Bridge crack detection method and device based on YOLO enhanced feature fusion and SIoU optimization

Abstract

The application relates to the field of artificial intelligence, in particular to a bridge crack detection method and device based on YOLO enhanced feature fusion and SIoU optimization, which comprises the following steps: collecting a bridge appearance image; extracting first multi-scale crack features through a cross-stage partial large kernel separable kernel attention module and a plurality of double convolution cross-stage local bottleneck modules; performing an aggregation-distribution mechanism through a neck network, performing feature alignment processing and feature fusion processing through a low-stage collection and distribution branch, and performing feature alignment processing and feature fusion processing through a high-stage collection and distribution branch; performing crack category prediction and boundary box position regression according to third multi-scale crack features, and outputting a bridge crack detection result. Significant information in the sequence is effectively captured, so that more targets to be detected can be recognized. The multi-scale feature fusion capability is enhanced, and the best delay / accuracy trade-off between different model sizes is realized.

Description

Technical Field

[0001] This application relates to a bridge crack detection method and device based on YOLO enhanced feature fusion and SIoU optimization, belonging to the field of artificial intelligence technology. Background Technology

[0002] Bridges are vital structures in transportation, and their structural health directly impacts the safety and smooth flow of the entire transportation network. With a vast number of existing bridges, safety monitoring of in-service bridges and the development of scientific assessment and maintenance methods have become a global research hotspot. Meanwhile, the construction standards of some older bridges are incompatible with the ever-increasing vehicle loads, further exacerbating bridge surface damage. Factors such as material degradation, fatigue accumulation, and harsh environmental erosion have led to a surge in various types of structural damage to bridges, including cracks, corrosion, and spalling, with cracks being a typical indicator of bridge surface damage. Crack detection and prevention are crucial for bridge maintenance. Timely detection and repair of cracks can prevent further deterioration, reducing maintenance costs while ensuring normal traffic flow and driver safety. Therefore, accurate crack detection in bridges is of great significance for improving the quality of bridge surface inspection.

[0003] Early methods for detecting bridge cracks typically relied on equipment platforms such as bridge inspection vehicles and scaffolding, using measuring instruments like crack rulers, crack microscopes, and vernier calipers, and employing manual visual measurement methods. These methods depended heavily on the knowledge and experience of inspectors. Traditional research can be categorized into five areas: wavelet transform, image thresholding, manual feature and classification, edge detection-based methods, and minimum path-based methods. While these methods are simple, convenient, and easy to implement, they generally suffer from problems such as being time-consuming and labor-intensive, relying on subjective judgment, having low detection efficiency, posing significant safety hazards, and having a high rate of missed defects. Furthermore, they are difficult to apply to complex bridge structures, thus failing to meet the core requirements of timeliness and accuracy in practical engineering applications.

[0004] Machine learning-based methods overcome the limitations of traditional manual inspection methods, enabling non-contact crack detection and identification, and are widely used in various fields. Machine learning-based methods improve the accuracy and efficiency of bridge crack detection through automated feature learning capabilities. Based on data labels and learning objectives, machine learning-based methods can be divided into supervised learning methods and unsupervised learning methods. Supervised learning uses labeled data to train models to establish a mapping relationship between input images and crack locations, offering advantages in accuracy and speed, but relying heavily on labeled data. Unsupervised learning identifies abnormal patterns through the inherent distribution characteristics of data, eliminating the need for labeling and reducing manual costs, but sacrificing quantization capabilities.

[0005] With the continuous development of computer technology, acquiring optical images of structural defects using high-definition cameras and then conducting intelligent detection of image targets using computer vision technology has gradually become a mainstream technical solution. Deep learning, as a common algorithm in the field of computer vision, can automatically adapt to complex cracks, avoid manually designing feature extractors, and possesses outstanding image processing capabilities, making it a research hotspot. Based on the processing stages, deep learning-based bridge crack detection methods can be simply divided into two categories: single-stage detection algorithms and two-stage detection algorithms. Two-stage methods divide target detection into two stages: generating candidate regions and classification and bounding box regression. Examples include Region-Based Convolutional Neural Networks (R-CNN), Fast Region-Based Convolutional Neural Networks (Fast R-CNN), and Faster Region-Based Convolutional Neural Networks (Faster R-CNN). Single-stage methods directly generate bounding boxes and classify data from the input image without generating candidate regions, such as YOLO (You Only Look Once) and Single Shot MultiBox Detector (SSD). In comparison, two-stage algorithms have higher accuracy but higher computational complexity, while single-stage algorithms are faster but have slightly lower accuracy.

[0006] Deep learning-based single-stage object detection methods have achieved breakthroughs in accuracy, speed, real-time performance, and deployment cost through end-to-end architectural innovation. They can replace machine learning-based methods and become the primary solution for bridge crack detection. Research shows that the YOLO family of algorithms performs well in many object detection networks and can be divided into two categories. The first category improves detection accuracy by modifying the YOLO network architecture, while the second category improves detection speed through lightweight design. These studies have effectively improved detection accuracy. However, when used to detect bridge cracks, their performance drops sharply due to the elongated nature of bridge cracks, the low contrast between the crack and the background, and the presence of various interference factors. Therefore, detecting bridge cracks remains challenging. Summary of the Invention

[0007] This application addresses the shortcomings of existing technologies by providing a bridge crack detection method and approach based on YOLO enhanced feature fusion and SIoU optimization.

[0008] On the one hand, a bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization is provided, the method comprising: Images of the bridge's exterior are acquired and input into a deep learning model; the deep learning model includes a backbone network, a neck network, and a detection head. The large kernel separable kernel attention module in the backbone network performs large kernel separable convolution processing and attention mechanism processing, and performs convolution processing through multiple double convolution cross-stage local bottleneck modules in the backbone network to extract the first multi-scale crack features. The neck network executes an aggregation-distribution mechanism, performs feature alignment and feature fusion on the first multi-scale crack features through low-stage aggregation and distribution branches to obtain the second multi-scale crack features, and performs feature alignment and feature fusion on the second multi-scale crack features through high-stage aggregation and distribution branches to obtain the third multi-scale crack features. The detection head predicts the crack category and regresses the bounding box position based on the third multi-scale crack features, and outputs the bridge crack detection results.

[0009] On the other hand, a bridge crack detection device based on YOLO enhanced feature fusion and SIoU optimization includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization as described above.

[0010] The beneficial effects of this application are: 1. A C2LSKA spatial attention module is proposed. The core design of C2LSKA is to capture feature information through multi-dimensional interaction and efficient computation to improve the model's feature recognition and focusing capabilities. By decomposing the large kernel attention into a series of separable small kernel attentions, this method effectively captures salient information in the sequence, enabling it to identify more targets to be detected. It has advantages in the trade-off between accuracy and speed and is suitable for dense prediction tasks in resource-constrained scenarios.

[0011] 2. The original upsampling / downsampling module in YOLO11 is replaced with a convergence-distribution (GD) mechanism. This GD architecture integrates local and global contextual information to suppress irrelevant background interference and sharpen bridge crack edges. By uniformly collecting features across layers within the global receptive domain and redistributing them to the appropriate levels, the GD mechanism establishes a more efficient feature interaction paradigm. This enhances multi-scale feature fusion capabilities while achieving an optimal latency / accuracy tradeoff across different model sizes.

