A multi-level concrete crack detection method based on a drone

Abstract

The present application belongs to the field of image data processing, and provides a multi-level concrete crack detection method based on unmanned aerial vehicles. In the acquisition stage, an edge computing platform based on unmanned aerial vehicles and a target detection model is used to perform real-time target detection on the surface of the building structure, so as to realize rapid identification of suspected crack areas and task-driven image acquisition. In the fine detection stage, a multi-scale modeling strategy is used, which integrates a multi-scale perception path of coarse-grained region screening, fine-grained boundary enhancement and post-processing optimization segmentation. A mixed attention mechanism MLCA is introduced to enhance the focusing ability of the model on detailed features and the semantic robustness in complex backgrounds. The overall consideration of the real-time requirements of the inspection task and the accuracy of the pixel-level semantic modeling provides theoretical support and method foundation for intelligent identification of large structure surface cracks in high-resolution scenarios.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image data processing, and particularly relates to a multi-level concrete crack detection method based on a UAV. BACKGROUND

[0002] With the increasing demand for structural life cycle management, health monitoring and early damage identification of large-scale concrete structures such as long-span bridges, high-rise and super high-rise buildings, and port terminals are becoming an important direction in the field of structural engineering. Cracks on the surface of concrete structures, as one of the earliest and most intuitive signs of deterioration, directly threaten the service performance, safety and life cycle management of the structure. Efficient and intelligent detection of cracks in large-scale concrete structures has become a key research topic in the field of structural health monitoring.

[0003] Traditional manual inspection methods generally have high labor intensity, low efficiency and high operation risk, especially in high-altitude operations or complex structural areas. At the same time, for large-scale structure surfaces, it is often difficult to achieve full coverage and continuous detection by relying on scaffolding or manual tools. These limitations have driven the engineering community to demand non-contact, automated and scalable detection methods. In this context, UAVs have become a promising technology path in structural detection due to their flexible deployment, high-resolution imaging capabilities and wide coverage advantages.

[0004] However, current crack detection methods based on UAVs, especially those relying on target detection models such as YOLO or semantic segmentation networks, still face two major challenges in practical applications: first, small cracks are difficult to detect due to their extremely small pixel coverage area and susceptibility to interference from concrete texture, stains or changes in lighting conditions; second, the massive computational burden brought about by high-resolution images makes real-time detection difficult, especially in scenarios where UAV endurance time and computing resources are limited.

[0005] Therefore, to address the rapid detection and fine segmentation of large-scale concrete infrastructure, a new crack perception framework for engineering applications is needed to overcome the limitations of existing detection methods that struggle to balance accuracy and efficiency. In this context, the present application fully leverages the advantages of UAV platforms in high-resolution image acquisition and flexible deployment to achieve both rapid localization of large-scale cracks and high-precision analysis of micro-scale cracks. SUMMARY

[0006] To solve the above technical problems, the present application provides a multi-level concrete crack detection method based on a UAV to solve the problems in the prior art. The technical solution adopted by the present application is as follows:

[0007] A multi-level concrete crack detection method based on a UAV, comprising the following steps:

[0008] Step 1, the surface image of the building is obtained by the UAV and is transmitted to the edge computing device in real time, the MLCA-EnhancedYOLO11 model deployed by the edge computing device is used for offline processing, potential crack areas are identified in the surface image, and corresponding regions of interest are generated; the regions of interest are re-collected and observed at close range by operating the UAV;

[0009] Step 2, the regions of interest are detected using a multi-scale detection strategy to generate a fused crack segmentation mask, and micro crack detection of the potential crack areas is realized;

[0010] Step 3, the crack segmentation mask is optimized using a morphological optimization operator and connected component analysis to obtain the final crack segmentation mask.

