Adhesive defect identification method based on MCED-YOLOv7 network

Abstract

The invention provides an adhesive defect identification method based on an MCED-YOLOv7 network, and the method comprises the steps: obtaining a to-be-identified image, and carrying out the data image processing through the linear interpolation; carrying out image feature processing by adopting an exponential amplitude limiting gain algorithm; based on a YOLOv7 architecture, an MLCA attention mechanism is introduced behind an ELAN module, and a CBAM attention mechanism is added to the convolution of an SPPCSPC module; an original CIOU is improved into an EIOU, a DYHEAD target detection head is added, and an MCED-YOLOv7 detection network is constructed and used for carrying out defect identification on an image to be identified. According to the invention, accurate detection of a normal adhesive, top void and base void in the insulation board can be realized, the accuracy and robustness of the detection network are improved, and the phenomena of false detection and excessive detection are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ground penetrating radar signal target recognition, and in particular to an adhesive defect recognition method based on a MCED-YOLOv7 network. Background Art

[0002] As an advanced non-destructive testing tool, Ground Penetrating Radar (GPR) has significant technical advantages such as high efficiency, high anti-interference level and strong penetration ability, and can obtain underground images efficiently and accurately. GPR technology realizes the detection and imaging of underground structures or objects by emitting electromagnetic waves and receiving underground reflected signals, providing important technical support for geological exploration, civil engineering, environmental monitoring and other fields. It has been widely used to detect abnormal objects underground. The electromagnetic waves emitted by the ground penetrating radar are reflected to form a B-scan profile after hitting the medium with different electromagnetic properties. Therefore, the detection of abnormal objects underground can be considered as the detection of abnormal GPR signals.

[0003] Ground penetrating radar technology plays an important role in the detection of building exterior wall insulation boards and adhesives. Its working principle is to transmit high-frequency electromagnetic waves to the target area through the transmitting antenna. When these electromagnetic waves propagate underground or inside the building, they will be reflected and refracted by different media (such as insulation boards, concrete, walls, etc.). When the electromagnetic waves encounter the adhesive layer between the insulation board and the wall, due to the difference in dielectric constant between the two, obvious reflection signals will be generated. After these reflection signals are received by the receiving antenna, after signal processing and interpretation, information about the thickness, continuity, uniformity, internal structure and other information of the adhesive layer can be obtained to determine whether it has defects. However, at present, the detection of the bonding state of the building exterior wall insulation layer mainly relies on manual interpretation of radar images, which is not only inefficient, but also easy to introduce human errors; at the same time, the background of the ground penetrating radar image is relatively complex, and it becomes a challenge to accurately judge the state of normal adhesive, the top hollowing between the insulation board and the adhesive, and the base hollowing between the wall and the adhesive. Therefore, how to use ground penetrating radar technology to accurately and automatically detect the bonding state of the building exterior wall insulation layer is still a problem worth studying.

[0004] In the prior art, there are the Faster R-CNN algorithm for intelligent identification of pipeline targets, 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 are no special algorithms for detecting normal adhesives, top voids, and base voids. In particular, when faced with multiple, scattered, and irregular target features presented in the B-scan image of the insulation board adhesive, the difficulty of the recognition task increases significantly. This complexity is not only due to the diversity of target morphology, but also due to their irregular distribution in the slices and possible overlap, which poses a great challenge to automated recognition.

[0005] Therefore, it is urgent to develop a target recognition method that can efficiently and accurately handle multiple and scattered targets such as insulation board adhesives, so as to accurately detect normal adhesives inside the insulation boards, top voids, and base voids, and realize real-time monitoring and evaluation of the bonding quality of the building exterior wall insulation layer, ensure the insulation effect and safety of the building exterior wall, and improve the quality control level of construction. Summary of the invention

[0006] Based on this, it is necessary to provide an adhesive defect recognition method based on the MCED-YOLOv7 network to address the above technical problems.