[0012] 3. The SIoU loss function is added to optimize the bounding box function loss, enhancing the model's ability to evaluate box accuracy. SIoU comprehensively evaluates bounding box alignment by integrating position, dimension, and angle information, providing a more comprehensive prediction method for bounding boxes in object detection. Through its shape loss term, SIoU effectively reduces the aspect ratio difference between the predicted and ground truth (GT) detection boxes, thereby improving the accuracy of shape evaluation. When significant aspect ratio changes occur, this mechanism applies a stronger penalty, forcing the model to prioritize shape correspondence. Furthermore, SIoU considers angle factors by the direction of the line connecting the centers of the predicted and ground truth bounding boxes, thus accelerating convergence and improving matching accuracy. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization in the embodiments of this application. Figure 2 This is a diagram of the YOLO11 network structure under one scenario in an embodiment of this application. Figure 3 This is a diagram of the CLGDS network structure used in a bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization, as described in one embodiment of this application. Figure 4 This is a structural diagram of the C2LSKA module under one scenario in an embodiment of this application; Figure 5 This is a schematic diagram of the GD mechanism under one scenario in an embodiment of this application; Figure 6 This is a schematic diagram of CIoU loss and SIoU loss under one scenario in the embodiments of this application; Figure 7 This is a schematic diagram of a dataset sample and labels in one scenario of an embodiment of this application; Figure 8 This is a schematic diagram of evaluation indicators under one scenario in the embodiments of this application; Figure 9 This is a loss function curve diagram under one scenario in the embodiments of this application; Figure 10This is a precision-recall curve diagram under one scenario in the embodiments of this application; Figure 11 This is a diagram illustrating other performance diagnostics under one scenario in this application embodiment; Figure 12 This is a schematic diagram of a confusion matrix under one scenario in an embodiment of this application; Figure 13 This is a schematic diagram of partial crack detection results under one scenario in an embodiment of this application; Figure 14 This is a comparison graph of the average accuracy curves of each algorithm in an ablation experiment under one scenario in this application embodiment; Figure 15 This is a schematic diagram of the crack detection results of the ablation experiment under one scenario in an embodiment of this application; Figure 16 This is a graph of the YOLO series algorithms under one scenario in an embodiment of this application. Figure 17 This is a graph showing the relationship between processing time and mAP@50-95 for a YOLO variant and an ablation model under one scenario in this application embodiment; Figure 18 This is a schematic diagram comparing the crack detection results of different types of models under one scenario in an embodiment of this application; Figure 19 This is a schematic diagram of partial results of bridge crack detection under one scenario in an embodiment of this application; Figure 20 This is a schematic diagram comparing the performance indicators of bridge crack detection under one scenario in the embodiments of this application; Figure 21 This is a schematic diagram of a bridge crack detection device based on YOLO enhanced feature fusion and SIoU optimization in an embodiment of this application. Detailed Implementation

[0014] The specific embodiments of this application are described in detail below. This application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed herein.

[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used is for describing particular embodiments only and is not intended to limit this application.

[0016] Bridge crack detection is a crucial component of bridge maintenance and a typical object detection task. SSD and YOLO, as typical single-stage object detection methods, both rely on convolutional feature maps to predict bounding boxes, but they differ in their data processing and usage. The core innovation of the YOLO-based object detection framework lies in reconstructing the detection task into a spatially discretized grid prediction problem, achieving optimized detection accuracy while maintaining real-time performance. It has now deeply penetrated diverse applications in computer vision. Since its inception, the YOLO series of algorithms has undergone three development phases: the foundational phase (v1-v3), the performance optimization phase (v4-v7), and the architectural innovation phase (v8 to present). To achieve a balance between speed and accuracy, YOLO's basic framework has been continuously explored and improved, primarily focusing on the following dimensions: backbone network topology reconstruction, cross-level feature interaction mechanisms, and loss function optimization paths.

[0017] The backbone network, as the core of feature extraction in object detection models, directly impacts detection accuracy and speed. The DarkNet series utilizes customized convolutions to construct deep networks, improving both speed and accuracy in object detection. CSPNet enhances the receptive field by introducing Cross-Stage Partial (CSP) structures and Spatial Pyramid Pooling (SPP) modules into DarkNet-53, achieving feature map segmentation and gradient path fusion, reducing repetitive gradient information, and improving learning efficiency. Regarding lightweight design, YOLO-Mobile and YOLO-Nano employ depthwise separable convolutions and pointwise grouped convolutions, respectively, replacing standard convolution strategies, significantly reducing the number of floating point operations (FLOPs). Embedding attention mechanisms and Transformers into the backbone network not only improves recall for small object detection but also enhances detection performance in occluded scenes.

[0018] The feature interaction mechanism aims to address challenges such as scale variation, occlusion, and small object detection. Its core lies in optimizing the fusion and enhancement of information from different levels, spatial locations, and semantics in feature maps. Regarding multi-scale feature fusion, the strategy has evolved from a single Feature Pyramid Network (FPN) to bidirectional PANet, and then to weighted bidirectional BiFPN, gradually improving the interaction capabilities of cross-scale features. Simultaneously, various attention mechanisms are introduced to achieve synergistic enhancement of spatial and channel features. To avoid feature learning conflicts between classification and regression tasks, a task decoupling strategy is adopted, enabling each task to learn interactive features independently. Furthermore, GOLD-YOLO employs an information aggregation-distribution mechanism, efficiently achieving the fusion of cross-level features.

[0019] The loss function, as the optimization objective of the object detection algorithm, directly affects the model's localization accuracy and classification reliability. This object detection task requires simultaneous optimization of bounding box localization (a regression problem) and object classification (a classification problem). YOLO unifies these into a regression framework, and its loss function typically includes three parts: bounding box localization loss, object confidence loss, and object classification loss. Bounding box localization loss has evolved from the early simple squared error (MSE) loss to geometrically perceptible Intersection over Union (IoU) loss and its variants, such as Generalized Intersection over Union (GIoU), Distance Intersection over Union (DIoU), and Complete Intersection over Union (CIoU). Regarding classification loss, early YOLO used Cross-Entropy (CE) loss to handle classification and confidence, but it struggled to effectively learn difficult samples (e.g., small objects, occluded objects). Focal loss (FL) improves the detection capability for imbalanced samples by mitigating the sample imbalance problem. However, the weight allocation strategy of focal loss is not flexible enough when dealing with positive and negative samples, which led to the development of variable focal loss (VFL). In summary, the design of YOLO's loss function has evolved from the initial simple weighted sum to a complex mechanism that integrates multiple strategies to dynamically adapt to different sample characteristics.

[0020] Based on this, by improving the design of the backbone network, neck network, and loss function, the YOLO series models have demonstrated generalization performance in general object detection tasks. However, traditional methods still have limitations in addressing the challenges of crack detection in bridges, such as the weak and ambiguous crack morphology, the high diversity of geometric structures, and the complexity of background interference.

[0021] Based on this, such as Figure 1 As shown, a bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization is provided. The method includes: S101: Acquire images of the bridge's exterior and input them into a deep learning model; the deep learning model includes a backbone network, a neck network, and a detection head.

[0022] The deep learning model's network architecture is based on the YOLO11 network architecture. Specifically, with the continuous growth in demand for real-time object detection, the YOLO series of algorithms has evolved as a benchmark in this field. Compared to the previous YOLOv8, YOLO11 continues the core concept of YOLO single-shot detection, further improving performance and flexibility through optimized network structure and training strategies. YOLO11 includes five variants with different network sizes: 11n, 11s, 11m, 11l, and 11x, with the number of parameters increasing sequentially from smallest to largest. Their differences lie in the number of parameters, computational complexity, and inference speed. YOLO11 adopts the three-level structure of YOLO: Backbone, Neck, and Head, ensuring consistency in basic operations during object detection and supporting visual tasks such as object detection, instance segmentation, and rotated bounding box detection. Therefore, the YOLO11 model can be selected for bridge crack detection.

[0023] YOLO11 builds upon YOLOv8 with deeper innovations. It introduces more efficient feature fusion strategies, such as the Path Aggregation Network (PAFPN) and the Neural Architecture Search Feature Pyramid Network (NAS-FPN). These strategies can more effectively fuse feature information at different scales, improving the model's ability to capture object details. Simultaneously, YOLO11 employs more refined bounding box prediction methods, such as Simple Online and Offline Tracking with Associated Detection (SimOTA) and Task-aligned One-stage Object Detection (TOOD). These methods can more accurately predict object bounding boxes, improving detection accuracy. Furthermore, YOLO11 further enhances the model's generalization ability and robustness by introducing self-supervised learning and knowledge distillation techniques.

[0024] like Figure 2 As shown, YOLO11 adopts a typical three-stage detection architecture, consisting of a backbone network, a feature fusion neck, and a detection head. Each module works together to achieve end-to-end optimization of classification from feature extraction to target localization.