[0011] Further, in step 1, the MLCA-EnhancedYOLO11 model is obtained based on the YOLOv11n network and mixed with a local channel attention mechanism, i.e., embedding an MLCA module in the high-level feature fusion stage of the YOLO11n network;

[0012] The processing flow of the MLCA module includes:

[0013] Step 1.1: performing local average pooling on the input feature map to extract local spatial features, performing global average pooling, and reshaping the feature map after global average pooling, to convert the input feature map into a one-dimensional vector encapsulating global information;

[0014] Step 1.2: performing one-dimensional convolution operation on the reshaped feature map in step 1.1, and then restoring the original spatial resolution through non-average pooling to obtain global pooling features; performing one-dimensional convolution operation on the feature map after local average pooling in step 1.1, and then reshaping to obtain local pooling features;

[0015] Step 1.3: fusing the global pooling features and the local pooling features through addition operation to obtain fused features;

[0016] Step 1.4: combining the fused features with the input feature map through multiplication operation to output enhanced feature map.

[0017] Further, in the one-dimensional convolution operation, the size of the convolution kernel is determined by the following formula:

[0018] (1)

[0019] wherein, is the channel number, and is a hyperparameter with a default value of 2. denotes is odd.

[0020] Further, step 2 comprises:

[0021] Step 2.1: coarse-grained candidate region screening and spatial stitching:

[0022] Let the input image size be , adopt fixed window size and step size , non-overlappingly divide the whole image to obtain a sub-image set ; each sub-image is input to the MLCA-EnhancedYOLO11 model trained in step 1 to output a confidence mask ; for each confidence mask , calculate its crack region response ratio: ; if , it is an effective crack candidate region, denoted as ; wherein is the judgment threshold of the crack region response ratio;

[0023] Then a stitching algorithm based on geometric adjacency and spatial consistency constraints is adopted: by calculating the geometric center and adjacency relationship of the candidate window, the crack fragments within the effective crack candidate region divided are recombined into continuous structural units;

[0024] Step 2.2: high-overlap rate sliding window scanning: dense sampling is performed in the candidate crack region with a smaller step size, so that the same boundary pixels in the image are repeatedly covered by multiple windows;

[0025] Step 2.3: multi-window semantic fusion: a semantic fusion mechanism is introduced, for each pixel, weighted fusion is performed in all the masks covering the pixel.

[0026] Further, in step 2.1, the stitching algorithm based on geometric adjacency and spatial consistency constraints comprises:

[0027] First, calculate the geometric center of each candidate window , judge whether any two candidate windows satisfy the following adjacency conditions:

[0028] Horizontal adjacency: (2)

[0029] Vertical adjacency: (3)

[0030] The candidate window pairs satisfying the adjacency conditions constitute a connection graph, and through recursive clustering, a set of structurally connected regions is formed; wherein, a continuous region representing the complete crack structure; and the center coordinates of the i th candidate window; and ; is the length of the fixed window, is the width of the fixed window; is the horizontal adjacency threshold, is the vertical adjacency threshold.

[0031] Further, step 2.2 includes: for each candidate , a local scan is performed using a sliding window with a window size of , and a step size of , is the step size of the sliding window; each sliding window image segment is input into the MLCA-Enhanced YOLO11 model trained in step 1 to obtain a preliminary crack segmentation mask .

[0032] Further, in step 2.3, for each pixel, all masks covering the pixel are fused by weighting, denoted as:

[0033] (4)

[0034] wherein represents the number of sliding windows in which the pixel is contained, is the crack response value of the th sliding window at the position; is the fused crack segmentation mask.

[0035] Further, in step 3, the morphological optimization operator includes: first performing an opening operation to remove false responses and isolated noise points by a method of first erosion and then dilation; then performing a closing operation by a method of first dilation and then erosion for filling small gaps and breaks in the crack boundary; and simultaneously combining an edge smoothing method to refine and correct the crack boundary to obtain the final crack segmentation mask.

[0036] The connected component analysis includes: based on 8-neighbor pixel connectivity, all pixels in the binary mask are labeled for connectivity.

