In this work, I present a comprehensive study on the development and field validation of an intelligent detection system for bridge surface defects, leveraging the synergy between unmanned aerial vehicles (UAVs) and deep learning. The proposed system integrates a custom UAV platform equipped with a high-resolution camera, a multi‑stage image preprocessing pipeline, and a Faster R‑CNN based defect recognition model. The entire workflow was applied to a major suspension bridge located in a mountainous region of southwestern China. Field results demonstrate that the system achieves high detection accuracy for multi‑scale defects such as exposed reinforcement, spalling, and local damage, while improving inspection efficiency by approximately 12 times compared to conventional manual methods. Through this real‑world case study, I aim to provide a practical reference for the intelligent upgrade of bridge inspection technologies.
1. Introduction
Bridges serve as critical nodes in transportation infrastructure, and their long‑term safety directly affects public security and economic development. With the total number of bridges in China exceeding one million and the aging problem becoming increasingly prominent, traditional inspection methods that rely on manual visual checks and contact‑based measurements can no longer meet modern maintenance demands. These conventional approaches suffer from low efficiency, high cost, significant risks of working at height, and strong subjectivity in the assessment results. In this context, combining the efficient data acquisition capability of UAV drones with the powerful pattern recognition ability of deep learning has become an urgent need for the industry. UAV drones offer excellent maneuverability and can rapidly capture high‑resolution images of hard‑to‑reach and hazardous areas. Meanwhile, deep learning has demonstrated great potential in automatically identifying bridge defects from images.
However, most existing research focuses on optimizing algorithm models in isolation. There is a lack of systematic application cases that seamlessly integrate data acquisition, intelligent recognition, and engineering management into a streamlined workflow. To bridge this gap, I carried out a full‑scale field study on an extra‑long suspension bridge. This paper presents the entire process, from field reconnaissance, UAV drone aerial photography, image preprocessing, intelligent defect detection, to result information management. By analyzing the effectiveness and value of this intelligent inspection path in a real project, I hope to offer a practical reference for advancing the intelligent upgrade of bridge inspection techniques.
2. System Design of UAV Visual Inspection
2.1 Data Acquisition System
The core data acquisition system consists of a DJI Matrice 350 RTK UAV drone equipped with a RIE-SF10 inspection camera. The UAV drone features a new‑generation flight control system, high‑precision RTK positioning, and omnidirectional obstacle avoidance, ensuring flight stability in complex canyon wind fields. The RIE-SF10 camera provides ultra‑high spatial resolution of 0.1 mm, a three‑axis stabilized gimbal, and intelligent zoom capability, enabling fine image collection of critical components such as steel box girder welds and tower anchorages. Before the mission, airspace simulation and field reconnaissance were performed based on a 3D bridge model. Safe takeoff/landing points and layered segmented routes were planned. For the main cable, steel box girder, and tower, different flight modes (close flight, oblique photography, and surrounding path) were adopted, maintaining an image overlap rate above 80% and a flight speed of 2 m/s. A dynamic trajectory correction scheme was integrated to cope with high‑altitude wind disturbances, and the camera zoom function was used to cover hidden areas. This approach achieved comprehensive, high‑precision data collection for the high‑rise and long‑span bridge, providing multi‑dimensional data support for defect recognition.
2.2 Image Preprocessing and Enhancement
Raw images acquired during bridge surface inspection often suffer from uneven quality, complex backgrounds, or non‑uniform illumination, which directly affect the recognition accuracy of deep learning models. To address these issues, I employed two core methods: geometric transformation and histogram equalization.
Geometric transformations increase data diversity through spatial changes. Rotation and flipping were the primary operations. The rotation transformation rotates an image around its center by a specific angle θ. The coordinate mapping is given by:
$$
\begin{bmatrix} x’ \\ y’ \end{bmatrix} = \begin{bmatrix} \cos\theta & -\sin\theta \\ \sin\theta & \cos\theta \end{bmatrix} \begin{bmatrix} x \\ y \end{bmatrix}
$$
where x, y are the input coordinates, x’, y’ are the output coordinates, and θ is the rotation angle in radians. Flipping includes horizontal and vertical flips. These operations effectively expand the original image dataset and improve sample diversity.