[0025] The backbone network consists of multiple components, including CBS, C3K2, SPPF, and C2PSA. It is primarily responsible for capturing feature information at different scales from the input image. Through operations such as convolution, the backbone transforms image information into efficient feature representations, providing input for the neck network. Compared to previous versions, YOLO11 eliminates the C3 module in the backbone network and adopts a CSP-based C2F module, thereby improving computational efficiency and more effective information extraction capabilities.

[0026] YOLO11's backbone network employs a deep, reconfigurable feature extraction architecture, achieving synergistic optimization of multi-scale feature representation capabilities and computational efficiency through two core innovations. The first is the replacement strategy of the Cross-Stage Partial Bottleneck (C3K2) module: replacing the traditional Cross-Stage Partial Bottleneck (C2f) structure, enhancing feature extraction flexibility through a dynamic branching mechanism, and expanding the effective receptive field while maintaining high parameter efficiency. The second is position-sensitive attention enhancement: embedding a position-sensitive attention mechanism (PSA) into the Spatial Pyramid Pooling-Fast (SPPF) layer, fusing multi-head attention with a Feed-Forward Network (FFN) module, explicitly modeling long-range dependencies of spatial features.

[0027] The Neck Network, located between the Backbone Network and the Detection Head, functions primarily to perform multi-scale fusion and semantic enhancement of the primary features extracted by the Backbone Network, thereby providing highly discriminative features for the Detection Head. YOLO11 employs a Feature Pyramid Network (FPN) as its neck architecture and embeds a C3K2 module. Through a cross-level feature aggregation mechanism, it optimizes the representation capabilities of targets at different scales and reconstructs the feature propagation path, improving the robustness of small target detection.

[0028] The detection head is the final part of the object detection network, directly responsible for generating the final detection results, including the object's category and location. YOLO11's detection head employs a decoupled task-specific head (TSDH) architecture, processing object classification and bounding box regression (BBR) tasks in parallel through independent branches. This alleviates the optimization conflict between classification and regression tasks and introduces a dynamic label assignment (DLA) strategy and a position-aware loss function (PAL) to adaptively optimize the matching accuracy of training samples for targets at different scales. The detection head deeply integrates multi-scale, high-dimensional features transmitted from the backbone and neck networks, achieving a synergistic improvement in classification confidence and localization accuracy under lightweight computational constraints, significantly enhancing generalization performance in complex scenes.

[0029] Based on this, such as Figure 3 As shown, the deep learning model for bridge crack detection proposed in this application mainly consists of three parts: a feature pyramid obtained by improving the backbone network; feature fusion achieved through the object detection domain (Gather-and-Distribute Mechanism, GD) mechanism; and a detection head that uses SCYLLA-IoU (SIoU) instead of the full crossover union ratio (CIoU).

[0030] S102: Perform large kernel separable convolution processing and attention mechanism processing through the cross-stage large kernel separable kernel attention module in the backbone network, and perform convolution processing through multiple double convolution cross-stage local bottleneck modules in the backbone network to extract the first multi-scale crack features.

[0031] Compared to traditional attention mechanisms, the CrossStage Partial Pyramid Squeeze Attention (C2PSA) module in the YOLO11 backbone enhances its ability to focus on complex occluded objects and key regions through multi-scale convolution and channel weighting, demonstrating outstanding performance in directional object detection tasks. Bridge surface cracks exhibit complex and varied shapes, lengths, widths, and orientations. When cracks are thin and have low contrast with the background, C2PSA is prone to losing crucial information during image processing, leading to decreased detection accuracy. Furthermore, the model's adaptability to changes in the shape and location of complex cracks is insufficient, making it difficult to accurately locate their centers and boundaries, further impacting accuracy. In addition, noise interference from uneven lighting, shadows, and surface textures during image acquisition also reduces the model's detection accuracy and stability.

[0032] To solve the above problems, such as Figure 4 As shown, a Cross-Stage Partial 2 with Large Separable Kernel Attention (C2LSKA) module is designed to improve the YOLO11 backbone. This module can also be described as a Cross-Stage Partial 2 with Large Separable Kernel Attention module, C2LSKA unit, etc. The PSA module in the C2PSA module is replaced by N cascaded Large Separable Kernel Attention (LSKA) modules. This decomposes the traditional 2D convolution into cascaded horizontal and vertical 1D convolution kernels, reducing computational complexity while effectively maintaining a large receptive field and enhancing the ability to model long-range spatial dependencies. Furthermore, by effectively suppressing noise interference, the robustness of the model under uneven lighting or complex texture backgrounds is improved.

[0033] Specifically, such as Figure 3 As shown, multiple C3K2 modules output their respective hierarchical features, which are then fused with the hierarchical features output by the C2LSKA module to obtain the first multi-scale crack features. For example, Figure 4 As shown, the C2LSKA module in the backbone network includes an initial convolutional layer, multi-level cross-stage partial large kernel separable kernel attention modules, spatial pyramid pooling fast layers, batch normalization layers, activation functions, and residual connection structures.

[0034] At this point, the bridge's appearance image is convolved through the initial convolutional layer to obtain a low-order feature map. Let the feature map output by the upper network be F. F first passes through the initial convolutional layer (Conv) to extract local features, resulting in the low-order feature map.

[0035] Through a multi-level, cross-stage, partially large-kernel separable attention module, feature segmentation is performed on the low-order feature map to obtain a first sub-feature map and a second sub-feature map; the first sub-feature map is then divided into two feature sub-maps through a split operation. Second sub-feature map .

[0036] Feature enhancement processing is performed on the first sub-feature map to obtain the enhanced feature map; where the first sub-feature map... The data is fed into N cascaded LSKA units for processing to obtain the output enhanced feature map. .

[0037] Feature fusion processing is performed on the enhanced feature map and the second sub-feature map to obtain the enhanced multi-scale feature map. With the second feature subgraph The features are stitched together to form a new enhanced multi-scale feature map. .at last, The output is then processed through another convolutional layer.

[0038] The enhanced multi-scale feature map is pooled by a fast layer of spatial pyramid pooling, and the corresponding hierarchical features in the first multi-scale crack feature are output through the residual connection structure.

[0039] Furthermore, the multi-level, cross-stage large kernel separable attention module includes multiple cascaded large kernel separable attention (LSKA) units. Each large kernel separable attention unit includes a large kernel depthwise separable convolution decomposition layer, a depthwise separable dilated convolutional layer, a channel attention generation layer, and a feature recalibration layer. The large kernel depthwise separable convolution decomposition layer and the depthwise separable dilated convolutional layer include two parallel one-dimensional depthwise separable convolutional layers, namely a horizontal one-dimensional convolutional kernel and a vertical one-dimensional convolutional kernel, respectively. The channel attention generation layer includes a single two-dimensional convolutional layer. The feature recalibration layer performs element-wise Hadamard product operations.

[0040] At this point, the first sub-feature map It requires processing through N cascaded large-kernel separable attention (LSKA) units. The processing flow of the first LSKA unit is as follows: Sensitive features in the horizontal and vertical directions are extracted by using two parallel one-dimensional depthwise separable convolutional layers in the large kernel depthwise separable convolutional decomposition layer. First, it is fed with a horizontal one-dimensional depthwise separable convolution kernel of size 1×(2d-1). Perform convolutions simultaneously with a vertical one-dimensional depthwise separable convolution kernel of size (2d-1)×1. Perform convolution to obtain the output features of a one-dimensional depthwise separable convolution. , as a sensitive feature.

[0041] Dilated convolution processing is performed using depthwise separable dilated convolutional layers. First, with a size of One-dimensional horizontal depth-separable dilatational convolution kernel and size are One-dimensional vertical depth separable dilated convolution kernel Perform convolution operations separately to obtain one-dimensional depthwise separable dilated convolution features. .