[0037] The present application has the following beneficial effects: the present application fully combines the advantages of unmanned aerial vehicle inspection in high-resolution image acquisition and flexible deployment, realizes the dual goals of rapid positioning of large-scale cracks and high-precision analysis of micro-scale cracks, and overall considers the real-time requirements of the inspection task and the accuracy of pixel-level semantic modeling, providing theoretical support and method foundation for intelligent identification of large-scale structural surface cracks in high-resolution scenes; and solves the problems of the existing detection methods in the balance between real-time and accuracy, micro-crack identification, boundary detail restoration and high-resolution image processing. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The overall flow framework of the present application is shown in the figure;

[0039] Figure 2 The MLCA-EnhancedYOLO11 framework is shown in the figure;

[0040] Figure 3 The basic module structure diagram is shown in the figure;

[0041] Figure 4 The unmanned aerial vehicle detection scene diagram is shown in the figure;

[0042] Figure 5 The multi-scale detection strategy diagram is shown in the figure;

[0043] Figure 6 The 8-neighbor pixel connectivity diagram is shown in the figure;

[0044] Figure 7 Test result comparison 1 is shown in the figure;

[0045] Figure 8 Test result comparison 2 is shown in the figure. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings of the embodiments of the present application. Figures 1-8 The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings of the embodiments of the present application.

[0047] To solve the problem that crack detection in unmanned aerial vehicle inspection cannot balance between large-scale real-time and local refinement, the present application constructs a multi-level concrete crack detection method based on unmanned aerial vehicles. The overall flow is shown in the figure. Figure 1As shown, the method core is divided into three stages: real-time acquisition and positioning, multi-scale fine modeling, and structure optimization and enhancement. The three are connected to form a complete detection link from coarse to fine and from fast to stable. Overall, in the acquisition stage, the application adopts an edge computing platform based on a UAV and a target detection model to perform real-time target detection on the surface of the building structure, achieving rapid identification of suspected crack areas and task-driven image acquisition; in the fine detection stage, a multi-scale modeling strategy is adopted, which integrates a multi-scale perception path of coarse-grained region screening, fine-grained boundary enhancement, and post-processing optimization segmentation. A mixed attention mechanism MLCA is introduced to enhance the model's focusing ability on detailed features and semantic robustness in complex backgrounds. The overall consideration of real-time requirements of the inspection task and the accuracy of pixel-level semantic modeling provides theoretical support and method foundation for intelligent identification of large-scale structure surface cracks in high-resolution scenarios.

[0048] As Figure 1 A multi-level concrete crack detection method based on a UAV, comprising the following steps:

[0049] Step 1: Obtain the surface image of the building by the UAV and real-time to the edge computing device, and perform offline processing by the MLCA-EnhancedYOLO11 model deployed by the edge computing device to identify potential crack areas in the surface image and generate the corresponding region of interest; and by operating the UAV, the region of interest is re-acquired and observed at close range;

[0050] Step 2: Detect the region of interest using a multi-scale detection strategy to generate a fused crack segmentation mask, achieving micro crack detection of potential crack areas;

[0051] Step 3: Optimize the crack segmentation mask using morphological optimization operators and connected component analysis to obtain the final crack segmentation mask.

[0052] The specific technical content of each step is described in detail below.

[0053] For step 1, in the UAV inspection scenario, the crack detection task faces two major challenges: first, the UAV platform is limited by flight endurance and load capacity, so real-time detection requires the model to have high-efficiency inference performance; second, the background of the building structure surface is complex, and the crack morphology and scale differ greatly, which puts higher requirements on the detection accuracy of the model. To address the above problems, the application constructs an MLCA-EnhancedYOLO11 edge computing module that integrates the YOLOv11n architecture and the mixed channel attention mechanism (MLCA, MixedLocalChannelAttention).

[0054] YOLO11 network as the latest version of YOLO series, has significant improvements in feature extraction and inference efficiency compared to previous models, which is the prior art; as shown in Figure 2 The architecture mainly consists of three core parts: Backbone, Neck, and Head. Backbone adopts C3k2 modules, effectively reducing the computational burden through small convolution kernels and feature map segmentation strategies, while introducing C2PSA modules to enhance spatial attention expression, enabling the model to adaptively focus on high salient regions. The Neck module fuses multi-scale features from different depth levels to ensure the scale robustness of detection. The Head module uses Depthwise Separable Convolution to further reduce parameter quantity and improve computational efficiency. Compared to YOLOv8, YOLO11 optimizes feature flow design and structural compactness, making it more suitable for edge computing platform deployment, while maintaining high precision and achieving significant real-time performance advantages.