Histogram equalization enhances image contrast by redistributing pixel intensity values. The process first computes the probability distribution of the gray‑level histogram:
$$
P_r(r_k) = \frac{n_k}{N}
$$
where rk represents the k-th gray level, nk is the number of pixels with that gray level, and N is the total number of pixels. Then the cumulative distribution function is calculated:
$$
s_k = \sum_{j=0}^{k} P_r(r_j)
$$
where sk is the value of the cumulative distribution at gray level rk. Finally, intensity remapping is applied. This method significantly enhances the contrast between defect regions and the background, making defect features more distinguishable. Consequently, the model’s robustness to complex environments is improved, and the false‑positive rate is reduced by 15%. These preprocessing steps provided a high‑quality data foundation for subsequent intelligent recognition, ensuring the identifiability of defect features and the stability of model training.
2.3 Defect Recognition Model
I selected the Faster R‑CNN object detection framework combined with an EfficientNet-B0 backbone network for bridge surface defect recognition. This model structure offers the dual advantages of lightness and high accuracy: EfficientNet-B0 optimizes network depth, width, and resolution through compound scaling, achieving efficient feature extraction while maintaining a small parameter count. The two‑stage design of Faster R‑CNN ensures high detection precision. The total computational cost of the model is reduced by about 60% compared to traditional networks, making it more suitable for deployment on mobile or edge computing platforms.
The recognition process consists of three core steps. First, the input image passes through the EfficientNet-B0 backbone to extract multi‑scale feature maps, using depth‑wise separable convolutions and MBConv modules to reduce computational redundancy. Then, candidate regions are generated on the feature maps. The model is optimized jointly by a classification loss and a regression loss. The total loss function is:
$$
L = \frac{1}{N_{cls}} \sum_i L_{cls}(p_i, p_i^*) + \lambda \frac{1}{N_{reg}} \sum_i p_i^* L_{reg}(t_i, t_i^*)
$$
where L is the total loss to be minimized, Lcls is the classification loss measuring the accuracy of class predictions, pi is the predicted probability of being an object, pi* is the ground‑truth label (1 for object, 0 for background), Ncls is the number of samples for classification loss, Lreg is the regression loss measuring bounding‑box localization accuracy, ti and ti* are the predicted and ground‑truth bounding‑box coordinate parameters, Nreg is the number of samples for regression loss, and λ is a hyper‑parameter balancing the two losses. Finally, the candidate regions are normalized through ROI Pooling and passed through fully connected layers to output defect categories with confidence scores and refined bounding boxes. The model was optimized for multi‑scale defects such as exposed reinforcement, spalling, and local damage by fusing shallow details with high‑level semantics through a Feature Pyramid Network, thereby enhancing small target detection capability.
2.4 Defect Localization and Classification
In the intelligent bridge surface defect recognition system, precise localization and accurate classification are core metrics for evaluating model performance. I used Intersection over Union (IoU) and Non‑Maximum Suppression (NMS) techniques to ensure accurate defect region localization, and adopted mean Average Precision (mAP) and recall as quantitative evaluation indicators. The localization mechanism measures the overlap between the predicted bounding box and the ground‑truth box using IoU, defined as:
$$
\text{IoU} = \frac{\text{Area of Overlap}}{\text{Area of Union}}
$$
where IoU is a dimensionless indicator of overlap, Area of Overlap is the area of overlap in square pixels, and Area of Union is the area of union in square pixels. I set the IoU threshold to 0.5: a detection is considered correct if the IoU between the predicted box and the ground‑truth box is at least 0.5. The NMS technique eliminates redundant detections by sorting boxes by confidence and suppressing those with high overlap, thus solving the problem of multiple detections for the same defect. The performance evaluation used mAP as the core metric, which is computed by integrating the precision‑recall curve.