[0042] An attention weight map is generated through a channel attention generation layer, and then applied to the second sub-feature map through a feature recalibration layer to obtain an enhanced feature map. Then... Perform 1×1 kernel convolution to integrate information between channels and obtain the attention weight map. Finally, and Perform Hadamard operation to obtain the enhanced feature map output by the current LSKA unit. .

[0043] At this point, the enhanced feature map output by each large kernel separable attention unit is used as the input to the next concatenated large kernel separable attention unit, until the final enhanced feature map is output by the last large kernel separable attention unit. Here, the enhanced feature map output by the first LSKA unit is used... As the input to the second LSKA unit, after N iterations of LSKA unit processing, the final enhanced feature map is output. .

[0044] S103: The aggregation-distribution mechanism is executed through the neck network. The first multi-scale crack features are processed by feature alignment and feature fusion through the low-stage aggregation branch to obtain the second multi-scale crack features. The second multi-scale crack features are then processed by feature alignment and feature fusion through the high-stage aggregation branch to obtain the third multi-scale crack features.

[0045] YOLO11 fuses information from different levels through an improved cross-stage bidirectional feature fusion strategy (borrowing from the PANet structure). However, this strategy has inherent flaws when fusing across layers: information fusion from non-adjacent layers needs to be "recursively" passed through adjacent layers, inevitably causing information loss.

[0046] This lengthy transmission mechanism leads to two key problems: 1. Small crack features gradually lose detail as they are transmitted to deeper layers; 2. Large crack features may lack sufficient semantic information when transmitted to shallower layers. These information losses collectively reduce the model's detection accuracy for cracks on bridge surfaces of different scales, limiting the effectiveness of the fusion strategy. Furthermore, the neck network's failure to fully fuse shallow detail information further hinders the model's ability to accurately detect small cracks within the scope of small targets, thus increasing the false negative rate.

[0047] To address the information loss issue in recursive information fusion, a Gather-and-Distribute (GD) mechanism is proposed and deployed in the neck network. The GD mechanism utilizes convolution and self-attention to achieve global fusion of cross-layer features and efficiently distributes the fused global contextual information to all feature levels. This design reduces feature information loss and improves the efficiency of the neck network in fusing multi-scale information (especially shallow details).

[0048] The GD mechanism achieves efficient information fusion through aggregation and distribution processes, such as... Figure 5As shown, its structure includes a low-level feature map distribution (Low-GD), a high-level feature map distribution (High-GD), and an information injection module. The Low-level feature map distribution (Low-GD) processes and fuses larger feature maps, focusing on improving the detection of large targets. The High-level feature map distribution (High-GD) processes and fuses smaller feature maps, focusing on enhancing the recognition of small targets. The information injection module efficiently distributes (injects) the aggregated global information to each feature level, improving the target feature representation capability. Each feature map distribution (Low-GD and High-GD) further integrates a Feature Alignment Module (FAM) and an Information Fusion Module (IFM). The Feature Alignment Module aligns the spatial or scale of features, while the Information Fusion Module effectively fuses the aligned features.

[0049] Specifically, such as Figure 4 and Figure 5 As shown, the Low-GD branch includes the Low-FAM module, the Low-IFM module, and the first injection module.

[0050] The local features corresponding to the four levels contained in the first multi-scale crack features are determined, which are referred to here as the feature pyramid (B2, B3, B4, B5) extracted from the backbone network.

[0051] Based on the Low-FAM module, the local features (B2, B3, B4, B5) corresponding to each level of the first multi-scale crack feature are aligned through interpolation and pooling.

[0052] Based on the Low-Information Fusion Module (Low-IFM), the first global feature is extracted through convolution processing and multiple reparameterization block (RepBlock) processing, and the first global feature is divided into two first sub-parts, namely LG-B3 and LG-B4.

[0053] Through the first injection module (the Inject module in Low-GD), the two first sub-parts are injected into the local features corresponding to the middle two levels of the four layers respectively, to obtain the first global fusion features of two cross-layer features. These are then combined with the local features corresponding to the highest level of the four layers to obtain the second multi-scale crack features of three layers. That is, LG-B3 and LG-B4 are distributed and injected into B3 and B4 respectively to obtain the global fusion features P3 and P4 of the cross-layer features. In this stage, feature B5 is directly output as P5, forming a feature pyramid (P3, P4, P5), which serves as the second multi-scale crack feature.

[0054] The High-GD branch includes the High-FAM module, the High-IFM module, and the second injection module.

[0055] The local features corresponding to the three levels contained in the second multi-scale crack features are identified, which are referred to here as the feature pyramid (P3, P4, P5) of the second multi-scale crack features.

[0056] Based on the High-FAM module, the local features (P3, P4, P5) corresponding to each level of the second multi-scale crack feature are aligned through interpolation and pooling.

[0057] Based on the High-IFM module, the second global feature is extracted through a multi-head attention mechanism and divided into two second sub-parts, namely HG-P4 and HG-P5.

[0058] The second injection module injects the two second sub-parts into the local features corresponding to the two highest levels of the three layers, respectively, to obtain the second global fusion features of two cross-layer features. These are then combined with the local features corresponding to the lowest level of the three layers to obtain the third multi-scale crack features of the three layers. HG-P4 and HG-P5 are distributed and injected into P4 and P5, respectively, to obtain the global multi-scale fusion features N4 and N5. In this stage, feature P3 is directly output as N3, forming the final feature pyramid (N3, N4, N5), which serves as the third multi-scale crack feature.

[0059] S104: Using the detection head, crack category prediction and bounding box position regression are performed based on the third multi-scale crack features, and the bridge crack detection results are output.

[0060] The loss functions used in the training phase of deep learning models include SIoU loss, binary cross-entropy loss, and distributed focus loss; binary cross-entropy loss is used to train crack category prediction; SIoU loss and distributed focus loss are used to train bounding box location regression.

[0061] In object detection, the loss function for bounding box regression directly impacts detection performance. YOLOv11's loss function consists of two parts: a classification loss and a regression loss function. The classification loss function uses Binary Cross-Entropy (BCE) loss, while the regression loss function uses Distribution Focal Loss (DFL) and Complete Intersection over Union (CIoU) loss. In bounding box regression tasks, CIoU loss considers the overlap area, center point distance, and aspect ratio of the bounding boxes. However, in bridge crack detection tasks, the size and shape of foreign objects vary significantly, and CIoU fails to effectively handle these differences and the balance between easy and difficult samples. In contrast, SIoU (SCYLLA-IoU, SIoU) regression loss integrates losses from angle, distance, shape, and IoU, dynamically adjusting the loss weights to improve bounding box regression performance. Therefore, SIoU provides a more robust bounding box regression method for bridge crack detection that is sensitive to the shape and size of the target by incorporating angle and shape information.

[0062] The SIoU loss function improves upon traditional CIoU by introducing two key mechanisms: 1. Center distance constraint and dynamic gradient adjustment: Constraining the center distance between the predicted and ground truth bounding boxes, and calculating the distance using the aspect ratio of their minimum bounding boxes. This mechanism dynamically reduces the gradient contribution intensity of high-quality samples (e.g., clear lateral cracks); 2. Angle-aware correction: Correcting the rotational deviation of oblique crack detection boxes through an angle-aware mechanism. These two mechanisms compensate for the shortcomings of CIoU, which relies solely on aspect ratio and center distance. Through this coupling mechanism of dynamic gradient adjustment and geometric constraints, SIoU effectively improves the model's convergence speed and detection accuracy. Figure 6 The calculation process of CIoU loss function and SIoU loss function is shown in the figure.

[0063] like Figure 6 As shown, where B and These are the crack prediction box and the annotation box, respectively. and d represents the height and width of the smallest bounding box containing the true and predicted values, respectively; d is the distance between the center points of the true and predicted bounding boxes; c is the length of the diagonal of the smallest region containing both boxes. The SIoU loss function is calculated as shown in Formula 1: Formula 1; where Δ is the distance loss function, Let be the shape loss function, IoU be the crossover ratio loss function, and Δ and . The calculation involves the angle loss function. .

[0064] In summary, the overall loss function of the algorithm is expressed as follows: ;in, , , For hyperparameters, For the complete loss function, Let BCE be the loss function. For DFL loss function, This is the SIoU loss function.