[0055] Although the YOLO11 architecture performs well in terms of efficiency and generalization, in the unmanned aerial vehicle aerial scene, cracks often exhibit elongated, low-contrast, and irregular morphological characteristics, which are easily disturbed by factors such as concrete surface texture, lighting changes, and stains, leading to insufficient stability of the model in small target recognition. Therefore, the MLCA module is embedded in the high-level feature fusion stage of the YOLO11 network to enhance the model's focusing ability on key area features.

[0056] As shown in Figure 3 The basic principle of the MLCA module. Initially, the input feature map is subjected to Local Average Pooling (LAP) to extract detailed local spatial features. Subsequently, Global Average Pooling (GAP) is applied, followed by reshaping, which converts the feature map into a one-dimensional vector encapsulating global information. The generated structure consists of two branches: the first branch contains global context information, while the second branch preserves local spatial information. For features obtained through GAP, one-dimensional convolution (Conv1d) is performed, and non-average pooling (UNAP) is used to restore the original spatial resolution. Similarly, LAP-generated features are subjected to one-dimensional convolution, followed by reshaping, in preparation for subsequent processing. To achieve information fusion, global and local pooled features are combined through additive operations, resulting in a hybrid attention effect that integrates global context into the feature map. Finally, the local and global processed features are resized to their original spatial dimensions. These refined features are then combined with the original input feature map through multiplicative operations, effectively enhancing the representational capacity of the MLCA module. This process reliably identifies crack damage even under challenging conditions such as rough surfaces and lighting changes.

[0057] The processing flow of the MLCA module includes:

[0058] Step 1.1: Perform local average pooling on the input feature map to extract local spatial features, perform global average pooling, and reshape the feature map after global average pooling, converting the input feature map into a one-dimensional vector encapsulating global information.

[0059] Step 1.2: Perform one-dimensional convolution operation on the reshaped feature map in step 1.1, and then recover the original spatial resolution through non-average pooling to obtain global pooling features; perform one-dimensional convolution operation on the feature map after local average pooling in step 1.1, and then reshape to obtain local pooling features.

[0060] Step 1.3: Fuse the global pooling features and the local pooling features through addition operation to obtain fusion features.

[0061] Step 1.4: Combine the fusion features with the input feature map through multiplication operation, and output the enhanced feature map.

[0062] Figure 3 In the formula, Conv1d represents one-dimensional convolution operation, where the kernel size k is proportional to the channel dimension C. This proportional relationship means that when capturing local cross-channel interaction information, this approach only considers the relationship between each channel and its k adjacent channels. The parameter k is selected according to formula (1):

[0063] (1)

[0064] where is the channel number, is the size of the convolution kernel, and is a hyperparameter with a default value of 2 represents is an odd number.

[0065] This structure enables the model to effectively suppress non-structural noise in complex backgrounds and highlight the prominent features of the crack region. The local attention part enhances the sensitivity to subtle textures and edge information, while the global attention plays a key role in maintaining overall structural consistency, thus achieving more balanced feature expression in complex scenes. Compared with traditional SE or CBAM attention modules, MLCA has higher representation ability in capturing cross-channel interaction and spatial details, while having lower computational complexity, greatly improving the running efficiency of the model on edge devices.

[0066] Overall, the MLCA-Enhanced YOLO11 significantly enhances the model's feature focusing ability and detection robustness in high-resolution UAV images while ensuring lightweight and real-time performance. By introducing a hybrid attention mechanism at the feature fusion layer, the model can not only more accurately distinguish cracks from complex backgrounds, but also maintain stable detection performance under weak texture and low contrast conditions. This improvement provides high-quality candidate region input for the subsequent multi-scale crack fine modeling stage and lays a solid foundation for intelligent identification of complex structural surface cracks.

