Adhesive Defect Identification Method Based on MCED-YOLOv7 Network

Abstract

This invention provides a method for adhesive defect identification based on an MCED-YOLOv7 network, comprising: acquiring an image to be identified; performing data image processing using linear interpolation; performing image feature processing using an exponentially limited gain algorithm; based on the YOLOv7 architecture, introducing an MLCA attention mechanism after the ELAN module and adding a CBAM attention mechanism to the convolution of the SPPCSPC module; improving the original CIOU to EIOU and adding a DYHEAD target detection head to construct an MCED-YOLOv7 detection network for defect identification of the image to be identified. This invention can achieve accurate detection of normal adhesive inside the insulation board, top voids, and base layer voids, improving the accuracy and robustness of the detection network and significantly reducing false positives and false negatives.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar signal target recognition technology, and in particular to a method for identifying adhesive defects based on the MCED-YOLOv7 network. Background Technology

[0002] Ground penetrating radar (GPR), as an advanced non-destructive testing tool, possesses significant technical advantages such as high efficiency, high anti-interference level, and strong penetration capability, enabling it to efficiently and accurately acquire subsurface images. GPR technology detects and images subsurface structures or objects by emitting electromagnetic waves and receiving reflected signals, providing crucial technical support for fields such as geological exploration, civil engineering, and environmental monitoring. It has been widely used to detect subsurface anomalies. The electromagnetic waves emitted by GPR are reflected after impacting media with different electromagnetic properties, forming a B-scan profile. Therefore, the detection of subsurface anomalies can be considered as the detection of anomalous GPR signals.

[0003] Ground-penetrating radar (GPR) technology plays a crucial role in the target detection of building exterior wall insulation boards and adhesives. Its working principle involves transmitting high-frequency electromagnetic waves to the target area via a transmitting antenna. These waves, propagating underground or within the building, are reflected and refracted by different media (such as insulation boards, concrete, and walls). When these electromagnetic waves encounter the adhesive layer between the insulation board and the wall, the difference in dielectric constant between the two generates a significant reflection signal. These reflected signals are received by a receiving antenna, and after signal processing and interpretation, information about the thickness, continuity, uniformity, and internal structure of the adhesive layer can be obtained to determine if any defects exist. However, current detection of the adhesion status of building exterior wall insulation layers mainly relies on manual interpretation of radar images, which is not only inefficient but also prone to human error. Furthermore, the background of GPR images is complex, making it challenging to accurately determine the adhesion status of normal adhesive, the top separation between the insulation board and the adhesive, and the base separation between the wall and the adhesive. Therefore, how to accurately and automatically detect the adhesion status of building exterior wall insulation layers using GPR technology remains a problem worthy of further research.

[0004] Existing technologies include the Faster R-CNN algorithm for intelligent pipeline target recognition, the YOLOv4 algorithm for road anomaly detection, and the Mask R-CNN algorithm for tunnel lining detection. However, in the field of building exterior wall insulation boards, there is no dedicated algorithm for detecting targets such as normal adhesive, top voids, and base layer voids. The difficulty of the recognition task increases significantly, especially when faced with multiple, scattered, and irregular target features presented in B-scan images of insulation board adhesives. This complexity stems not only from the diversity of target morphology but also from their irregular distribution and potential overlap within the slices, posing a significant challenge to automated recognition.

[0005] Therefore, there is an urgent need to develop a target identification method that can efficiently and accurately handle multiple and dispersed targets such as insulation board adhesive, in order to accurately detect normal adhesive inside the insulation board, top voids, and base layer voids, so as to realize real-time monitoring and evaluation of the bonding quality of building exterior wall insulation layer, ensure the insulation effect and safety of building exterior walls, and improve the quality control level of building construction. Summary of the Invention

[0006] Therefore, it is necessary to provide a method for identifying adhesive defects based on the MCED-YOLOv7 network to address the aforementioned technical problems.

[0007] A method for identifying adhesive defects based on the MCED-YOLOv7 network includes the following steps: acquiring the image to be identified and performing data image processing using linear interpolation; performing image feature processing using an exponentially limited gain algorithm; based on the YOLOv7 architecture, introducing an MLCA attention mechanism after the ELAN module and adding a CBAM attention mechanism to the convolution of the SPPCSPC module; improving the original CIOU to EIOU and adding a DYHEAD target detection head to construct the MCED-YOLOv7 detection network for defect identification of the image to be identified.