[0065] In summary, a triple-enhanced YOLO11 framework for robust bridge crack detection in complex environments is proposed. Traditional YOLO network architectures use C2PSA modules to enhance the model's ability to extract and process multi-scale features, but their ability to capture subtle crack features is limited. LSKA units are highly adaptable to different convolutional kernel sizes, while C2LSKA, by replacing the PSA module in the C2PSA module with an LSKA module, proposes a heterogeneous sensing module capable of sensing cracks in structures of different scales.

[0066] The C2PSA module enhances the model's ability to extract and process multi-scale features, but it exhibits limited ability to capture fine crack features. In contrast, the LSKA unit demonstrates strong adaptability to convolutional kernels of different sizes. C2LSKA integrates the C2PSA structure with LSKA, proposing a heterogeneous perception module capable of detecting cracks with different structural scales. To further sharpen bridge crack edges and suppress interference from irrelevant background regions, a feature attention fusion block aggregation-distribution (GD) is designed to fuse crack features from different network layers. Considering the complex geometric properties of cracks and the superior geometric perception capability of scale-invariant IoU with rotation alignment, SIoU is adopted as the regression objective function. In summary, the proposed C2LSKA-GD-SIoU has the advantages of capturing fine features, fusing local detail features and global semantic features, and extracting geometric perception capabilities for bridge cracks.

[0067] 1. A C2LSKA module is proposed. The core design of C2LSKA is to capture feature information through multi-dimensional interaction and efficient computation to improve the model's feature recognition and focusing capabilities. By decomposing a large kernel attention into a series of separable small kernel attentions, this method effectively captures salient information in the sequence, enabling it to identify more targets to be detected. It has advantages in the trade-off between accuracy and speed and is suitable for intensive prediction tasks in resource-constrained scenarios.

[0068] 2. The original upsampling / downsampling module in YOLO11 is replaced with a convergence-distribution (GD) mechanism. This GD architecture integrates local and global contextual information to suppress irrelevant background interference and sharpen bridge crack edges. By uniformly collecting features across layers within the global receptive domain and redistributing them to the appropriate levels, the GD mechanism establishes a more efficient feature interaction paradigm. This enhances multi-scale feature fusion capabilities while achieving an optimal latency / accuracy tradeoff across different model sizes.

[0069] 3. The SIoU loss function is introduced to effectively improve the model's localization accuracy by refining the bounding box regression process. SIoU comprehensively evaluates bounding box alignment by integrating position, dimension, and angle information, providing a more comprehensive prediction method for ground truth bounding box matching in object detection. Through its shape loss term, SIoU effectively reduces the aspect ratio difference between predicted and ground truth bounding boxes, thereby improving the accuracy of shape evaluation. When significant aspect ratio changes occur, this mechanism applies a stronger penalty, forcing the model to prioritize shape correspondence. Furthermore, SIoU considers angle factors by the direction of the line connecting the centers of predicted and ground truth bounding boxes, thereby accelerating convergence and improving matching accuracy.

[0070] In one embodiment, experiments are conducted on the method embodiments described above to verify their effectiveness.

[0071] First, the test and training datasets are described, including high-resolution images of the bridge surface. Then, the implementation details of the crack recognition network are given, and the evaluation criteria are described.

[0072] After determining the algorithm structure, a large number of suitable crack images need to be collected as the training and validation sets for the algorithm. The data provided in this paper comes from an open-source platform. The original dataset contains 1424 bridge appearance images, each with a size of 1024×1024 pixels, such as... Figure 7 As shown in section (a) of the document. Considering that some original images contain redundant objects or highly similar backgrounds, this application prefers images with different features and significant variations.

[0073] EasyDL was used to label the crack locations, and the labeled images will be used for subsequent model training and prediction. Some labeled images in the dataset are shown below. Figure 7 As shown in section (b) of the dataset. The dataset is then divided into training and test sets in a 7:3 ratio.

[0074] Before training the model, experimental parameters need to be determined and a virtual environment configured. In this embodiment, experiments were conducted on a computer running Ubuntu 22.04. Model training and validation were performed on an NVIDIA GeForce RTX 4070 Ti SUPERGPU (16GB) and an Intel Core i9-14900KF CPU. The deep learning environment was built using PyCharm, Python 3.8, and PyTorch 2.3.1 + cu121.

[0075] The specific parameters for model training are as follows: Image size is a key factor in defining the features of the input data. In the YOLO algorithm, to maintain the efficiency and real-time performance of the algorithm, a fixed value is usually required as the input. The image input size is set to 640*640. Batch size determines the number of samples used in each parameter update. The training batch size is 32, which means that 32 images are used in each iteration. An epoch refers to the number of times the entire training set is input into the neural network for training, which is 2000 times; the learning rate is the amount of update of the network weights in the optimization algorithm, which determines whether and when the loss function converges to a local minimum. Therefore, choosing an appropriate learning rate is crucial for model training. The initial learning rate for this training process is 0.01. `auto` is selected as the optimizer, and the optimizer parameters are automatically adjusted according to the performance metrics of the training process. In this experiment, the stochastic gradient descent algorithm is used to optimize the training method. This application also establishes an early stopping mechanism; if the effect does not improve after 100 epochs, training will stop.

[0076] When evaluating target detection, targets can only be correctly identified and labeled when the IoU (Intersection over Union) is greater than 0.5. After completing the model experiments, it is crucial to evaluate the model's ability to accurately identify and locate cracks in the test images.

[0077] The evaluation focuses primarily on detection accuracy and model inference speed. Before introducing specific metrics, this application first defines a confusion matrix, which provides a structured summary of classification predictions, illustrating the distribution between predicted class labels and actual true labels. The data composition calculation in the confusion matrix is ​​the foundation for all subsequent evaluation metrics. For binary classification tasks, the confusion matrix contains four basic elements: true positive (TP), false positive (FP), true negative (TN), and false negative (FN), encompassing all possible prediction scenarios. TP refers to a sample that is actually positive and correctly predicted as positive by the model, such as... Figure 8 As shown in section (b) of the diagram. TN refers to samples that are actually negative but were correctly predicted to be negative by the model, such as... Figure 8 As shown in section (d) of the model. FP represents a sample that is actually negative but the model incorrectly predicts it as positive, such as... Figure 8 As shown in section (c) of the diagram. FN represents a sample that was actually positive but was incorrectly predicted as negative by the model, such as... Figure 8 As shown in part (e).

[0078] This paper uses accuracy (P), recall (R), mAP@50, and mAP@50-95 as evaluation criteria for model trials, with mAP@50 and mAP@50-95 being the primary evaluation metrics. P is the core metric for measuring the accuracy of the detection model, defined as the proportion of true positives to all predicted positives (including true positives and false positives), as shown in Formula 3. A higher value indicates a lower false positive rate.

[0079] Formula 3.

[0080] R, also known as sensitivity or true positive rate (TPR), is a metric for measuring the integrity of a detection model. It is defined as the proportion of correctly identified positive examples out of all true positive examples, as shown in Formula 4. A higher recall rate indicates that the model misses fewer true positive samples (such as cracks), but its level is unrelated to whether the model produces false positives.

[0081] Formula 4.

[0082] Average precision (AP) is the area under the precision-recall curve (AUC) for each class, while mean average precision (mAP) is the average of AP across all classes. Generally, higher AP and mAP values ​​indicate better overall classifier performance. AP and mAP are calculated as shown in formula (V).

[0083] ; Formula 5.

[0084] F1 is a single metric that comprehensively evaluates a model's precision and recall. Its value is the harmonic mean of the two, as shown in Formula 6. This metric seeks a balance between the two types of errors (false positives and false negatives).