[0067] In the inspection task of the UAV, the present application constructs a real-time closed-loop process of "UAV image acquisition-edge detection-ROI extraction-artificial feedback adjustment". First, the UAV platform is equipped with high-resolution imaging and stable control components to collect the target structure surface from multiple angles. The collected images are transmitted in real time to the edge computing device through the image transmission link. The MLCA-Enhanced YOLO11 model deployed on the edge performs fast inference on the image stream, completing the detection and positioning of candidate crack regions. The detection results are generated into regions of interest by the ROI extraction module, and task prompt information is formed. The UAV detection scene is as shown in Figure 4 .

[0068] Subsequently, the operator views the ROI and confidence distribution output by the edge in real time through the terminal interface, and adjusts the flight path, shooting angle or focal length setting of the UAV accordingly to supplement the collection and close-range observation of the key areas. The closed-loop mechanism of "collection-detection-feedback-re-collection" enables the inspection process to maintain wide coverage while focusing on suspected crack areas for intensive imaging and attention, thereby significantly improving the detection specificity, accuracy and operation efficiency.

[0069] For step 2, within the located ROI, a hierarchical multi-scale detection strategy is further adopted: first, coarse-grained sliding window scanning and adjacency splicing are used to ensure the structural continuity of cracks across regions; then high-overlap sliding windows are introduced in the candidate region to achieve pixel-level capture of small cracks; finally, weighted fusion of multi-window results is used to effectively improve the consistency of boundary discrimination and detail analysis ability. This stage establishes a balance between global and local, taking into account detection efficiency and fine modeling of crack geometric features, and is the core innovation of the present application.

[0070] To realize high-precision semantic segmentation of crack regions in high-resolution building surface images, the application proposes a multi-scale crack fine modeling strategy in the fine detection stage. The method takes "candidate region positioning-dense reconstruction enhancement" as the core idea, integrates coarse-grained screening and fine-grained semantic completion process, effectively balances detection efficiency and accuracy, and shows excellent structure sensitivity and segmentation consistency in small cracks, broken boundaries and complex texture background. The overall process is shown in Figure 5 .

[0071] Step 2 includes the following steps:

[0072] Step 2.1: Coarse-grained candidate region screening and spatial splicing: In high-resolution building surface images, crack regions are often sparsely distributed, and large-area background regions occupy most of the pixels. If pixel-level segmentation is directly performed on the full image, not only the calculation cost is huge, but also redundant reasoning is easy to cause, which reduces the overall detection efficiency. Therefore, the application proposes a candidate region screening strategy based on coarse-grained sliding window to quickly lock potential crack regions and effectively reduce the calculation burden of subsequent fine modeling.

[0073] Specifically, let the input image size be , adopt a fixed window size and a step size , non-overlappingly divide the full image to obtain a sub-image set . Each sub-image is input into the MLCA-EnhancedYOLO11 model trained in step 1, and the confidence mask is output. For each confidence mask , calculate its crack region response ratio: ; if , it is an effective crack candidate region, denoted as . Wherein is the judgment threshold of the crack region response ratio; this stage completes the quick locking of the high-response region of the crack, significantly reducing the search space of the subsequent fine segmentation stage.

[0074] However, conventional sliding window scanning is easy to cause cross-regional cracks to be cut, damaging their geometric connectivity and affecting the subsequent segmentation accuracy. To solve this problem, the application innovatively introduces a splicing algorithm based on geometric adjacency and spatial consistency constraints: by calculating the geometric center and adjacency relationship of the candidate window, the divided crack fragments are recombined into continuous structural units. This method effectively restores the cross-regional connectivity of the crack and avoids the "fragmentation" problem in traditional candidate region screening, providing complete context information for subsequent high-overlap fine segmentation. The specific method is as follows: first, calculate the geometric center of each candidate window ; judge whether any two windows satisfy the following adjacency conditions:

[0075] Horizontal adjacency: (2)

[0076] Vertical adjacency: (3)

[0077] The window pairs satisfying the adjacency condition constitute a connected graph, and a set of structurally connected regions is formed by recursive clustering Each represents a continuous region that can be regarded as a complete crack structure; and are the center coordinates of the first i candidate window; and ; and are the length and width of the fixed window, i.e. the actual length and width; this mechanism effectively overcomes the problem of structural fragmentation caused by window segmentation, not only significantly reducing the computational burden, but also ensuring the spatial integrity of the crack structure. It provides complete contextual semantic support for the next step of fine semantic modeling.