[0007] An adhesive defect recognition method based on MCED-YOLOv7 network includes the following steps: obtaining an image to be recognized, and using linear interpolation to perform data image processing; using an exponential limiting gain algorithm to process image features; based on the YOLOv7 architecture, introducing an MLCA attention mechanism after an ELAN module, and adding a CBAM attention mechanism to the convolution of an SPPCSPC module; improving the original CIOU to EIOU, and adding a DYHEAD target detection head to construct a MCED-YOLOv7 detection network for defect recognition of the image to be recognized.

[0008] In one embodiment, the exponential clipping 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 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 point, x0 is the time or data point index, and y max is the maximum value of the gain function, y0 is the gain function, and y′0 is the transposed gain function.

[0014] In one of the embodiments, based on the YOLOv7 architecture, the MLCA attention mechanism is introduced after the ELAN module, and the CBAM attention mechanism is added to the convolution of the SPPCSPC module, including: based on the YOLOv7 architecture, cross-layer connection and hybrid convolution are introduced in the ELAN module, and the MLCA attention mechanism is introduced after the ELAN module, combining local features and global features as well as channel and spatial feature information to identify feature areas of the target at different scales; adding multiple parallel maximum pooling operations to the convolution of the SPPCSPC module to extract multi-scale features, and adding the CBAM attention mechanism to the SPPCSPC module to integrate channel attention and spatial attention, and dynamically adjust the weight distribution of different areas according to the feature information of the input image.

[0015] In one of the embodiments, the original CIOU is improved to EIOU, and a DYHEAD target detection head is added to construct a MCED-YOLOv7 detection network, including: improving the CIOU in the YOLOv7 algorithm to an EIOU loss function; in the scale-aware feature layer, the spatial position of spatial perception, and the output channel of task perception, a multi-head self-attention mechanism is coherently combined through the DYHEAD target detection head; combining the EIOU loss function and the DYHEAD target detection head with attention to construct a MCED-YOLOv7 detection network; and detecting the image to be identified through the MCED-YOLOv7 detection network to identify adhesive defects.

[0016] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: by performing data image processing on the acquired image to be identified through linear interpolation, the detail expression of the image can be significantly improved, and the image distortion caused by rasterization can be effectively suppressed; the exponential limiting gain algorithm is used for image feature processing, so that the features of the three situations of normal adhesive, top void and base void in the B-scan image are clearer, which is conducive to 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 context information in the feature maps of different levels, thereby improving the recognition ability of the target object. , and added the CBAM attention mechanism to the convolution of the SPPCSPC module to further enhance the selectivity and robustness of the feature representation; the original CIOU loss function was improved to EIOU to more accurately describe the overlap between the prediction box and the target box, thereby optimizing the positioning accuracy of the model; the DYHEAD target detection head was added to adaptively adjust its structure and parameters to adapt to target objects of different scales and shapes, thereby improving the detection performance. Finally, the MCED-YOLOv7 detection network was constructed to achieve accurate recognition of image defects to be identified, which improved the accuracy and robustness of the detection network and significantly improved the false detection and multiple detection phenomena. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of an adhesive defect recognition method based on 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] FIG3(a) is a C-scan ground penetrating radar image of a top hollowing out in one embodiment (the arrow points to the B-scan slice imaging direction, and the dotted frame is the adhesive with the top hollowing out); FIG3(b) is a B-scan slice image in the X and Y directions as indicated by the arrows under the no-gain; FIG3(c) is a B-scan slice image in the X and Y directions as indicated by the arrows under the piecewise constant coefficient gain; FIG3(d) is a B-scan slice image in the X and Y directions as indicated by the arrows under the exponential linear combination gain; FIG3(e) is a B-scan slice image in the X and Y directions as indicated by the arrows under the logarithmic limiting gain; and FIG3(f) is a B-scan slice image in the X and Y directions as indicated by the arrows under the exponential limiting gain (the dotted frame is the normal adhesive, and the dotted frame is the top hollowing out).