[0008] In one embodiment, the exponentially limited gain algorithm includes:

[0009]

[0010] y max =8

[0011] y0(y0>y max )=y max

[0012] Data = Data × y′0

[0013] In the formula, Data is the raw data to be processed, x0 = 1:length(Data) creates a vector x0, whose elements range from 1 to the length of Data, representing the index of the data point, x0 is the time or data point index, and y max Let y0 be the maximum value of the gain function, y0 be the gain function, and y′0 be the gain function after its transpose.

[0014] In one embodiment, the YOLOv7-based architecture, which introduces an MLCA attention mechanism after the ELAN module and a CBAM attention mechanism in the convolution of the SPPCSPC module, includes: introducing cross-layer connections and hybrid convolutions in the ELAN module based on the YOLOv7 architecture, and introducing an MLCA attention mechanism after the ELAN module to combine information from local and global features as well as channel and spatial features to identify feature regions of the target at different scales; adding parallel multiple max pooling operations in the convolution of the SPPCSPC module to extract multi-scale features, and adding a CBAM attention mechanism in the SPPCSPC module to integrate channel attention and spatial attention, dynamically adjusting the weight allocation of different regions according to the feature information of the input image.

[0015] In one embodiment, the improvement of the original CIOU to EIOU and the addition of a DYHEAD target detection head to construct the MCED-YOLOv7 detection network includes: improving the CIOU in the YOLOv7 algorithm to an EIOU loss function; coherently combining a multi-head self-attention mechanism through the DYHEAD target detection head within the scale-aware feature layer, the spatially aware spatial location, and the task-aware output channel; constructing the MCED-YOLOv7 detection network by combining the EIOU loss function and the DYHEAD target detection head with attention; and detecting the image to be identified through the MCED-YOLOv7 detection network to identify adhesive defects.

[0016] Compared to existing technologies, the advantages and beneficial effects of this invention are as follows: By performing data image processing on the acquired image to be identified using linear interpolation, the detail representation of the image can be significantly improved, effectively suppressing image distortion caused by rasterization; the exponentially limited gain algorithm is used for image feature processing, making the features of normal adhesive, top delamination, and base layer delamination in the B-scan image clearer, which is beneficial for subsequent target recognition; based on the YOLOv7 algorithm framework, the MLCA attention mechanism is introduced after the ELAN module, which can better capture and utilize the contextual information in feature maps at different levels, thereby improving the ability to recognize target objects. Furthermore, a CBAM attention mechanism was added to the convolution of the SPPCSPC module to further enhance the selectivity and robustness of feature representation; the original CIOU loss function was improved to EIOU to more accurately describe the overlap between the predicted box and the target box, thereby optimizing the model's localization accuracy; the DYHEAD target detection head was added, which can adaptively adjust its structure and parameters to adapt to target objects of different scales and shapes, thereby improving detection performance. Finally, the MCED-YOLOv7 detection network was constructed to achieve accurate identification of defects in the image to be identified, improving the accuracy and robustness of the detection network and significantly reducing false positives and multiple detections. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an adhesive defect identification method based on the MCED-YOLOv7 network in one embodiment.

[0018] Figure 2(a) is a B-scan slice diagram without interpolation in one embodiment; Figure 2(b) is a B-scan slice diagram using nearest neighbor interpolation; Figure 2(c) is a B-scan slice diagram using natural spline interpolation; Figure 2(d) is a B-scan slice diagram using linear interpolation; Figure 2(e) is a B-scan slice diagram using bicubic interpolation.

[0019] Figure 3(a) is a top-free C-scan ground-penetrating radar image in one embodiment (the arrows point to the B-scan slice imaging direction, and the dashed box represents the top-free adhesive); Figure 3(b) is a B-scan slice image in the X and Y directions as indicated by the down arrow without gain; Figure 3(c) is a B-scan slice image in the X and Y directions as indicated by the down arrow of piecewise constant gain; Figure 3(d) is a B-scan slice image in the X and Y directions as indicated by the arrow of exponential linear combined gain; Figure 3(e) is a B-scan slice image in the X and Y directions as indicated by the down arrow of logarithmically limited gain; Figure 3(f) is a B-scan slice image in the X and Y directions as indicated by the down arrow of exponentially limited gain (the dashed box represents normal adhesive, and the dotted-line box represents top-free).