[0078] Step 2.2: High overlap rate sliding window scanning: After the preliminary locking of the candidate region, the integrity of the crack boundary becomes a key factor affecting the quality of segmentation. However, the conventional sliding window method usually uses a larger step or a low overlap rate division, although it can improve the computational efficiency, but often leads to the crack being truncated at the window edge, resulting in the problem of boundary missing and structural discontinuity. This defect is particularly prominent in the detection of small cracks and cross-region cracks, seriously affecting the geometric consistency of the segmentation result.

[0079] To solve this problem, the invention proposes a high overlap rate sliding window scanning mechanism. A smaller step is used for dense sampling within the candidate crack region, so that the same boundary pixels in the image are repeatedly covered by multiple windows. With this redundant sampling method, crack edge information can be perceived and confirmed multiple times between different windows, effectively alleviating the boundary missing phenomenon caused by window cutting.

[0080] Specifically, for each candidate crack region , a sliding window with a window size of is used for local scanning, and the step size is set to , is the step size of the sliding window, so as to ensure that the image region is redundantly covered in multiple windows. Each sliding window image segment is input into the MLCA-EnhancedYOLO11 model trained in step 1 to obtain the preliminary pixel-level crack mask By increasing the overlap rate between windows, the system actively utilizes the redundant information across windows to compensate and repair the boundary semantics, making the model's discrimination in the crack edge area more continuous and stable.

[0081] Step 2.3: Multi-window semantic fusion: After high-overlap sliding window scanning, the same pixel point Usually, the segmentation results of multiple windows will be repeated. This redundancy not only brings information repetition, but also provides an opportunity to further improve the stability and refinement of the segmentation results. To fully utilize this redundancy feature, the invention introduces a semantic fusion mechanism: for each pixel, the weighted fusion is performed in all masks covering the location to improve the boundary continuity and the ability to capture small cracks. Specifically, it is expressed as:

[0082] (4)

[0083] Wherein represents the pixel is contained in the number of sliding windows, is the crack response value of the th sliding window at this location. The fused mask expresses the crack determination probability under local multi-view, which can significantly improve the classification stability and consistency of the boundary area. The overlap rate can be calculated by formula (5) through the relationship between window size and step size.

[0084] (5)

[0085] By weighting and fusing multiple prediction results covering the same pixel, the system can integrate the redundant information across windows to form a globally consistent crack determination. This mechanism is particularly prominent in the boundary area: the multi-view information provided by the overlapping windows effectively compensates for the local bias of single-window reasoning, thereby significantly enhancing the continuity and stability of the crack boundary.

[0086] For step 3, after multi-scale fine modeling, the system can generate a preliminary crack segmentation mask within the ROI range. However, due to the inevitable light changes, material texture interference and model reasoning uncertainty in high-resolution images, the original mask often has problems such as boundary discontinuity, small area noise or isolated false positives. These defects not only affect the complete expression of crack geometric features, but also may weaken the accuracy of subsequent structure state evaluation. Therefore, this research introduces a joint post-processing mechanism of morphological optimization operator and connected component analysis after the mask generation to enhance the structural coherence and engineering applicability of the crack segmentation results.

[0087] Morphological optimization operator: In order to further improve the structural continuity and robustness of the mask, the present application introduces morphological optimization processing to the fusion result. First, the opening operation is performed to remove small area false response and isolated noise points by the method of erosion first and then dilation, so as to reduce the false detection of the model on the background texture, and then the closing operation is performed to fill the small gap and fracture of the crack boundary after dilation and then erosion, so as to improve the continuity of the line. At the same time, combined with the edge smoothing processing method, the crack boundary is refined and corrected to make it more consistent with the actual geometric shape, and the final high-quality crack mask is obtained.

[0088] Connected component analysis: in order to further eliminate isolated noise and false detection areas in the crack segmentation result, the present application introduces connected component analysis (CCA) after morphological optimization. This method is based on the 8-neighbor pixel connectivity as shown in Figure 6 , all pixels in the binary mask are labeled for connectivity. In this way, isolated small area false detection areas can be effectively identified and removed, and crack fragments with structural integrity and engineering significance are retained. This process ensures the global rationality of the mask result in spatial expression, and avoids the interference of noise areas on subsequent analysis.