[0020] FIG4(a) is a C-scan ground penetrating radar image of base layer voids in one embodiment (the arrow points to the B-scan slice imaging direction, and the dotted frame is the adhesive of the base layer voids); FIG4(b) is a B-scan slice image in the X and Y directions as indicated by the arrows under no gain; FIG4(c) is a B-scan slice image in the X and Y directions as indicated by the arrows under the piecewise constant coefficient gain; FIG4(d) is a B-scan slice image in the X and Y directions as indicated by the arrows under the exponential linear combination gain; FIG4(e) is a B-scan slice image in the X and Y directions as indicated by the arrows under the logarithmic limiting gain; and FIG4(f) is a B-scan slice image in the X and Y directions as indicated by the arrows under the exponential limiting gain (the dotted frame is the normal adhesive, and the dotted frame is the top voids).

[0021] FIG. 5( a ) is a diagram showing the true value of the original YOLOv7 model training in one embodiment; FIG. 5( b ) is a diagram showing the predicted value result of the original YOLOv7 model training (the red circle indicates multiple false positives).

[0022] FIG6(a) is a structural diagram of MLAC in one embodiment; FIG6(b) is a structural diagram of CBAM (including the internal structures of spatial attention and channel attention).

[0023] Figure 7(a) is a training confusion matrix diagram of the improved model MCED-YOLOv7 in one embodiment; Figure 7(b) is a training F1 curve diagram of the improved model MCED-YOLOv7; Figure 7(c) is a training Precision curve diagram of the improved model MCED-YOLOv7; Figure 7(d) is a training Recall curve diagram of the improved model MCED-YOLOv7; Figure 7(e) is a training PR curve diagram of the improved model MCED-YOLOv7; Figure 7(f) is a curve diagram of various indicators during the training process of the improved model MCED-YOLOv7; Figure 7(g) is a visual prediction result and true value comparison diagram of the improved model MCED-YOLOv7 (multiple inspections are in the red circle). DETAILED DESCRIPTION

[0024] Before describing the specific embodiments of the present invention, the overall concept of the present invention is described as follows:

[0025] The present invention is mainly developed based on the target recognition process of ground penetrating radar signals. At present, the detection of the bonding state of the building exterior wall insulation layer mainly relies on manual interpretation of radar images, which is not only time-consuming but also prone to errors. At the same time, the background of the ground penetrating radar image is relatively complex. How to accurately judge whether the insulation board and the adhesive, and the wall and the adhesive are in a normal bonding state or there is a void state is still a difficult problem. At present, there is no special algorithm for the detection of normal adhesives, top voids and base voids. In particular, when faced with multiple, scattered and irregular target features presented in the B-scan image of the insulation board adhesive, the difficulty of the recognition task is significantly increased. This complexity is not only due to the diversity of target morphology, but also due to their irregular distribution in the slices and possible overlapping phenomena, which brings great challenges to automatic recognition.

[0026] Therefore, the present invention proposes an adhesive defect recognition method based on the MCED-YOLOv7 network. For the detection of defect status, linear interpolation is first used in the data image processing stage to improve the rasterization phenomenon, smooth the echo and highlight the details; then the improved exponential limiting gain algorithm is used to make the features in the B-scan image more obvious; in order to improve the false detection and multiple detection phenomena, on the basis of the YOLOv7 algorithm used, the MLCA attention mechanism is introduced after the ELAN module, and the CBAM attention mechanism is added to the convolution after the SPPCSPC module; at the same time, the original CIOU is improved to EIOU, and the DYHEAD target detection head is added, and the MCED-YOLOv7 improved algorithm is proposed. The improved MCED-YOLOv7 model is 1.71% higher than the original YOLOv7 mAP@0.5:0.95, and the multiple detection false detection phenomenon is greatly reduced, which verifies the feasibility and accuracy of the model and significantly improves the recognition accuracy and efficiency of the model for insulation board adhesive defects. The above-mentioned detection method can accurately determine whether there is a normal bonding state or void phenomenon between the insulation board and the adhesive, and between the wall and the adhesive. It can realize real-time monitoring and evaluation of the bonding quality of the building exterior wall insulation layer, ensure the insulation effect and safety of the building exterior wall, and improve the quality control level of the construction.

[0027] After introducing the overall concept of the present invention, in order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail by specific implementation methods in combination with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

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

[0029] Step S101, obtaining an image to be recognized, and performing data image processing using linear interpolation.

