The attention-weighted features of is the image at position The original characteristics of , and is the weight parameter, is the bias term, and Respectively indicate the location Original features of the right and bottom pixels; Based on the weighted feature map, key features including crack width, length and shape defect information are extracted using the formula: ； Get crack and defect characterization ,in, is the feature data extracted from the image, is the attention weight of each point, Indicates the integration of image pixels; The crack edge refinement step is specifically as follows: Based on the crack feature map, the conditional random field model is applied to perform preliminary edge refinement, calculate the dependency relationship and edge probability between adjacent pixels, and use the formula: ； Generate preliminary crack edges ,in, represents the conditional random field model, Represents the crack characteristic map, Represents the parameter settings of the CRF model; Based on the preliminary crack edge, it is integrated with the original image and the visibility of the boundary between the crack and the component is optimized using the formula: ； Get the refined crack edge map ,in, represents the edge sharpening function, represents the initial crack edge, represents the original image, and represent the intensity parameters of enhancement and sharpening, respectively; The steps for obtaining the structural feature defect data set are specifically as follows: Integrate the differentiated hierarchical feature maps, extract crack features of multiple sizes and shapes, and use linear transformation to standardize feature data using the formula: ； Generate feature set ,in, yes The mean of yes The standard deviation of is the total number of pixels in the image, is the refined crack edge map; Based on the feature set, combined with the original image, the formula is adopted: ； Generate structural feature defect dataset ,in, is the feature set, is the original image, is the weight factor; The adjustment steps of the learning rate parameters are specifically as follows: Based on the structural feature defect data set, the basic learning rate for model training is set using the formula: ； Get the base learning rate ,in, is the adjustment coefficient, which is set according to the initial response sensitivity of the model. is the total number of training cycles; Based on the basic learning rate, the cosine annealing strategy is applied to dynamically adjust the learning rate using the formula: ； Generate an adjusted learning rate where Is the current cycle The learning rate, is the base learning rate, is the current training cycle, is the total number of training cycles, is the learning rate decay exponent; Apply the adjusted learning rate to update the weights, the formula is: ； Get the updated model weights ,in, is the historical weight, Is the current cycle The learning rate, is the gradient of the loss function; The optimization steps for the defect recognition accuracy are specifically as follows: Based on the new data of the target concrete type, the basic parameters of the training model are determined using the formula: ； Get the initial parameter settings of the model ,in, is the fitness parameter of the model, is the historical performance index, is the characteristic coefficient of the new data, It is an indicator of the diversity of the data; Based on the preliminary parameter setting of the model, the model weight and learning rate are adjusted to match the new concrete type, using the formula: ； Get the adjusted model parameters ,in, are the model weights, is the learning rate, is the fitness parameter of the model, is the historical performance index, It is an indicator of the diversity of the data; Based on the adjusted model parameters, validation tests were performed on new concrete data to evaluate the accuracy of the model using the formula: ； Get the accuracy evaluation results of the model ,in, are the adjusted model parameters, is the base of natural logarithms; The steps for obtaining the defect identification logic result are specifically as follows: Based on the optimized defect recognition model, combined with the input data through the model, the formula is adopted: ； Generate preliminary defect detection results ,in, is the input defect data, is the defect recognition model, is to adjust the zoom factor, is the bias term, is the base of natural logarithm; Based on the preliminary defect detection results, compared with the current defect data, the loss function and optimizer parameters of the model are automatically adjusted using the formula: ； Get the adjusted optimizer parameters ,in, is the initial parameter, is the preliminary defect detection result, is the current defect data, is the loss function, is the learning rate, is the gradient of the loss function with respect to the parameters; The defect recognition model is retrained using the adjusted optimizer parameters to optimize the recognition capability of similar defects in the future, using the formula: ； Get defect identification logic results ,in, is the defect recognition model, is the input defect data, are the tuned optimizer parameters.

2. A concrete component defect automatic identification system, characterized in that: According to the method for automatic identification of concrete component defects according to claim 1, the system comprises: The initialization processing module standardizes the color, illumination and resolution of the input image based on the initialization detection model, performs image processing by adjusting the brightness and contrast parameters, and uses the attention mechanism to identify the preliminary features of cracks and defects to generate a crack feature map; The crack feature refinement module uses a conditional random field to refine the crack edge based on the crack feature map, integrates multi-level differential feature maps, and identifies multi-size and multi-morphological defects in the map to generate a structural feature defect data set; The learning rate optimization module applies a cosine annealing strategy based on the structural feature defect data set, dynamically adjusts the learning rate, updates the weights, and performs transfer learning on new data of the target concrete type to optimize recognition accuracy and generate an optimized defect recognition model; The model adjustment module adjusts the loss function parameters of the model based on the optimized defect recognition model and the comparison between the defect detection results and the current defect data, optimizes the model performance, and generates adjusted model parameters; The result optimization module optimizes the logic processing of defect identification based on the adjusted model parameters, and automatically adjusts the optimizer parameters to optimize the recognition capability of similar defects in the future and generate defect identification logic results.

Images