[0089] In summary, step 3 optimizes and enhances the crack mask based on multi-scale fine modeling to further ensure the engineering usability of the result. The system can effectively overcome the noise interference and boundary defects in the original mask, realize high-precision and stable crack semantic expression. This design not only improves the overall quality of the detection result, but also lays a solid foundation for subsequent crack width measurement, extension prediction and structure safety evaluation,

[0090] In order to further verify the applicability and boundary expression ability of the method proposed in the present application in actual high-resolution images, the present application selects representative unmanned aerial vehicle real image samples to compare and analyze the visualization effect of crack segmentation results under different model structures. The fusion version of YOLO11 and MLCA attention mechanism of the present application is selected as the baseline model for crack detection of real building wall images. The output corresponds to the segmentation mask, and the crack morphology, boundary continuity and response area integrity are qualitatively analyzed. Figure 7 、 Figure 8 The comparative results of crack segmentation are shown.

[0091] The figure shows the detection results on a typical concrete structure surface image. It can be observed that without introducing the base model of the present application, the model output often has the following problems: first, the response of small cracks is insufficient, which is manifested as mask missing or breaking, making it difficult to capture fine and narrow crack structures; second, the edge prediction is discontinuous, especially in areas with complex background texture or severe light changes, the model often appears blurred or interrupted in detection; third, the structure continuity is poor, and the cracks across regions appear "segmented" response, affecting the accurate restoration of the overall crack morphology.

[0092] In contrast, after integrating the detection method proposed in the present application, the model output is significantly improved: the crack details are completely preserved, the boundary transition is natural, and the structure connectivity and noise suppression ability are maintained in complex scenes. The performance improvement is mainly due to the synergistic effect of the following mechanisms: 1. Interest region screening guided by coarse-grained detection: through the first-stage coarse sliding window scanning and threshold screening mechanism, the model can effectively eliminate invalid background regions, concentrate resources on suspected crack regions, and improve the overall recognition efficiency and robustness; 2. Spatial reconstruction mechanism of interest region: introduce candidate region stitching strategy based on geometric adjacency and structural consistency constraints in the divided image blocks, enhance the spatial continuous modeling ability of the model for cross-window crack structures; 3. High overlap sliding window and semantic redundancy fusion: dense sliding window sampling is performed in the candidate region, and the boundary response consistency is improved by using multiple semantic fusion in overlapping regions, effectively alleviating the problems of cracking, deviation, etc. caused by window cutting and branch reasoning.

[0093] In addition, from the perspective of boundary detail performance, the integrated model has stronger adaptability to complex texture interference, can better suppress false responses caused by non-structural cracks (such as wall color difference, surface erosion lines), and improve the semantic purity and discrimination stability of the mask output.

[0094] Therefore, the detection method proposed in the present application realizes boundary refinement, morphological continuity and response region accuracy at the structure expression level, provides effective technical support for high-fidelity crack modeling in complex backgrounds, and verifies the practical application value of the method in high-resolution structure image intelligent perception tasks.