[0030] Specifically, the image to be identified of the insulation board adhesive is obtained. In the data processing and image optimization stage, the linear interpolation technology is used to improve the rasterization phenomenon and smooth the echo signal, which significantly improves the image detail expression and effectively suppresses the image distortion caused by rasterization. The linear interpolation method reconstructs the image in a refined manner by calculating the linear relationship between adjacent pixels, thereby further highlighting the subtle features of the image while maintaining the stability of the overall image structure. The linear interpolation technology is used to improve the rasterization phenomenon and smooth the echo signal, which significantly improves the image detail expression and effectively suppresses the image distortion caused by rasterization.

[0031] Comparing various interpolation processing methods, the results are shown in Figure 2. No interpolation, nearest neighbor interpolation, and natural spline interpolation (two-dimensional) have poor results, and the results are still similar to rasterization, while linear interpolation and bicubic interpolation have relatively impressive results, with smooth features and prominent details, and can initially distinguish the slight difference between normal adhesive and top empty echo. Since linear interpolation takes less time than bicubic interpolation and can achieve the same effect, linear interpolation can be used to obtain the best processing effect.

[0032] Step S102: using an exponential clipping gain algorithm to perform image feature processing.

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

[0034] According to the comparative experimental results, the use of exponential limiting gain function makes the features of the three situations in the B-scan image, namely normal bonding, top void between adhesive and insulation board, and base void between adhesive and wall, clearer. The gain is small at the beginning, which suppresses the direct wave disturbance under certain circumstances. It gradually increases with time, but will not exceed the set maximum gain multiple, which can make the features of concern more prominent and reduce the gain of unnecessary parts, which is more conducive to subsequent target recognition.

[0035] Among them, the exponential clipping 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 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 point, x0 is the time or data point index, and y max is the maximum value of the gain function, y0 is the gain function, and y′0 is the transposed gain function.

[0041] Specifically, is the formula for the gain function, using an exponential function with 8 as the base to determine the starting level of the gain and the maximum level that can be achieved (without setting y max In the case of restrictions), 1.2 is the growth coefficient, which determines how fast the gain function grows with x0, x0 is the time or data point index, which increases over time, 10 -2 is a scaling factor used to adjust the scale of x0, thereby controlling the rate of change of the gain function so that the gain grows slowly in the initial stage and gradually accelerates. This formula indicates that the gain increases exponentially over time (data point index).

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

[0043] Data=Data×y′0 The constructed gain function y0 (after transposition) is multiplied element by element by the original data Data, and each data point in Data will obtain a corresponding gain according to its position in time or data point index.

[0044] The principle of the exponential clipping gain algorithm is to construct a gain function that changes with time and apply it to the data Data. The gain function has the characteristic of exponential growth, and the maximum gain multiple is set to avoid excessive gain. In this way, dynamic gain adjustment of the data can be achieved, so that the data gain is small at the beginning, and then gradually increases over time, but never exceeds the set maximum gain multiple. The burr phenomenon of the B-scan echo image in the X and Y directions under the original segmented constant coefficient gain is relatively serious, which affects observation; the exponential clipping gain function makes the characteristics of the three situations of normal adhesive, top hollowing between adhesive and insulation board, and base hollowing between adhesive and wall in the B-scan image clearer. The gain is small at the beginning, and direct wave interference is suppressed under certain circumstances. As time gradually increases, but does not exceed the set maximum gain multiple, the area of ​​interest can be made more prominent, and the gain of unnecessary areas becomes smaller, which can be more conducive to subsequent target recognition.

[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 original YOLOv7 model training real value, as shown in Figure 5, the red circle in Figure 5(b) is a multiple false detection. Normal adhesive and base layer voids are both easy to be misjudged, and multiple detection frames are easy to appear. Therefore, in order to improve the false detection and multiple detection phenomena, based on the YOLOv7 algorithm, the MLCA attention mechanism is introduced after the ELAN module, and the CBAM attention mechanism is added to the convolution of the SPPCSPC module, so as to be more suitable for the detection and judgment of the three scenarios of insulation board adhesive.

