Abstract
Recognition of defects in concrete infrastructure, especially in bridges, is a costly and time-consuming crucial first step in the assessment of the structural integrity. Large variation in appearance of the concrete material, changing illumination and weather conditions, a variety of possible surface markings as well as the possibility for different types of defects to overlap, make it a challenging real-world task. In this work we introduce the novel COncrete DEfect BRidge IMage dataset (CODEBRIM) for multi-target classification of five commonly appearing concrete defects. We investigate and compare two reinforcement learning based meta-learning approaches, MetaQNN and efficient neural architecture search, to find suitable convolutional neural network architectures for this challenging multi-class multi-target task. We show that learned architectures have fewer overall parameters in addition to yielding better multi-target accuracy in comparison to popular neural architectures from the literature evaluated in the context of our application.
Introduction
Inspecting concrete bridges for structural integrity is a slow, expensive, and manual process. While deep learning has revolutionized object recognition, its application to infrastructure maintenance has lagged. Most existing datasets focus solely on a single defect type—usually cracks—and ignore the messy reality of the field: defects often overlap. For instance, a crack might appear inside a spalled area with exposed rebar, requiring a model capable of multi-target classification rather than simple single-label categorization.
Furthermore, standard neural networks (like those designed for ImageNet) are optimized for identifying distinct objects such as dogs or cars. These architectures may not be optimal for identifying subtle material textures and degradations. This paper tackles the challenge of multi-target classification where a single image patch might contain multiple, non-exclusive defect categories, and seeks to find the best neural architecture for this specific domain.
The CODEBRIM Dataset
The core contribution of this work is the CODEBRIM (COncrete DEfect BRidge IMage) dataset. It consists of 1,590 high-resolution images sourced from 30 unique bridges, providing a diverse set of environmental conditions and scales. The dataset includes 5,354 defect bounding boxes and 2,506 background boxes.
The researchers identified five mutually non-exclusive defect types: crack, spallation, efflorescence (calcium leaching), exposed reinforcement bars, and corrosion (stains). Images were acquired at varying scales and weather conditions, including some captured via Unmanned Aerial Vehicles (UAVs) to access hard-to-reach areas, ensuring the dataset reflects real-world inspection challenges.
Methodology: Meta-Learning and NAS
The authors move away from manually designing "human-intuition" architectures and instead use Meta-Learning, specifically Neural Architecture Search (NAS). They compare two reinforcement learning-based approaches:
- MetaQNN: Uses Q-Learning to "learn" how to build increasingly accurate CNNs layer by layer.
- ENAS (Efficient Neural Architecture Search): A faster approach where different candidate architectures share weights, drastically reducing the computational cost of the search.
To handle the multi-target nature of the task, these methods were adapted to use a Sigmoid activation and Binary Cross-Entropy loss for each class. This ensures the model can predict multiple defects simultaneously. They also utilized Stochastic Gradient Descent with Warm Restarts (SGDWR) to find better local minima during training.
Results and Discussion
The meta-learned architectures achieved multi-target test accuracies of up to ~72%, outperforming standard baselines like AlexNet and VGG-A. A significant finding was the "ImageNet Bias" gap: improvements in standard benchmarks do not always translate to specialized material tasks. Some "superior" literature models actually performed worse or overfit heavily on concrete defects.
One of the most striking results was the parameter efficiency of the discovered models. The search algorithms favored architectures that were surprisingly shallow and narrow compared to modern giants. For example, the MetaQNN-2 model uses only 1.22 million parameters, while VGG-D uses over 134 million parameters to achieve lower accuracy on the same task. Even the high-performing DenseNet-121 was matched or beaten by the NAS-generated models, which had significantly lower computational overhead.
Conclusion
This research demonstrates that meta-learning can produce specialized CNN architectures that are both more accurate and significantly more efficient than general-purpose models for infrastructure inspection. While this study focuses on classifying pre-cropped image patches, future work will involve integrating these models into full object detection pipelines and exploring semantic segmentation for pixel-level defect severity analysis. The efficiency of these meta-learned models makes them ideal candidates for real-time inspection via UAVs.
Summary updated