[0095] The above embodiments only describe the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications, variations, modifications and replacements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A multi-level unmanned aerial vehicle (UAV) based concrete crack detection method, comprising: The method comprises the following steps: Step 1: obtaining the surface image of the building by the unmanned aerial vehicle and transmitting the surface image to the edge computing device in real time, performing offline processing on the surface image by the MLCA-EnhancedYOLO11 model deployed by the edge computing device, identifying the potential crack area in the surface image, and generating the corresponding region of interest; through the operation of the unmanned aerial vehicle, the region of interest is re-acquired and observed at close range; Step 2: detecting the region of interest by using a multi-scale detection strategy to generate a fused crack segmentation mask, and realizing micro crack detection of the potential crack area; Step 3: optimizing the crack segmentation mask by using a morphological optimization operator and connected component analysis to obtain the final crack segmentation mask; In step 1, the MLCA-EnhancedYOLO11 model is obtained based on the YOLOv11n network and mixed with a local channel attention mechanism, that is, embedding the MLCA module in the high-level feature fusion stage of the YOLO11n network; The processing flow of the MLCA module comprises: Step 1.1: performing local average pooling on the input feature map to extract local spatial features, performing global average pooling, and reshaping the feature map after global average pooling to convert the input feature map into a one-dimensional vector encapsulating global information; Step 1.2: performing one-dimensional convolution operation on the reshaped feature map in step 1.1, and then restoring the original spatial resolution through non-average pooling to obtain global pooling features; performing one-dimensional convolution operation on the feature map after local average pooling in step 1.1, and then reshaping to obtain local pooling features; Step 1.3: fusing the global pooling features and the local pooling features by addition operation to obtain fused features; Step 1.4: combining the fused features with the input feature map by multiplication operation to output the enhanced feature map; Step 2 comprises: Step 2.1: coarse-grained candidate region screening and spatial splicing: Let the input image size be , the fixed window size be and the step size be , the whole image is non-overlappingly divided to obtain a sub-image set ; each sub-image is input into the MLCA-EnhancedYOLO11 model trained in step 1 to output a confidence mask ; for each confidence mask , the crack region response ratio thereof is calculated: ; if , it is an effective crack candidate region, denoted as ; wherein is a crack region response ratio determination threshold. Then a splicing algorithm based on geometric adjacency and spatial consistency constraints is used: the geometric center and adjacency relationship of the candidate window are calculated, and the crack segments in the effective crack candidate region divided are recombined into continuous structural units; Step 2.2: high overlap rate sliding window scanning: dense sampling is performed in the candidate crack region, so that the same boundary pixels in the image are repeatedly covered by multiple sliding windows; Step 2.3: multi-window semantic fusion: a semantic fusion mechanism is introduced, and for each pixel point, all the masks covering the pixel point are fused by weighting.

2. The method of claim 1, wherein the method is based on a multi-level unmanned aerial vehicle (UAV) for concrete crack detection. In one-dimensional convolution operation, the size of the convolution kernel is determined by the following equation: （1） wherein, is a channel number, with is a hyperparameter defaulting to 2, denotes is an odd number.

3. The method of claim 1, wherein the method is based on a multi-level unmanned aerial vehicle (UAV) for concrete crack detection. In step 2.1, the splicing algorithm based on geometric adjacency and spatial consistency constraints comprises: First, the geometric center of each candidate window is calculated determines whether any two candidate windows satisfy the following adjacency condition: Horizontal abutment: (2) vertically abutting: (3) The candidate window pairs satisfying the adjacency condition constitute a connected graph, and a set of structurally connected regions is formed by recursive clustering ; wherein, represents a continuous region of the complete fracture structure; and is the center coordinate of the i th candidate window; and ; is the length of the fixed window, is the width of the fixed window; is the horizontal adjacency threshold value, is the vertical adjacency threshold value.

4. The method of claim 3, wherein the method further comprises: Step 2.2 includes: for each candidate , a local scan is performed using a sliding window with a window size of , and a step size of , is the step size of the sliding window; each sliding window image segment is input into the MLCA-Enhanced YOLO11 model trained in step 1 to obtain a preliminary crack segmentation mask .

5. The method of claim 4, wherein the method further comprises: In step 2.3, for each pixel point, all the masks covering the pixel point are fused by weighting, which is expressed as: (4) wherein represents a pixel the number of sliding windows is the crack response value of the th sliding window at the position; is the fused crack segmentation mask.

6. The method of claim 1, wherein the method is based on a multi-level unmanned aerial vehicle (UAV) for concrete crack detection. In step 3, the morphological optimization operator comprises: first, performing an opening operation to remove false responses and isolated noise points by erosion followed by dilation; then performing a closing operation, first dilation and then erosion, to fill the gaps and breaks in the crack boundary; at the same time, an edge smoothing method is used to refine and correct the crack boundary to obtain the final crack segmentation mask; Connected component analysis includes: based on 8-neighborhood pixel connectivity, performing connectivity labeling on all pixels in the binarized mask.

Images