[0020] Figure 4(a) is a C-scan ground-penetrating radar image of the base layer void in one embodiment (the arrows point to the B-scan slice imaging direction, and the area inside the dashed box is the adhesive for the base layer void); Figure 4(b) is a B-scan slice image in the X and Y directions as indicated by the down arrow without gain; Figure 4(c) is a B-scan slice image in the X and Y directions as indicated by the down arrow of piecewise constant coefficient gain; Figure 4(d) is a B-scan slice image in the X and Y directions as indicated by the arrow of exponential linear combined gain; Figure 4(e) is a B-scan slice image in the X and Y directions as indicated by the down arrow of logarithmically limited gain; Figure 4(f) is a B-scan slice image in the X and Y directions as indicated by the down arrow of exponentially limited gain (the area inside the dashed box is the normal adhesive, and the area inside the dotted-line box is the top void).

[0021] Figure 5(a) shows the actual values ​​of the original YOLOv7 model training in one embodiment; Figure 5(b) shows the predicted values ​​of the original YOLOv7 model training (the red circle indicates multiple detections and false positives).

[0022] Figure 6(a) is a structural diagram of MLAC in one embodiment; Figure 6(b) is a structural diagram of CBAM (including the internal structure of spatial attention and channel attention).

[0023] Figure 7(a) shows the training confusion matrix of the improved model MCED-YOLOv7 in one embodiment; Figure 7(b) shows the training F1 curve of the improved model MCED-YOLOv7; Figure 7(c) shows the training precision curve of the improved model MCED-YOLOv7; Figure 7(d) shows the training recall curve of the improved model MCED-YOLOv7; Figure 7(e) shows the training PR curve of the improved model MCED-YOLOv7; Figure 7(f) shows the curves of various indicators during the training process of the improved model MCED-YOLOv7; Figure 7(g) shows the comparison between the visual prediction results and the true values ​​of the improved model MCED-YOLOv7 (multiple detection is indicated in the red circle). Detailed Implementation

[0024] Before describing the specific embodiments of the present invention, the overall concept of the present invention will be explained as follows:

[0025] This invention is primarily based on the target recognition process of ground-penetrating radar (GPR) signals. Currently, the detection of the bonding status of building exterior wall insulation layers mainly relies on manual interpretation of radar images, which is not only time-consuming but also prone to errors. Furthermore, the background of GPR images is quite complex, making it difficult to accurately determine whether the bonding between the insulation board and the adhesive, or between the wall and the adhesive, is normal or whether there are gaps. Currently, there are no dedicated algorithms for detecting normal adhesive bonding, gaps at the top, and gaps in the base layer. The difficulty of the recognition task increases significantly, especially when faced with multiple, scattered, and irregular target features presented in the B-scan image of the insulation board adhesive. This complexity stems not only from the diversity of target morphology but also from their irregular distribution and potential overlap within the slice, posing a significant challenge to automated recognition.

[0026] Therefore, this invention proposes a method for adhesive defect identification based on the MCED-YOLOv7 network. For defect state detection, linear interpolation is first used in the data image processing stage to improve rasterization, smooth echoes, and highlight details. Then, an improved exponentially limited gain algorithm is used to make features in the B-scan image more obvious. To improve false positives and false negatives, an MLCA attention mechanism is introduced after the ELAN module and a CBAM attention mechanism is added to the convolution after the SPPCSPC module, based on the YOLOv7 algorithm used. Simultaneously, the original CIOU is improved to EIOU, and a DYHEAD target detection head is added, resulting in an improved MCED-YOLOv7 algorithm. The improved MCED-YOLOv7 model achieves a 1.71% improvement in mAP@0.5 compared to the original YOLOv7, significantly reducing false positives and false negatives, verifying the model's feasibility and accuracy, and significantly improving the model's accuracy and efficiency in identifying adhesive defects in insulation boards. The aforementioned testing method can accurately determine whether there is a normal bonding state or voids between the insulation board and the adhesive, as well as between the wall and the adhesive. It can realize real-time monitoring and evaluation of the bonding quality of the building's exterior wall insulation layer, ensure the insulation effect and safety of the building's exterior walls, and improve the quality control level of building construction.

[0027] Having introduced the overall concept of the present invention, to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0028] In one embodiment, such as Figure 1 As shown, a method for identifying adhesive defects based on the MCED-YOLOv7 network is provided, including the following steps:

[0029] Step S101: Obtain the image to be recognized and perform data image processing using linear interpolation.

[0030] Specifically, an image of the insulation board adhesive to be identified was acquired. During the data processing and image optimization stages, linear interpolation technology was employed to improve rasterization and smooth the echo signal. This significantly improved the image's detail and effectively suppressed image distortion caused by rasterization. Linear interpolation calculates the linear relationship between adjacent pixels to perform refined image reconstruction, thereby further highlighting subtle features while maintaining the overall stability of the image structure. The use of linear interpolation to improve rasterization and smooth the echo signal significantly improved the image's detail and effectively suppressed image distortion caused by rasterization.