3. Engineering Application of Intelligent Bridge Surface Defect Detection
3.1 Project Overview
The field study was conducted on a super‑long suspension bridge located in a mountainous province in southwestern China. This bridge is the first extra‑long span suspension bridge in the region, with a total length of 2235.16 m, a main span of 1196 m, a deck width of 33.5 m, a design load of Highway‑Class I, and a seismic fortification intensity of IX degrees. The deck is 285 m above the river level. The main structure is a twin‑tower single‑span steel box girder suspension bridge. The site lies in a strong seismic zone of a high‑plateau mountainous area, with steep terrain, deep valleys, a northern subtropical mountain monsoon climate, annual precipitation of 1350‑1500 mm, a rainy season with frequent fog, and complex geological conditions including interbedded basalt and clastic rocks, abundant groundwater, and well‑developed rock joints. The bridge exhibits typical defects such as fatigue cracks in steel box girder welds, corrosion in the cable system, and concrete spalling. Many inspection blind spots exist and safety risks are high, making traditional manual inspection insufficient. Selecting this bridge as the engineering background is ideal because it is representative of bridges in southwestern mountainous regions, and it provides a realistic scenario for validating UAV‑based 3D reconstruction, intelligent flight planning, and defect recognition technologies. The engineering urgency is high, data acquisition feasibility is good, and this study has significant implications for advancing intelligent bridge inspection.
3.2 Data Collection
During the field data acquisition, a total of 1478 raw images of the bridge surface were captured. After rigorous screening, 1336 high‑quality images were retained, yielding an effective image retention rate of 90.4%. Using geometric transformations and histogram equalization, I performed batch enhancement on these 1336 images, generating an expanded dataset of 5652 images — an augmentation factor of 324%. In the data labeling stage, three uniformly trained technicians independently annotated the images using the LabelImg tool. The inter‑annotator consistency for the three defect categories (exposed reinforcement, spalling, and local damage) reached 92.6%. The final dataset consisted of 5652 annotation files containing 12,874 valid bounding boxes, with an average of 2.28 defect targets per image. The dataset was randomly split into training and validation sets at a 9:1 ratio, providing a solid foundation for accurate model training and reliable evaluation. The validation set was constructed to maintain a balanced proportion of images from different bridge parts, ensuring the comprehensiveness and representativeness of the evaluation results. Notably, the dataset deliberately includes multi‑scale defects: large‑scale defects account for 28.3%, medium‑scale for 45.6%, and small‑scale for 26.1%. This diversity is crucial for enabling the model to handle defects of varying sizes in real projects.
The following table summarizes the dataset composition:
| Item | Value |
|---|---|
| Raw images captured | 1478 |
| Qualified images after screening | 1336 |
| Effective retention rate | 90.4% |
| Augmented images after preprocessing | 5652 |
| Augmentation factor | 324% |
| Number of annotation files | 5652 |
| Total valid bounding boxes | 12,874 |
| Average defects per image | 2.28 |
| Inter‑annotator consistency | 92.6% |
| Training‑validation split | 9:1 |
| Large‑scale defect ratio | 28.3% |
| Medium‑scale defect ratio | 45.6% |
| Small‑scale defect ratio | 26.1% |
3.3 Detection Result Generation
During the detection and result management phase, the trained model processed all 5652 preprocessed images in batch, automatically outputting structured data containing detailed defect information. For each image, the system recorded the bounding‑box coordinates, defect category, and confidence score for every detected object. The average confidence across all detections was 0.82, with a maximum of 0.96 and a minimum of 0.61, providing a reliable machine‑readable basis for subsequent quantitative analysis and statistics. Then, through a custom‑developed bridge defect information management system, these structured detection results were linked with the corresponding original images and defect‑annotated images. The system completed the batch import of all data within 10 minutes, processing 5652 records and associating 16,956 image files, with a generated database storage size of approximately 45.2 GB. The system automatically bound this batch of data to 12 basic attribute fields of the target bridge, establishing a complete database architecture including six linked tables: a bridge basic information table, a defect detection record table, and an image storage path table, among others. The bridge defect digital archive was thus created.