[0047] Among them, step S103 includes: based on the YOLOv7 architecture, introducing cross-layer connections and mixed convolutions in the ELAN module, and introducing the MLCA attention mechanism after the ELAN module, combining local features and global features as well as channel and spatial feature information to identify the feature areas of the target at different scales; adding multiple parallel maximum pooling operations to the convolution of the SPPCSPC module to extract multi-scale features, and adding the CBAM attention mechanism to the SPPCSPC module to integrate channel attention and spatial attention, and dynamically adjust the weight distribution of different areas according to the feature information of the input image.

[0048] Specifically, in the YOLOv7 target detection algorithm, the ELAN module is a key component, which is designed to improve the efficiency and accuracy of the network model in target detection tasks. The ELAN module optimizes the network structure and implements an efficient layer aggregation method, thereby improving the model's feature extraction capabilities without increasing excessive computational costs. The module improves the connection method in the network so that information between different layers can flow and merge better, thereby improving the model's perception of target features; the ELAN module introduces technical means such as cross-layer connections and hybrid convolutions to further improve the performance of the model.

[0049] The SPPCSPC (Spatial Pyramid Pooling Fast with Context-Sensitive PyramidConvolution) module is an important component of the YOLOv7 network. It is mainly used to solve the problem of extracting repeated features of images in convolutional neural networks and reduce or eliminate image distortion problems that may be caused by image processing operations. This module adds multiple parallel maximum pooling operations to a series of convolutions, and improves 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 voids, and base voids.

[0051] MLCA (Mixed Local Channel Attention) is a lightweight attention mechanism that combines local and global features as well as channel and spatial feature information. It can significantly improve detection accuracy while adding a small amount of parameters. The combination of local and global feature information helps the model identify targets at different scales.

[0052] CBAM (Convolutional Block Attention Module) is an efficient and general attention mechanism that provides a lightweight attention enhancement function for the model by integrating channel attention and spatial attention. It can improve the perception ability of the model, enhance the feature representation of different channels, and extract key information at different positions in space.

[0053] In order to further improve the recognition ability of YOLOv7 for normal adhesive, top void and base void, MLCA and CBAM attention mechanisms were added after the ELAN module and SPPCSPC module respectively. After adding the two attention mechanisms, the model can quickly locate the key areas in complex input information, thereby improving the recognition accuracy and efficiency of the model.

[0054] The MLCA attention mechanism is added after the ELAN module, so that the model can more accurately focus on the feature areas of these targets when dealing with 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). Compared with adding attention mechanisms such as BiFormer, CBAM, EMA, or GE, adding the MLCA attention mechanism after the ELAN module has a greater improvement on the model. The MLCA attention mechanism can dynamically adjust 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 the feature information closely related to the target. In this way, the model can more accurately identify these targets and reduce the probability of false detection and multiple detection.

[0055] Table 1(a)

[0056]

[0057] By integrating the CBAM attention mechanism after the SPPCSPC convolution module, the model can give priority to the areas closely related to the target when performing multi-scale feature extraction. The experimental results of adding different attention mechanisms after the SPPCSPC convolution module are shown in Table 1(b). Compared with adding attention mechanisms such as BiFormer, MLCA, EMA or GE, adding the CBAM attention mechanism after the SPPCSPC module has a greater improvement on the model. This improvement enables the model to more accurately gather target features during feature extraction while reducing the focus on background information. In addition, the CBAM attention mechanism can dynamically adjust the weight distribution of different regions according to the feature information of the input image. Therefore, the model can more flexibly cope with target recognition tasks in various complex scenarios.

[0058] CBAM channel attention focuses on the channel dimension of the feature map and learns the weight of each channel to enhance the representation ability of key channel features. It emphasizes the channel characteristics of continuity and smoothness of normal echoes, and strengthens the channel characteristics of weakened top empty echoes and different changes in the layered intensity in the depth direction of the hyperbola, with synchronized edge intensity contrasts and different hyperbolic shapes. It strengthens the channel characteristics of stronger base empty echoes and different changes in the layered intensity in the depth direction of the hyperbola, with different edge intensity contrasts and different hyperbolic shapes.