[0085] Formula Six; Inference time refers to the time required for a model to process a single input and generate its output (prediction). It is a key metric for evaluating model efficiency and real-time performance. Frames per second (FPS) is a closely related metric that quantifies the speed at which a model processes sequential inputs, specifically measuring the number of frames processed per second. FPS is widely used to evaluate the throughput of computer vision models. A higher FPS value indicates faster processing speed and the ability to achieve real-time applications. Its formula is shown in Formula 7: Formula 7; FrameNum and ElapsedTime represent the total number of images processed and the total inference time consumed by the model, respectively. FLOPs (floating-point operations) are a key metric for quantifying the computational complexity of a deep learning model, representing the number of floating-point operations required to process a single input sample. A lower FLOPs value indicates higher computational efficiency.

[0086] Therefore, the analysis of experimental results begins with an analysis of the training results. Then, the ablation study is introduced and compared with mainstream baseline target detection methods.

[0087] The loss function values ​​for both the training and validation sets consist of three main components: bounding box loss (box_loss), classification loss (cls_loss), and distribution focus loss (dfl_lors). Box_loss quantifies the difference between the predicted bounding box and the ground truth bounding box. Minimizing box_loss enhances the model's localization ability, producing more accurate bounding box predictions. cls_loss evaluates the deviation between the predicted object class and the actual object class. By minimizing cls_loss, the model gains enhanced discriminative ability, improving classification accuracy. dfl_loss utilizes keypoints to estimate the object's orientation and angle information. Minimizing dfl_loss improves the model's ability to capture accurate object orientation and fine-grained details, thus improving overall performance.

[0088] Figure 9 Parts (a) and (b) respectively show the loss functions of the YOLO11 algorithm and the proposed algorithm on the training and validation sets. Clearly, the loss curve decreases rapidly at the beginning, then the decreasing trend gradually slows down, and finally stabilizes. When the number of iterations reaches 207, the loss on the training set stabilizes at the following... Stability on the validation set The following is a summary of the process. Therefore, the training process of the network is considered complete. At the termination of the iteration, the final loss values ​​of the baseline YOLO11 algorithm were 0.75186 (bounding box loss), 0.80983 (classification loss), and 1.21941 (dfl loss). In contrast, the algorithm with an iteration step size of 345 achieved a bounding box loss of 0.73462, a cls loss of 0.76312, and a dfl loss of 1.21718. Comparative analysis shows that the proposed algorithm has a bounding box loss of 0.01724, a cls loss of 0.04671, and a dfl loss of 0.00223. These improvements indicate that the predictions generated by the proposed method are closer to the ground truth, demonstrating a gradual enhancement in the model's ability to identify cracks in images. Furthermore, Figure 9Parts (c), (d), (e), and (f) describe the precision, recall, mAP@50, and the mAP@50-95 curve over 100 training epochs. Both metrics show rapid initial growth, followed by steady growth, eventually converging around 182 epochs. This trend indicates that the model's performance gradually improves and its crack detection capability enhances as training progresses.

[0089] Figure 10 The mAP comparison results between the proposed algorithm and YOLO11 are presented. It can be observed that, in the trade-off between precision and recall, the curve of the proposed algorithm is closer to the upper right corner, indicating its superior performance. Thanks to the improved model design, the crack detection algorithm proposed in this paper achieves an excellent mAP of 98.4%.

[0090] Figure 10 The curves for mAP@50 of YOLO11 and the proposed algorithm are shown. Figure 10 The mAP@50 curves of YOLO11 and the proposed algorithm were compared. Both curves exhibit a typical inverse relationship between recall and precision, with the curve closer to the upper right corner indicating better overall detection performance. The proposed crack detection algorithm achieved an mAP@50 of 92.3%, representing a 2.1 percentage point performance improvement compared to the baseline model. Recall decreases as precision increases, and the curve closer to the upper right corner indicates superior detection performance. The proposed crack detection algorithm achieves an mAP@50 value of 92.3, a 2.1% improvement compared to YOLO11.

[0091] To comprehensively evaluate model performance, we calculated F1-confidence, precision-confidence (PC), and recall-confidence (RC) curves as quantitative supplements. Figure 11 These curves visually reveal the model's overall performance in crack detection and classification: the F1 curve reflects the trade-off between precision and recall; the PC curve shows that precision increases monotonically with the confidence threshold, and the F1 score reaches a peak of 0.89 at a threshold of 0.320; the broad area under the RC curve confirms that the model has high recall and low false negative rate.

[0092] Figure 12 The confusion matrix of a simplified bridge crack detection algorithm is presented, where the horizontal axis represents the true value and the vertical axis corresponds to the predicted label. A diagonally dominant distribution indicates an overall detection accuracy of 93%, quantitatively validating the model's ability to identify structural defects.

[0093] Figure 13A comparative analysis of crack detection results between the proposed algorithm and the YOLO11 algorithm is presented. As shown in the figure, the proposed algorithm has higher accuracy and precision, especially in identifying fine cracks, which is of great value for bridge structure maintenance. Figure 13 The crack detection performance of the proposed algorithm was compared with that of the original YOLO11. Compared with YOLO11, the proposed algorithm has higher accuracy and precision in identifying crack details, significantly improving the maintenance efficiency and safety of bridge structures.

[0094] In addition, ablation experiments were conducted to verify the effectiveness of the proposed network and its C2LSKA, GD, and SIoU modules in improving the performance of the YOLO11 baseline for bridge crack detection. For ease of comparison, only these components were modified, while the other network structures remained unchanged. To verify the effectiveness of the proposed C2LSKA, GD, and SIoU modules, we conducted ablation experiments on the YOLO11 baseline. In the experiments, only the aforementioned modules were added or replaced sequentially, while the remaining network structures remained consistent.

[0095] The ablation study aimed to demonstrate the individual contributions of three key enhancement modules to the proposed algorithm: 1. Integrating the designed C2LSKA module into the backbone network; 2. Implementing the GD mechanism in the neck network; and 3. Employing the SIoU loss function in the head network. Compared to the baseline detector in the accuracy metrics, the initial YOLO11 architecture achieved a 58.7% mAP@50-95 value, with each modification contributing incrementally to performance. Adding the C2LSKA module improved the algorithm's mAP@50-95 value by 1%. The inclusion of the GD mechanism resulted in a 3.7% increase in the algorithm's mAP@50-95 value. The integration of SIoU was a crucial improvement, increasing the algorithm's mAP@50-95 value by 2.9%. The synergistic integration of all three enhancements achieved a final mAP@50-95 of 64.2%, an absolute improvement of 5.5% over the baseline. Table 1 systematically details these ablation results, highlighting the cumulative optimization effect of each component.

[0096] Table 1 Ablation Experiment Results 1

[0097] Ablation experiments show that the addition of the C2LSKA module improves the mAP@50-95 value of the YOLO11 algorithm by 1%. This improvement is attributed to several factors, including C2LSKA's ability to fuse multi-scale feature maps and its ability to capture feature information across different receptive domains through max-pooling operations of varying sizes. C2LSKA's ability to capture features and fuse multi-scale feature maps is crucial. While the original C2PSA shows significant performance in multi-scale fusion, it incurs substantial computational overhead. The introduced C2LSKA module employs a unique architecture design that preserves multi-scale receptive fields to capture rich contextual information while reducing computational costs. This integration of multi-scale processing and deep context awareness enables the module to generate more comprehensive feature representations. Specifically, large-kernel convolutions in LSKA cover multi-dimensional input regions, thereby obtaining broader contextual relevance. In complex object detection scenarios, this module effectively addresses feature relationships between dimensions, improving detection accuracy. To prevent potential feature fragmentation caused by multi-dimensional convolutions, LSKA implements feature fusion operations at its tail, ensuring the continuity and integrity of information. As a plug-and-play component, C2LSKA can be customized and optimized according to task or dataset requirements. This flexibility not only enhances model generalization but also allows for parameter configuration adjustments for specific scenarios (e.g., small / occluded object detection), further improving detection accuracy by enhancing attention to detailed features.