[0031] A comparison of various interpolation methods is shown in Figure 2. No interpolation, nearest neighbor interpolation, and natural spline interpolation (2D) produce poor results, resembling rasterization. Linear interpolation and bicubic interpolation, however, show more significant results, producing smoother features, more detailed images, and the ability to initially distinguish subtle differences between normal adhesive and top void echoes. Since linear interpolation requires less time than bicubic interpolation while achieving the same effect, linear interpolation yields the best processing results.

[0032] Step S102: Image feature processing is performed using the exponentially limited gain algorithm.

[0033] Specifically, the original piecewise constant gain method was improved into an exponentially limited automatic gain method. Top-free and base-free data were applied, and comparative experiments were conducted on no-gain, piecewise constant-coefficient gain, exponentially linear combination gain, logarithmically limited gain, and exponentially limited gain methods. The results of the comparative experiments are shown in Figures 3 and 4. Figures 3(a) and 4(a) show the C-scan ground-penetrating radar images of top-free and base-free data, respectively. The arrows point to the B-scan slice imaging direction, and the dashed boxes represent the adhesives used in top-free and base-free data.

[0034] The comparative experimental results show that using the exponentially limited gain function makes the features of three situations in the B-scan image clearer: normal bonding, top separation between the adhesive and the insulation board, and separation between the adhesive and the base layer of the wall. The gain is small at the beginning, which suppresses direct wave interference to a certain extent. It gradually increases over time, but will not exceed the set maximum gain multiple. This makes the features of interest more prominent, while reducing the gain of unnecessary parts, which is more conducive to subsequent target recognition.

[0035] The exponentially limiting gain algorithm includes:

[0036]

[0037] y max =8

[0038] y0(y0>y max )=y max

[0039] Data = Data × y′0

[0040] In the formula, Data is the raw data to be processed, x0 = 1:length(Data) creates a vector x0, whose elements range from 1 to the length of Data, representing the index of the data point, x0 is the time or data point index, and y max Let y0 be the maximum value of the gain function, y0 be the gain function, and y′0 be the gain function after its transpose.

[0041] Specifically, The formula for the gain function uses an exponential function with 8 as the base, which determines the initial gain level and the maximum possible gain level (without setting y). max (Under constraints), 1.2 is the growth factor, which determines the rate at which the gain function grows with x0, where x0 is the time or data point index, increasing over time. -2 is the scaling factor used to adjust the scale of x0, thereby controlling the rate of change of the gain function, so that the gain increases slowly in the initial stage and gradually accelerates. This formula means that the gain increases exponentially with time (data point index).

[0042] y max =8 sets the maximum value of the gain function to 8, meaning the gain will not exceed 8 times, which will exceed y. max The value of the element is set to y max This ensures that the gain does not exceed the set maximum value.

[0043] Data = Data × y′0 multiplies the constructed gain function y0 (after transpose) element by element by the original data Data. Each data point in Data will receive a corresponding gain based on its position in time or data point index.

[0044] The principle of this exponentially limited gain algorithm is to construct a time-varying gain function and apply it to the data. The gain function has an exponential growth characteristic, and a maximum gain factor is set to avoid excessive gain. In this way, dynamic gain adjustment of the data can be achieved, so that the gain is small at the beginning and gradually increases over time, but never exceeds the set maximum gain factor. Under the original piecewise constant gain, the B-scan echo image in the X and Y directions has severe spikes, affecting observation. The exponentially limited gain function makes the characteristics of three situations in the B-scan image clearer: normal adhesive, top separation between adhesive and insulation board, and base separation between adhesive and wall. The initial small gain suppresses direct wave interference to a certain extent. As the gain gradually increases over time, but never exceeds the set maximum gain factor, the area of ​​interest is highlighted more, while the gain in unnecessary areas is reduced, which is more conducive to subsequent target identification.

[0045] Step S103: Based on the YOLOv7 architecture, introduce the MLCA attention mechanism after the ELAN module and add the CBAM attention mechanism to the convolution of the SPPCSPC module.

[0046] Specifically, due to the training of the original YOLOv7 model, as shown in Figure 5, the areas circled in red in Figure 5(b) represent false positives and false negatives. Normal adhesive and base layer delamination are easily misclassified, and multiple detection boxes are prone to appear. Therefore, to improve the false positives and false negatives, an MLCA attention mechanism is introduced after the ELAN module and a CBAM attention mechanism is added to the convolution of the SPPCSPC module, based on the YOLOv7 algorithm, to better suit the detection and judgment of the three scenarios of insulation board adhesive.