[0059] CBAM channel attention focuses on the spatial position of the feature map and learns the importance of each position, thereby enhancing the response of key spatial positions. For normal echoes, the response of the entire waveform area is enhanced to maintain its continuity and smoothness; for top-out echoes, special attention is paid to areas with weakened intensity and different hyperbolic shapes, and irrelevant areas are suppressed; for base-out echoes, positions with intensity changes and different echo shapes are strengthened to reduce interference from other areas.

[0060] Combining the channel and spatial attention of CBAM, the channel attention enhances the model's ability to distinguish different echo features in the channel dimension (different intensities of stratification in the depth direction in the B-scan echo map, and different hyperbolic shapes in the three cases), while the spatial attention improves the model's attention to key features in spatial position (positions with obvious intensity changes, the size of the intensity difference determines the three cases; the differences in the hyperbola itself in the depth direction and the differences in contrast with the surrounding background).

[0061] Table 1(b)

[0062]

[0063] In order to comprehensively evaluate the recognition effect of the YOLOv7 model in three different scenarios, four key indicators, namely precision, recall, mAP@0.5 and mAP@0.5:0.95, are used to comprehensively judge the performance of the model. The above evaluation indicators can be used to measure the accuracy of the model prediction results and the degree of consistency between the actual labels.

[0064] For the two evaluation indicators, precision and recall, precision is an indicator to measure the proportion of true positive samples in the model's prediction results. It reflects the proportion of true positive samples in the instances where the model predicts the results as positive samples when identifying the target. Through precision, we can understand the degree to which the model grasps the true positive samples during the recognition process. Recall measures the proportion of true positive samples that the model can correctly identify among all true positive samples, reflecting the model's ability to capture actual targets. The higher the recall rate, the fewer targets the model misses during the recognition process.

[0065] In YOLOv7, mAP@0.5 is also called mAP50, which is the mAP value calculated when the intersection-over-union ratio threshold is set to 0.5. In the calculation process of mAP@0.5, the AP (Average Precision) of all images of each target category is first calculated. The calculation of AP involves interpolating the points on the Precision-Recall curve and taking the average value. In theory, AP can be calculated by integration, but usually an approximate or interpolation method is used to calculate it. The average precision is obtained, and the AP values ​​of all categories are averaged to get mAP@0.5. This indicator mainly focuses on the performance of the algorithm when the IoU threshold is 0.5, and is suitable for quick evaluation of the basic performance of the algorithm.

[0066] mAP@0.5:0.95 is more stringent than mAP@0.5, requiring the algorithm to maintain good performance under multiple IoU thresholds. The calculation design of mAP@0.5:0.95 calculates mAP at multiple values ​​of IoU threshold from 0.5 to 0.95 (with a step size of 0.05) and averages them. The formula is:

[0067]

[0068] Where N is the total amount of data, k is the index of each sample point, △r(k) = r(k)-r(k-1), C is the total number of categories, AP i Represents the AP value of the i-th category. This evaluation indicator requires the algorithm to accurately predict the location and size of the target under different IoU thresholds. Since mAP@0.5:0.95 involves the calculation of multiple IoU thresholds, its results can better reflect the generalization ability of the algorithm in different scenarios.

[0069] Step S104, the original CIOU is improved to EIOU, and the DYHEAD target detection head is added to construct a MCED-YOLOv7 detection network for defect recognition of the image to be recognized.

[0070] Specifically, in the field of target detection, IoU is used to measure the degree of overlap between the predicted box and the real box. In order to improve the accuracy and robustness of target 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, so that the model pays more attention to the accurate position and features of the target, and more accurately describes the degree of overlap between the predicted box and the target box, thereby improving the detection accuracy and optimizing the positioning accuracy of the model; at the same time, the DYHEAD (target detection head based on attention mechanism) target detection head of YOLOv7 is introduced. The detection head can adaptively adjust the structure and parameters to adapt to target objects of different scales and shapes, thereby improving the detection performance. Finally, the MCED-YOLOv7 detection network is constructed, through which the adhesive defects of the image to be identified are accurately identified.