[0098] The synergy between the GD mechanism and C2LSKA further improves YOLO11's mAP@50-95. Its core innovation lies in integrating multi-level backbone features through dual-path feature integration (low GD / high GD), achieving enhanced feature map representation. This mechanism combines convolutional operations with self-attention, simultaneously capturing local and global spatial information, thus alleviating the spatial modeling limitations inherent in single-channel attention mechanisms. To address the information loss problem during cross-layer feature interaction, the GD mechanism innovatively introduces three components: FAM, IFM, and Inject. FAM collects and aligns cross-level features, IFM promotes the mixing and exchange of multi-source features, and Inject enhances global representation capabilities. In summary, these components significantly enhance the model's ability to capture key image details and contextual information. Simultaneously, the GD mechanism performs a dual nonlinear recalibration operation, compressing channel redundancy information and focusing on task-critical features. This dual process enhances feature discrimination capabilities, providing more accurate decision-making for classification and localization tasks.

[0099] Based on the YOLO11 baseline framework, the end-to-end integration of the C2LSKA module, GD mechanism, and SIoU loss function resulted in a 5.5% absolute improvement in mAP@50-95 for bridge crack detection compared to the original model. This significant gain validates the module's effectiveness and its potential for collaborative optimization. The SIoU loss function effectively integrates the angle, distance, and shape information of the target in the localization task, overcoming the limitations of traditional IoU in scale-varying situations. This integrated approach overcomes the shortcomings of existing IoU-based methods that rely on adding new geometric constraints to accelerate convergence and improve detection performance. SIoU reformulates the loss function by incorporating a scale-invariant design principle that considers the vector angle between the ground truth and the predicted bounding box. Its scaling normalization module converts absolute dimensions to relative scales during IoU calculations, eliminating the influence of target size on the evaluation results. The dynamic weighting mechanism in SIoU automatically adjusts the penalty weights for position and shape errors according to the target scale, preventing small objects from being over-penalized due to absolute positional deviations. By introducing geometric constraints such as the center distance of the detection box and the aspect ratio into the IoU calculation, SIoU improves the accuracy of the detection box selection in dense scenes and significantly improves the detection of targets with extreme aspect ratios (such as cracks).

[0100] In terms of model performance, the baseline YOLO11 network achieves 333.3 FPS. Introducing the C2LSKA module significantly reduces the model size while maintaining acceptable performance. As shown in Tables 1 and 2, using the C2LSKA module alone increases the inference speed of the baseline model to 555.5 FPS. Although the introduction of the GD mechanism and SIoU loss reduces the inference speed by 133.3 FPS and 39.2 FPS respectively, they significantly improve the model's recall, precision, and mAP metric. Through the synergistic integration of the three modules, the final framework improves inference accuracy while maintaining an inference speed of 232.6 FPS. Experiments show that this integrated scheme outperforms the original YOLO11 model and any independent model in terms of model size and performance balance, validating the effectiveness of the overall architecture.

[0101] Table 2 Ablation Experiment Results 2

[0102] In the proposed algorithm's network structure, the C2LSKA module, GD mechanism, and SIoU loss function do not operate in isolation but rather collaboratively, jointly enhancing feature extraction and detection capabilities. C2LSKA enriches feature inputs through multi-scale context fusion while reducing redundancy. The GD mechanism enhances feature map representation by integrating hierarchical features, thereby improving semantic consistency. SIoU enhances localization and discrimination by introducing geometric factors (angle, distance, shape). The synergistic effect of these three modules constitutes the main catalyst for the performance enhancement of YOLO11.

[0103] Figure 14 Line graphs are used to illustrate the ablation study results. YOLO11_01 represents the integration of the C2LSKA module in the backbone network, YOLO11_02 represents the integration of the GD mechanism, and YOLO11_03 corresponds to the SIoU loss function replacement. To further visualize the contribution of each component of the proposed network in the final segmentation results, Figure 15 The ablation study results are presented: (a) a comparison of the backbone network with and without the C2LSKA module, (b) a performance comparison of some neck components before and after integrating the GD mechanism, and (c) a comparison of detection results of some head networks with and without the SIoU loss function. Visual evidence shows that adding these modules significantly improves crack detection capabilities.

[0104] To verify the effectiveness of the proposed bridge crack detection method, comparative experiments were conducted with mainstream object detection methods. All experiments used the same dataset and training environment to ensure the reliability and reproducibility of the results, including hardware configuration, hyperparameter settings, and data augmentation strategies. The benchmark object detection algorithms are compared as follows: SSD is a classic end-to-end single-level detection framework that achieves an effective balance between detection accuracy and speed through multi-scale feature mapping and a predefined default box mechanism. Its key innovation lies in eliminating the region proposal step, thereby significantly improving inference efficiency. Faster R-CNN proposes an end-to-end trainable two-stage framework that solves the inherent computational redundancy problem of traditional methods (such as R-CNN and Fast R-CNN) by achieving weight sharing between the Region Proposal Network (RPN) and the object detection network, and achieves substantial improvements in both detection accuracy and inference speed. YOLOv5: Designed by Ultralytics, YOLOv5 improves upon the YOLOv3 / v4 architecture through modular reconstruction, training optimization, and deployment enhancements. This industry-adopted framework significantly improves the accuracy of small target recognition while maintaining real-time detection efficiency. YOLOv8: Introduced by Ultralytics in 2023, YOLOv8 implements an anchor-free method within a unified multi-task architecture. Its core innovation combines a scalable backbone network with a task decoupling head, achieving simultaneous state-of-the-art accuracy and efficient inference for detection, segmentation, and pose estimation tasks. YOLO11 retains the efficiency of a single-stage detector while addressing three key challenges: missing small targets, long-tail distribution adaptation, and multimodal fusion. By deeply integrating the strengths of Transformer and CNN, it constructs a lightweight yet accurate dual-track architecture. First, comparative experiments were conducted between the proposed algorithm and the widely adopted fast R-CNN and SSD object detection methods. As shown in Table 3, in terms of accuracy, the proposed algorithm achieved a 92.3% mAP@50 score, outperforming the faster R-CNN (91.24%) and SSD (90.57%). Simultaneously, the recall rate was 91.4% and the precision was 87.8%. In terms of performance, the proposed algorithm achieved a GFLOPs score of 28.4, outperforming the faster R-CNN (30.43) and SSD (210.54). It also achieved 26.4 MB of parameters. It demonstrates strong overall performance in the object detection task. Table 3 shows the strong overall performance in the object detection task.

[0105] Table 3 Accuracy Comparison Table

[0106] To verify the superiority of the proposed algorithm, comparative experiments were conducted with YOLOv5, YOLOv8, and YOLO11, and the results are shown in Tables 4 and 5. Table 4 shows that the model proposed in this application outperforms the selected comparison methods in terms of precision, recall, F1 score, mAP@50, and mAP@50-95. Clearly, the method proposed in this application achieves 64.2% mAP@50-95, significantly outperforming YOLOv5 (52.2%), YOLOv8 (56.4%), and YOLO11 (58.6%), with differences of 10.2, 6.0, and 3.8 percentage points, respectively. Table 5 presents a comparison of the proposed algorithm with YOLO series in terms of model performance metrics. The model proposed in this application outperforms the selected comparison methods in terms of FPS, GFLOPs, and Params. Figure 5 As can be seen, the model in this application outperforms the selected comparison method in terms of FPS, GFLOPs, and Params. This model contains 13,609,699 parameters, exceeding the baseline YOLO11. The increase in the number of parameters is due to the introduction of the GD mechanism. To balance model size and detection accuracy, this application integrates the C2LSKA module into the backbone network for parameter reduction.

[0107] Table 4 Comparison Experiment Table 1

[0108] Table 5 Comparison Experiment Table 2

[0109] Figure 16 The line graph further compares the detection accuracy (mAP) of each algorithm, showing that the proposed algorithm maintains a leading mAP value throughout the entire training cycle. Furthermore, considering both model processing time and mAP@50-95 values, Figure 17 The scatter plot illustrates that the proposed algorithm demonstrates a competitive advantage in terms of speed-accuracy tradeoff.