[0047] Step S103 includes: based on the YOLOv7 architecture, introducing cross-layer connections and hybrid convolutions in the ELAN module, and introducing an MLCA attention mechanism after the ELAN module to combine local and global features as well as channel and spatial features to identify the feature regions of the target at different scales; adding parallel multiple max pooling operations to the convolutions in the SPPCSPC module to extract multi-scale features, and adding a CBAM attention mechanism to the SPPCSPC module to integrate channel attention and spatial attention, and dynamically adjusting the weight allocation of different regions according to the feature information of the input image.

[0048] Specifically, in the YOLOv7 object detection algorithm, the ELAN module is a key component, designed to improve the efficiency and accuracy of the network model in object detection tasks. The ELAN module optimizes the network structure, implementing an efficient layer aggregation method, thereby improving the model's feature extraction capability without significantly increasing computational cost. This module improves the connections within the network, enabling better information flow and fusion between different layers, thus enhancing the model's ability to perceive target features. Furthermore, the ELAN module introduces techniques such as cross-layer connections and hybrid convolutions to further improve the model's performance.

[0049] The SPPCSPC (Spatial Pyramid Pooling Fast with Context-Sensitive Pyramid Convolution) module is an important component of the YOLOv7 network. It primarily addresses the problem of extracting repetitive features from images in convolutional neural networks and reduces or eliminates image distortion that may result from image processing operations. This module incorporates multiple parallel max-pooling operations into a sequence of convolutions, improving the efficiency and accuracy of feature extraction through parallel processing.

[0050] In the YOLOv7 network, the SPPCSPC module is of great significance for extracting multi-scale features, enabling the model to obtain target information at different scales and improving the model's adaptability to changes in target size, so as to facilitate the identification of three types of targets: normal adhesive, top void, and base void.

[0051] MLCA (Mixed Local Channel Attention) is a lightweight attention mechanism that combines local and global features, as well as channel and spatial features. It can significantly improve detection accuracy with a small increase in the number of parameters. Combining local and global feature information helps the model recognize targets at different scales.

[0052] CBAM (Convolutional Block Attention Module) is an efficient and general attention mechanism that integrates channel attention and spatial attention to provide lightweight attention enhancement for the model. It can improve the model's perception ability, enhance the feature representation of different channels, and extract key information at different locations in space.

[0053] To further improve YOLOv7's ability to identify three targets—normal adhesive, top void, and base void—MLCA and CBAM attention mechanisms were added after the ELAN and SPPCSPC modules, respectively. The addition of these two attention mechanisms enables the model to quickly locate key areas in complex input information, thereby improving the model's recognition accuracy and efficiency.

[0054] Adding the MLCA attention mechanism after the ELAN module enables the model to more accurately focus on the feature regions of three targets: normal adhesive, top voids, and base voids. The experimental results of adding different attention mechanisms after the ELAN module are shown in Table 1(a). Adding the MLCA attention mechanism after the ELAN module significantly improves the model compared to adding attention mechanisms such as BiFormer, CBAM, EMA, or GE. The MLCA attention mechanism dynamically adjusts the weight distribution of different regions in the model based on the feature information of the input image, thereby guiding the model to focus more on feature information closely related to the target. In this way, the model can more accurately identify these targets and reduce the probability of false positives and false negatives.

[0055] Table 1(a)

[0056]

[0057] Integrating the CBAM attention mechanism after the SPPCSPC convolutional module allows the model to prioritize regions closely related to the target during multi-scale feature extraction. Experimental results with different attention mechanisms after the SPPCSPC convolutional module are shown in Table 1(b). Adding the CBAM attention mechanism after the SPPCSPC module significantly improves the model compared to adding BiFormer, MLCA, EMA, or GE attention mechanisms. This improvement enables the model to more accurately aggregate target features during feature extraction while reducing attention to background information. Furthermore, the CBAM attention mechanism can dynamically adjust the weight allocation of different regions based on the feature information of the input image, thus enabling the model to more flexibly handle target recognition tasks in various complex scenarios.

[0058] CBAM channel attention focuses on the channel dimension of the feature map, learning the weight of each channel to enhance the representation capability of key channel features. For normal echoes, it emphasizes the continuity and smoothness of channel features. For top void echoes, it emphasizes channel features with weakened enhancement intensity and different layered intensity changes in the depth direction of the hyperbola, with synchronous edge intensity contrast and different hyperbola shapes. For bottom void echoes, it emphasizes channel features with stronger enhancement intensity and different layered intensity changes in the depth direction of the hyperbola, with different edge intensity contrast and different hyperbola shapes.