The following table presents the detection performance statistics on the validation set:
| Metric | Value |
|---|---|
| Average confidence | 0.82 |
| Maximum confidence | 0.96 |
| Minimum confidence | 0.61 |
| mAP (IoU=0.5) | 0.785 |
| Recall | 0.812 |
| Precision | 0.796 |
| F1 score | 0.804 |
| Processing time per image (GPU) | 0.12 s |
| Total processing time (5652 images) | 11.3 min |
3.4 Defect Detection Results and Discussion
To visually demonstrate the practical effect of the intelligent detection model, I selected representative image pairs from the target bridge for comparison. For the exposed reinforcement defect, the model successfully identified slender bare steel areas on the bottom of the girder. Even under shadow interference, the detection box was precisely located with a high confidence score. Similarly, for the spalling defect, the model accurately distinguished concrete detachment regions from the normal surface at the deck connections. For the local damage defect, the model effectively identified concrete corner breakage at the girder ends.
On complex backgrounds such as textured pier surfaces, the model exhibited strong anti‑interference ability. However, occasional false positives occurred when dirt stains on the bridge deck were misclassified as spalling. This was likely due to poor lighting conditions or high similarity between the stain and defect features. In addition, for very small or low‑contrast defects, such as tiny exposed reinforcement points, the model sometimes missed detection, revealing current limitations under extreme conditions. Overall, the intelligent recognition system achieved efficient and accurate defect localization in most scenarios, providing reliable support for engineering applications.
4. Comparison with Traditional Methods
To quantitatively evaluate the advantages of the proposed UAV‑deep learning approach, I compared it with conventional manual inspection methods in terms of efficiency, safety, and objectivity. The comparison is summarized in the following table:
| Aspect | Traditional Manual Method | UAV + Deep Learning Method |
|---|---|---|
| Inspection time for full bridge | Approximately 12 days (2 teams, 6 people) | 1 day (1 team, 3 people, including flight and computing) |
| Detection efficiency improvement | Baseline | ~12× |
| Safety risk | High (working at height, traffic interference) | Low (remote operation, no need for scaffolding) |
| Objectivity | Subjective, inspector‑dependent | Consistent algorithmic output |
| Coverage of hidden areas | Very difficult | Easy with close‑up zoom and flexible routes |
| Data digitization | Manual notes and photos | Automated digital records linked to GIS |
| Cost per inspection (est.) | High (labor, traffic control, equipment rental) | Moderate (UAV depreciation, software license) |
The results clearly demonstrate that the proposed system greatly reduces inspection time and improves safety, while providing objective and repeatable results. Although the initial equipment investment may be higher, the long‑term savings in labor and reduced disruption to traffic make it a cost‑effective solution for routine bridge inspections.
5. Conclusion
In this study, I successfully integrated UAV drone remote sensing technology with deep learning models to construct a complete intelligent detection system for bridge surface defects. The system was validated through a field application on a super‑long suspension bridge in a mountainous region. The results show that the system can efficiently identify multi‑scale defects including exposed reinforcement, spalling, and local damage, with a detection efficiency approximately 12 times higher than traditional methods, while significantly reducing the risk of working at height. By combining data acquisition, intelligent recognition, and information management into a streamlined workflow, the system enables digital archiving and visual analysis of inspection results, providing scientific support for bridge maintenance decisions.
Nevertheless, the study also reveals that the system’s adaptability to small‑scale defects and complex lighting conditions still requires improvement. Future research will focus on multi‑modal data fusion (e.g., combining infrared thermal imaging or LiDAR), lightweight model deployment for edge computing, and integration with bridge management systems (BMS) to further enhance accuracy and practicality. This work aims to push bridge inspection toward a more intelligent, standardized, and efficient future.
Through this real‑world case, I hope to provide a concrete reference for researchers and practitioners in the field of bridge engineering, demonstrating the tangible benefits of combining UAV drones with deep learning for infrastructure health monitoring.