[0110] In YOLO, confidence level is a value between 0 and 1, indicating how certain the model is about the detected target. This method shows good results in bridge crack detection. Figure 18 The crack bounding boxes predicted using the developed model, YOLOv5, YOLOv8, YOLO11, and the proposed method are shown. Figure 18 The first column displays the original images containing different crack types, while the second to fifth columns show the detection results from YOLOv5, YOLOv8, YOLO11, and the proposed algorithm, respectively. Subfigure (a) compares the detection performance of transverse cracks, (b) compares the detection performance of longitudinal cracks, and (c) evaluates the detection efficiency of complex crack morphologies (e.g., mesh / diagonal cracks). Experimental results show that the algorithm has strong robustness in detecting multi-scale irregular cracks with morphological variations such as branching and discontinuity.

[0111] Figure 19 The algorithm's representative detection results on the validation set are shown. Figure 19 As shown, the algorithm in this application achieves high accuracy in detecting typical cracks (lateral, longitudinal, and mesh-like) (92.3% mAP@50). The method remains robust in challenging scenarios, including elongated cracks and low-contrast cracks. The generalization ability of the multi-scale feature fusion mechanism is verified.

[0112] In summary, under strict model complexity constraints, this method significantly improves detection accuracy (5.5% absolute gain at mAP@50-95). Unlike traditional two-stage detectors (such as SSDs) which face a fundamental accuracy-efficiency trade-off, this solution effectively addresses two key challenges: identifying fine cracks and suppressing crack-like background noise. Its technical advantages have been rigorously validated in demanding engineering environments such as bridge crack detection.

[0113] This application uses a public bridge crack detection dataset to evaluate the model's generalization ability, including images of transverse, longitudinal, diagonal, and network cracks. For example... Figure 20 As shown, the proposed model significantly improves upon the baseline YOLO11 in both quantitative metrics and relative gains. After 300 training iterations, the validation set F1 score and mAP50 scores reached 89.6% and 92.3%, respectively. Compared directly to YOLO11, the model achieves an absolute improvement of 1.8% in F1 score, a 2.2% improvement in mAP50, and a 5.6% improvement in mAP@50-95. These results validate its superior performance for complex crack detection tasks.

[0114] like Figure 21As shown in the embodiments of this application, a bridge crack detection device based on YOLO enhanced feature fusion and SIoU optimization is also provided, including: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization as described in any of the above embodiments.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are exhaustively listed. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] For those skilled in the art, various modifications and improvements can be made without departing from the concept of this application, and these all fall within the scope of protection of this application. The scope of protection of this application is determined by the appended claims.

Claims

1. A bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization, characterized in that, The method includes: Images of the bridge's exterior are acquired and input into a deep learning model; the deep learning model includes a backbone network, a neck network, and a detection head. The large kernel separable kernel attention module in the backbone network performs large kernel separable convolution processing and attention mechanism processing, and performs convolution processing through multiple double convolution cross-stage local bottleneck modules in the backbone network to extract the first multi-scale crack features. The neck network executes an aggregation-distribution mechanism, performs feature alignment and feature fusion on the first multi-scale crack features through low-stage aggregation and distribution branches to obtain the second multi-scale crack features, and performs feature alignment and feature fusion on the second multi-scale crack features through high-stage aggregation and distribution branches to obtain the third multi-scale crack features. The detection head predicts the crack category and regresses the bounding box position based on the third multi-scale crack features, and outputs the bridge crack detection results.

2. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 1, characterized in that, The backbone network includes an initial convolutional layer, multi-level cross-stage partial large kernel separable kernel attention modules, spatial pyramid pooling fast layers, batch normalization layers, activation functions, and residual connection structures. The large kernel separable kernel attention module in the cross-stage portion of the backbone network performs large kernel separable convolution processing and attention mechanism processing, specifically including: The bridge appearance image is processed by the initial convolutional layer to obtain a low-order feature map. The low-order feature map is segmented using the multi-level, cross-stage, partially large kernel separable kernel attention module to obtain a first sub-feature map and a second sub-feature map. Feature enhancement is then performed on the first sub-feature map to obtain an enhanced feature map. Finally, feature fusion is performed on the enhanced feature map and the second sub-feature map to obtain an enhanced multi-scale feature map. The enhanced multi-scale feature map is pooled using the spatial pyramid pooling fast layer, and the corresponding hierarchical features in the first multi-scale crack feature are output through the residual connection structure.

3. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 2, characterized in that, The multi-level, cross-stage large-kernel separable attention module includes multiple cascaded large-kernel separable attention units. Each large-kernel separable attention unit includes a large-kernel depthwise separable convolution decomposition layer, a depthwise separable dilated convolutional layer, a channel attention generation layer, and a feature recalibration layer. The large-kernel depthwise separable convolution decomposition layer and the depthwise separable dilated convolutional layer include two parallel one-dimensional depthwise separable convolutional layers, namely a horizontal one-dimensional convolutional kernel and a vertical one-dimensional convolutional kernel. The channel attention generation layer includes a single two-dimensional convolutional layer. The feature recalibration layer performs element-wise Hadamard product operations.

4. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 3, characterized in that, Feature enhancement processing is performed on the first sub-feature map to obtain an enhanced feature map, specifically including: The two parallel one-dimensional depthwise separable convolutional layers in the large kernel depthwise separable convolutional decomposition layer are used to extract sensitive features in the horizontal and vertical directions, respectively. Dilated convolution processing is performed using the depth-separable dilated convolutional layer; An attention weight map is generated through the channel attention generation layer, and the attention weight map is applied to the second sub-feature map through the feature recalibration layer to obtain an enhanced feature map.

5. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 4, characterized in that, The method further includes: The enhanced feature map output by each large kernel separable attention unit is used as the input to the next concatenated large kernel separable attention unit, until the final enhanced feature map is output by the last large kernel separable attention unit.

6. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 1, characterized in that, The low-stage aggregation branch includes a low-stage feature alignment module, a low-stage information fusion module, and a first injection module; The neck network executes an aggregation-distribution mechanism, performing feature alignment and feature fusion processing on the first multi-scale crack features through low-stage aggregation and distribution branches to obtain the second multi-scale crack features, specifically including: Determine the local features corresponding to the four levels contained in the first multi-scale crack features; Based on the low-stage feature alignment module, the local features corresponding to each level of the first multi-scale crack feature are aligned through interpolation and pooling. Based on the low-stage information fusion module, the first global feature is extracted through convolution processing and multiple reparameterized block processing, and the first global feature is divided into two first sub-parts; The first injection module injects the two first sub-parts into the local features corresponding to the middle two levels of the four levels to obtain the first global fusion feature of the two cross-layer features, and combines it with the local features corresponding to the highest level of the four levels to obtain the second multi-scale crack feature of the three levels.

7. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 6, characterized in that, The high-stage aggregation branch includes a high-stage feature alignment module, a high-stage information fusion module, and a second injection module; The third multi-scale crack features are obtained by performing feature alignment and feature fusion processing on the second multi-scale crack features through high-stage clustering and dispersing branches, specifically including: Determine the local features corresponding to the three levels contained in the second multi-scale crack features; Based on the high-stage feature alignment module, the local features corresponding to each level of the second multi-scale crack feature are aligned through interpolation and pooling. Based on the high-stage information fusion module, a second global feature is extracted through a multi-head attention mechanism, and the second global feature is divided into two second sub-parts; The second injection module injects the two second sub-parts into the local features corresponding to the two highest levels of the three levels to obtain the second global fusion features of the two cross-layer features. These features are then combined with the local features corresponding to the lowest level of the three levels to obtain the third multi-scale crack features of the three levels.

8. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 1, characterized in that, The loss functions used by the deep learning model during the training phase include SIoU loss, binary cross-entropy loss, and distribution focus loss. The binary cross-entropy loss is used to train crack category prediction; the SIoU loss and the distribution focus loss are used to train bounding box position regression.

9. The bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization according to claim 8, characterized in that, The formula for the SIoU loss is: Where Δ is the distance loss function, Let be the shape loss function, IoU be the crossover ratio loss function, and Δ and . The calculation involves the angle loss function. .

10. A bridge crack detection device based on YOLO enhanced feature fusion and SIoU optimization, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the bridge crack detection method based on YOLO enhanced feature fusion and SIoU optimization as described in any one of claims 1 to 9.

Images