[0059] CBAM channel attention focuses on the spatial location of the feature map, learning the importance of each location to enhance the response at key spatial locations. For normal echoes, it enhances the response across the entire waveform region, maintaining its continuity and smoothness; for top-void echoes, it pays special attention to areas with weakened intensity and different hyperbolic shapes, suppressing irrelevant areas; for bottom-void echoes, it strengthens locations with varying intensity and different echo shapes, reducing interference from other areas.

[0060] By combining channel and spatial attention in CBAM, channel attention enhances the model's ability to distinguish different echo features (different intensities of layering in the depth direction in B-scan echo maps, and different hyperbola shapes in the three cases) in the channel dimension, while spatial attention improves the model's attention to key features in spatial location (locations with significant intensity changes, the magnitude of intensity differences determining the three cases; differences in the hyperbola itself in the depth direction and differences in contrast with the surrounding background).

[0061] Table 1(b)

[0062]

[0063] To comprehensively evaluate the recognition performance of the YOLOv7 model in three different scenarios, four key metrics were selected: precision, recall, mAP@0.5, and mAP@0.5:0.95 to evaluate the model's performance. These evaluation metrics can be used to measure the accuracy of the model's predictions and the degree of agreement between the predictions and the actual labels.

[0064] For the evaluation metrics of precision and recall, precision measures the proportion of true positive instances among the model's predictions. It reflects the percentage of instances where the model predicts a positive result when it identifies a target. Precision indicates the model's ability to accurately identify true positive instances. Recall, on the other hand, measures the proportion of true positive instances that the model correctly identifies. It reflects the model's ability to capture actual targets; a higher recall indicates that the model misses fewer targets during the identification process.

[0065] In YOLOv7, mAP@0.5, also known as mAP50, is the mAP value calculated when the Intersection over Union (IoU) threshold is set to 0.5. The calculation of mAP@0.5 first involves calculating the AP (Average Precision) for all images of each target category. AP calculation involves interpolating points on the Precision-Recall curve and averaging the results. Theoretically, AP can be calculated through integration, but approximation or interpolation methods are typically used to obtain the average precision. The AP values ​​of all categories are then averaged to obtain mAP@0.5. This metric primarily focuses on the algorithm's performance at an IoU threshold of 0.5 and is suitable for quickly evaluating the algorithm's basic performance.

[0066] The mAP@0.5:0.95 approach is more stringent than mAP@0.5, requiring the algorithm to maintain good performance across multiple IoU thresholds. The mAP@0.5:0.95 calculation design computes mAP at multiple IoU thresholds ranging from 0.5 to 0.95 (with a step size of 0.05) and then averages the results, using the following formula:

[0067]

[0068] In the formula, N is the total amount of data, k is the index of each sample point, Δr(k) = r(k) - r(k-1), C is the total number of categories, and AP i This represents the AP value for the i-th class. This evaluation metric requires the algorithm to accurately predict the target's location and size under different IoU thresholds. Since mAP@0.5:0.95 involves the calculation of multiple IoU thresholds, its result better reflects the algorithm's generalization ability in different scenarios.

[0069] Step S104: Improve the original CIOU to EIOU and add the DYHEAD target detection head to construct the MCED-YOLOv7 detection network, which is used to identify defects in the image to be identified.

[0070] Specifically, in the field of object detection, IoU is used to measure the degree of overlap between the predicted bounding box and the ground truth bounding box. To improve the accuracy and robustness of object detection, the original CIOU of YOLOv7 is improved to EIOU. The EIOU loss function can fully consider the position and feature information of the target bounding box, making the model pay more attention to the accurate position and features of the target, and more accurately describe the degree of overlap between the predicted box and the target box, thereby improving the detection accuracy and optimizing the model's localization accuracy. At the same time, the DYHEAD (attention-based object detection head) of YOLOv7 is introduced. This detection head can adaptively adjust its structure and parameters to adapt to target objects of different scales and shapes, thereby improving detection performance. Finally, the MCED-YOLOv7 detection network is constructed, which is used to accurately identify adhesive defects in the image to be identified.

[0071] Step S104 includes: improving the CIOU loss function in the YOLOv7 algorithm to EIOU; coherently combining a multi-head self-attention mechanism through the DYHEAD target detection head within the scale-aware feature layer, the spatially aware spatial location, and the task-aware output channel; constructing the MCED-YOLOv7 detection network by combining the EIOU loss function and the DYHEAD target detection head with attention; and detecting the adhesive defect in the image to be recognized through the MCED-YOLOv7 detection network.