[0071] Among them, step S104 includes: improving the CIOU in the YOLOv7 algorithm to an EIOU loss function; coherently combining a multi-head self-attention mechanism through a DYHEAD target detection head in the scale-aware feature layer, the spatial position of spatial perception, and the output channel of task perception; combining the EIOU loss function and the DYHEAD target detection head with attention to construct a MCED-YOLOv7 detection network; detecting the image to be identified through the MCED-YOLOv7 detection network to identify adhesive defects.

[0072] In one embodiment, in order to verify the effectiveness of the EIOU loss function, the original CIOU is replaced by SIOU, EIOU and WiseIOU respectively, and the improved results of identifying three targets: normal adhesive, top void and base void are shown in Table 1(c). When improving IOU, EIOU has a greater improvement on the model than CIOU, SIOU and WiseIOU. When the CIOU loss function in YOLOv7 is changed to the EIOU loss function, the multiple detection and false detection phenomena of the adhesive echo (normal, top void, base void in three different situations) in the B-scan slice of the ground penetrating radar are improved, and the EIOU loss function can achieve more detailed consideration and processing of the echo ripple features. The effect of EIOU on the different characteristics of these three echoes:

[0073] Considering the continuity of normal echoes, the length and width of the hyperbola of normal adhesives are limited to a certain range. EIOU not only considers the intersection-over-union ratio of the predicted box and the true box, but also considers the difference in the aspect ratio between the predicted box and the true box, which helps the model better capture the continuity and regularity of normal echoes and reduce the situation where normal echoes are misjudged as abnormal echoes.

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

[0075] Irregularity processing: For the irregularity of the intensity of the hyperbola in the top void echo in the depth direction (intensity weakening and strengthening, the difference in depth direction with the surrounding intensity becomes larger), and the irregularity of the hyperbola shape in the depth direction (compared with the normal adhesive echo, the hyperbola line of the top void is thinner, especially the echo of the top layer, while the bottom line of the base void is thicker than normal), the aspect ratio difference term in the EIOU loss function can help the model identify features more sensitively. By optimizing the shape and position of the prediction box, the model can more accurately locate the top void area and reduce false detection and multiple detection.

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

[0077] Table 1(c)

[0078]

[0079] In order to further improve the performance of YOLOv7, DYHEAD is introduced. Based on the attention mechanism, the model can pay more attention to the key feature areas. After the DYHEAD module is introduced in the target detection head of YOLOv7, the training results are shown in Table 1(d). In the detection of the B-scan echo map of the ground penetrating radar, the introduction of DYHEAD in YOLOv7 can improve the multiple detection and false detection of adhesive echoes (including normal, top empty, and base empty). DYHEAD can accurately capture the characteristics of the echo ripples and effectively detect different echo features.

[0080] Table 1(d)

[0081]

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

[0083] Application of attention mechanism: DYHEAD significantly improves the representation ability of the target detection head by coherently combining the multi-head self-attention mechanism in the feature layer of scale perception, the spatial position of space perception, and the output channel of task perception. The attention mechanism can automatically learn and pay attention to the differences between the depth-direction layering capabilities in the B-scan echo image, the differences between the length, thickness, and thickness of different salient layers of the hyperbola. The edge intensity transition has different features compared with the surroundings, which improves the detection accuracy of different echo features.

[0084] Focus on different echo characteristics: For normal echoes, DYHEAD can focus on its continuity and smoothness characteristics, accurately identify and locate normal areas. For top void and base void echoes, DYHEAD can pay special attention to the thickness change rules and intensity reduction and enhancement characteristics of the hyperbolic protrusion layer in the waveform, and effectively identify the void area.