[0072] In one embodiment, to verify the effectiveness of the EIOU loss function, the original CIOU was replaced with SIOU, EIOU, and WiseIOU respectively. The improved results for identifying three targets—normal adhesive, top void, and base void—are shown in Table 1(c). When improving IOU, EIOU provides a greater improvement to the model compared to CIOU, SIOU, and WiseIOU. When the CIOU loss function in YOLOv7 is changed to the EIOU loss function, the multi-detection and false-detection phenomena of adhesive echoes (echoes of three different cases: normal, top void, and base void) in the ground-penetrating radar B-scan slice image are improved. The EIOU loss function can achieve more refined consideration and processing of echo texture features. The effect of EIOU on these three different echo features:

[0073] To ensure the continuity of normal echoes, the length and width of the hyperbola of normal adhesive are limited to a certain range. EIOU not only considers the intersection-union ratio of the predicted box and the ground truth box, but also the aspect ratio difference between the predicted box and the ground truth box. This helps the model to better capture the continuity and regularity of normal echoes and reduce the situation of misclassifying normal echoes as abnormal echoes.

[0074] For top void echo and base void echo:

[0075] Irregularity Handling: For the irregularities in the intensity of the hyperbola in the depth direction of the top void echo (intensity weakening and strengthening, and the difference in intensity compared to the surrounding area in the depth direction becomes larger), and the irregularities in the shape of the hyperbola in the depth direction (compared to normal adhesive echoes, the hyperbola lines in the top void are thinner, especially in the topmost layer, while the lines at the bottom of the base void are thicker than normal), the aspect ratio difference term in the EIOU loss function can help the model more sensitively identify features. By optimizing the shape and position of the prediction box, the model can more accurately locate the top void region, reducing false positives and false negatives.

[0076] Intensity variation focus: The EIOU loss function also considers the distance between the midpoint of the predicted box and the ground truth box. Although this factor has limited effect in directly detecting hyperbolic intensity variations in B-scan images, it can indirectly focus on the intensity variation between the adhesive echo and the surrounding background echo by optimizing the position of the predicted box, thereby enhancing the ability to identify top void echoes.

[0077] Table 1(c)

[0078]

[0079] To further improve the performance of YOLOv7, DYHEAD was introduced. Based on an attention mechanism, this enhances the model's focus on key feature regions. The training results after introducing the DYHEAD module into the YOLOv7 target detection head are shown in Table 1(d). In the detection of ground-penetrating radar B-scan echo maps, the introduction of DYHEAD in YOLOv7 improves the detection of multiple and false positives in adhesive echoes (including normal, top void, and base void cases). DYHEAD can accurately capture echo texture characteristics and effectively detect different echo features.

[0080] Table 1(d)

[0081]

[0082] The DYHEAD target detection head improves the detection of three types of echoes in the following ways:

[0083] Application of the attention mechanism: DYHEAD significantly improves the representational capability of the target detection head by coherently combining a multi-head self-attention mechanism within scale-aware feature layers, spatially-aware location, and task-aware output channels. The attention mechanism automatically learns and focuses on differences in depth-direction layering capabilities in B-scan echo images, as well as variations in the length and thickness of different hyperbolic protrusion layers. It also addresses the differences in edge intensity contrast with the surrounding area during edge intensity transitions, improving the detection accuracy for different echo features.

[0084] Focus on different echo characteristics: For normal echoes, DYHEAD can focus on their continuity and smoothness, accurately identify and locate normal areas. For top void and base void echoes, DYHEAD can pay special attention to the regularity of the thickness variation of the hyperbolic protruding layer and the characteristics of intensity weakening and strengthening in the waveform, effectively identifying void areas.

[0085] This invention integrates improved modules, combining the introduction of an MLCA attention mechanism after ELAN, the addition of a CBAM attention mechanism after SPPCSPC, an improved EIOU, and the addition of a DYHEAD detection head with attention, to propose the MCED-YOLOv7 (M represents the added MLCA attention mechanism, C represents the CBAM attention mechanism, E represents the improved EIOU, and D represents the added DYHEAD) target detection algorithm for insulation board adhesives. The ablation experiment results for each module of MCED-YOLOv7 are shown in Table 1(e), and the visualization results are shown in Figure 7. Here, "normal" represents normal adhesive, "top" represents top detachment, and "base" represents base layer detachment. Figure 7(g) shows a comparison between the visualized prediction results and the actual values ​​of the improved model MCED-YOLOv7, with multiple detections indicated by red circles. Experimental results show that the improved MCED-YOLOv7 model improves the mAP@0.5:0.95 by 1.71% compared to the original YOLOv7, and the visualization results show a significant reduction in multiple detections and false positives, verifying the effectiveness of the improved model.

[0086] Through the above improvements, the MCED-YOLOv7 model has demonstrated higher accuracy and robustness in target detection tasks. Furthermore, in the visualization results, the false detection and multiple detection phenomena for adhesive echoes (including normal, top void, and base void) have been significantly improved, verifying the effectiveness and practicality of the MCED-YOLOv7 model in the field of target detection.

[0087] Table 1(e)

[0088]

[0089] In this embodiment, linear interpolation is used to process the acquired image to be identified, which significantly improves the detail representation of the image and effectively suppresses image distortion caused by rasterization. The exponentially limited gain algorithm is used for image feature processing, making the features of normal adhesive, top delamination, and base layer delamination in the B-scan image clearer, which is beneficial for subsequent target recognition. Based on the YOLOv7 algorithm framework, an MLCA attention mechanism is introduced after the ELAN module, which can better capture and utilize the contextual information in feature maps at different levels, thereby improving the ability to recognize target objects and in SPPCS. The CBAM attention mechanism is added to the convolution of the PC module to further enhance the selectivity and robustness of feature representation; the original CIOU loss function is improved to EIOU to more accurately describe the overlap between the predicted box and the target box, thereby optimizing the model's localization accuracy; the DYHEAD target detection head is added, which can adaptively adjust its structure and parameters to adapt to target objects of different scales and shapes, thereby improving detection performance. Finally, the MCED-YOLOv7 detection network is constructed to achieve accurate identification of defects in the image to be identified, improving the accuracy and robustness of the detection network and significantly reducing false positives and multiple detections.

[0090] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0091] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a computer storage medium (ROM / RAM, magnetic disk, optical disk) for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, the present invention is not limited to any particular hardware and software combination.

[0092] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for adhesive defect recognition based on MCED-YOLOv7 network, characterized in that, The method comprises the following steps: acquiring an image to be identified, and performing data image processing by using linear interpolation; performing image feature processing by using an exponential limiting gain algorithm; based on a YOLOv7 architecture, introducing an MLCA attention mechanism after an ELAN module, and adding a CBAM attention mechanism in convolution of an SPPCSPC module; improving original CIOU to EIOU, adding a DYHEAD target detection head, and constructing an MCED-YOLOv7 detection network for defect identification of the image to be identified.

2. The adhesive defect identification method based on the MCED-YOLOv7 network according to claim 1, characterized in that, The exponential limiting gain algorithm comprises: y max ＝8 y0 (y0 > y max ) = y max Data = Data x y'0 where Data is the original data to be processed, x0 = 1 : length(Data) creates a vector x0 whose elements range from 1 to the length of Data, indicating the index of the data points, x0 is the time or data point index, y max is the maximum value of the gain function, y0 is the gain function, and y'0 is the transposed gain function.

3. The adhesive defect recognition method based on the MCED-YOLOv7 network according to claim 1, characterized in that, The MLCA attention mechanism introduced after the ELAN module and the CBAM attention mechanism added in the convolution of the SPPCSPC module based on the YOLOv7 architecture comprises: based on the YOLOv7 architecture, introducing cross-layer connection and mixed convolution in the ELAN module, and introducing the MLCA attention mechanism after the ELAN module, combining information of local features and global features, channel features and spatial features, and identifying feature regions of targets at different scales; adding parallel multiple maximum pooling operations in the convolution of the SPPCSPC module for extracting multi-scale features, and adding the CBAM attention mechanism in the SPPCSPC module for integrating channel attention and spatial attention and dynamically adjusting weight distribution of different regions according to feature information of the input image.

4. The adhesive defect identification method based on the MCED-YOLOv7 network according to claim 1, characterized in that, The MCED-YOLOv7 detection network constructed by improving the original CIOU to EIOU and adding the DYHEAD target detection head comprises: improving CIOU in the YOLOv7 algorithm to an EIOU loss function; in a scale-aware feature layer, a space-aware spatial position and a task-aware output channel, combining a DYHEAD target detection head and a multi-head self-attention mechanism; combining the EIOU loss function and the DYHEAD target detection head with attention, and constructing the MCED-YOLOv7 detection network; detecting the image to be identified by using the MCED-YOLOv7 detection network, and identifying adhesive defects.

Images