[0085] The present invention integrates the improved modules, introduces the MLCA attention mechanism after ELAN, adds the CBAM attention mechanism after SPPCSPC, improves EIOU and adds the DYHEAD detection head with attention, and proposes 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) insulation board adhesive target detection algorithm. The ablation experiment results of each module of MCED-YOLOv7 are shown in Table 1(e), and the visualization result diagram is shown in Figure 7, where normal represents normal adhesive, top represents top hollowing, and base represents base hollowing. Figure 7(g) is a comparison diagram of the visualization prediction results and true values ​​of the improved model MCED-YOLOv7, and the red circle represents multiple detections. The experimental results show that the improved MCED-YOLOv7 model is 1.71% higher than the original YOLOv7 mAP@0.5:0.95, and the visualization result has greatly reduced the phenomenon of multiple false detections, which can verify the effectiveness of the improved model.

[0086] Through the above improvements, the MCED-YOLOv7 model has shown higher accuracy and robustness in target detection tasks. In the visualization results, the false detection and multiple detection phenomena of 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 can significantly improve the detail expression of the image and effectively suppress the image distortion caused by rasterization; the exponential limiting gain algorithm is used to process the image features, so that the features of the three situations of normal adhesive, top void and base void in the B-scan image are clearer, which is conducive to 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 context information in the feature maps of different levels, thereby improving the recognition ability of the target object and The CBAM attention mechanism is added to the convolution of the PC module to further enhance the selectivity and robustness of the feature representation; the original CIOU loss function is improved to EIOU to more accurately describe the overlap between the prediction box and the target box, thereby optimizing the positioning accuracy of the model; the DYHEAD target detection head is added to adaptively adjust its structure and parameters to adapt to target objects of different scales and shapes, thereby improving the detection performance. Finally, the MCED-YOLOv7 detection network is constructed to achieve accurate recognition of image defects to be identified, which improves the accuracy and robustness of the detection network and significantly improves the false detection and multiple detection phenomena.

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

[0091] Obviously, those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a computer storage medium (ROM / RAM, magnetic disk, optical disk) and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than that here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Therefore, the present invention is not limited to any specific combination of hardware and software.

[0092] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for identifying adhesive defects based on MCED-YOLOv7 network, characterized in that: The following steps are involved: Get the image to be identified and use linear interpolation to process the data image; The exponential limiting gain algorithm is used to process image features; Based on the YOLOv7 architecture, the MLCA attention mechanism is introduced after the ELAN module, and the CBAM attention mechanism is added to the convolution of the SPPCSPC module; The original CIOU is improved to EIOU, and the DYHEAD target detection head is added to construct the MCED-YOLOv7 detection network for defect recognition of the image to be recognized.

2. The adhesive defect recognition method based on MCED-YOLOv7 network according to claim 1 is characterized in that: The exponential clipping gain algorithm includes: y max ＝8 y0(y0>y max )=y max Data=Data×y′0 In the formula, 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 point, x0 is the time or data point index, and 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 MCED-YOLOv7 network according to claim 1 is characterized in that: Based on the YOLOv7 architecture, the MLCA attention mechanism is introduced after the ELAN module, and the CBAM attention mechanism is added to the convolution of the SPPCSPC module, including: Based on the YOLOv7 architecture, cross-layer connections and mixed convolutions are introduced in the ELAN module, and the MLCA attention mechanism is introduced after the ELAN module to combine local features, global features, channel and spatial features to identify the characteristic areas of the target at different scales; Multiple parallel maximum pooling operations are added to the convolution of the SPPCSPC module to extract multi-scale features. A CBAM attention mechanism is added to the SPPCSPC module to integrate channel attention and spatial attention, and dynamically adjust the weight distribution of different regions according to the feature information of the input image.

4. The adhesive defect recognition method based on MCED-YOLOv7 network according to claim 1 is characterized in that: The original CIOU is improved to EIOU, and the DYHEAD target detection head is added to construct the MCED-YOLOv7 detection network, including: Improve the CIOU in the YOLOv7 algorithm to the EIOU loss function; In the scale-aware feature layer, the spatial position of the space perception, and the task-aware output channel, the DYHEAD object detection head is combined with a multi-head self-attention mechanism; Combining the EIOU loss function and the DYHEAD object detection head with attention, the MCED-YOLOv7 detection network is constructed; The image to be identified is detected by using the MCED-YOLOv7 detection network to identify adhesive defects.

Images