Deep learning corrosion detection with confidence

Corrosion poses a significant global economic challenge, estimated to cost 3–4% of GDP annually. Automated corrosion detection offers a potential solution for cost savings and risk mitigation. Recent advancements in deep learning and computer vision have spurred research in this area. However, a lack of publicly available, adequately sized and diverse datasets hinders the development of generalized models, leading to high rates of false positives and negatives in real-world applications. Previous studies using deep learning for corrosion detection have achieved limited success due to reliance on small, private datasets. This paper addresses this limitation by introducing a new deep learning corrosion detector capable of pixel-level segmentation, enhanced by three Bayesian variants providing pixel-level uncertainty estimates. These estimates, representing confidence levels, aim to improve decision-making in practical applications. The study utilizes a newly collected dataset comprising 225 images for training and validation.

Existing research on automated corrosion detection leverages deep learning, computer vision, and increased computing power. A comprehensive review by Nash, Drummond, and Birbilis (2018) summarizes deep learning applications in materials degradation, including corrosion detection. Prior work by the authors demonstrated the potential of deep learning for semantic segmentation in corrosion detection, but these models suffered from limitations including an F1-Score of only 0.55 due to insufficient training data and a tendency towards false positives when encountering data outside the training distribution. State-of-the-art models, while achieving an F1-score of approximately 0.71, also suffer from these issues due to the lack of a large, diverse dataset. The absence of publicly available labeled corrosion image datasets prevents the creation of generalized models. This study addresses the need for uncertainty estimation in practical applications by using Bayesian deep learning methods to quantify both epistemic and aleatoric uncertainty.

This study utilizes the High-Resolution Network (HRNetV2) architecture, modified for semantic segmentation, as the base model. Three Bayesian variants are implemented to provide uncertainty estimates: variational inference, Monte Carlo dropout, and an ensemble method. Variational inference introduces variational convolutional layers to sample from the normal distribution of convolution kernel weights. Monte Carlo dropout applies dropout during training and inference. The ensemble method trains multiple HRNetV2 models from different initializations. All three variants output both a corrosion prediction map and an aleatoric uncertainty map (treated as log variance). The Bayesian binary-cross-entropy loss function is used during training to optimize both prediction accuracy and aleatoric uncertainty. The dataset consists of 225 expertly labeled images of corrosion from an industrial site. Tenfold cross-validation is used for training, with pre-trained weights from HRNetV2 on the MS-COCO Stuff dataset used for transfer learning. The model is trained for 80 epochs using binary-cross-entropy loss, followed by further epochs using Bayesian binary-cross-entropy loss. During inference, the input image is passed through the model multiple times, and the mean and standard deviation of the outputs are used to estimate the prediction and epistemic uncertainty. A threshold of 0.75 is used to classify pixels as corrosion or background. The F1-Score is calculated for both raw and uncertainty-adjusted outputs. Sparsity curves are also generated to evaluate the uncertainty metrics.

The three Bayesian variants demonstrate high accuracy in corrosion detection, exceeding previous state-of-the-art performance. The average F1-scores across tenfold cross-validation are: Variational (0.84), Monte-Carlo dropout (0.80), and Ensemble (0.86). The ensemble method consistently performs best, particularly in terms of uncertainty quantification, which is considered most useful in practice. The epistemic uncertainty maps effectively highlight areas of low confidence (higher uncertainty in regions where corrosion is absent), while aleatoric uncertainty is higher in areas with shadows, overexposure, or poor image quality, which are also areas with less accurate corrosion predictions. Analysis of F1-Score vs. threshold curves reveals that uncertainty adjustment improves performance for some models and datasets, while other models may not benefit greatly. Sparsity curves indicate epistemic uncertainty tracks closely with the binary-cross-entropy loss. Notably, even with a relatively small training dataset (225 images), the models achieved impressive results, surpassing the estimated human accuracy for the task. When tested on an out-of-distribution (OoD) dataset of 14 images, the models showed lower F1-scores, particularly in images with dispersed corrosion, shadows, and overexposure, demonstrating limitations for diverse real world conditions. False positives were predominantly observed for foliage, water, and text.

This study demonstrates that incorporating Bayesian methods into deep learning significantly improves the reliability of corrosion detection models. The high accuracy achieved, even with a small dataset, highlights the potential of transfer learning and the effectiveness of the HRNetV2 architecture. The inclusion of uncertainty quantification addresses a critical limitation of previous deep learning approaches. The ensemble method emerges as the most robust and informative, providing higher accuracy and more reliable uncertainty estimates for improved decision-making. The findings suggest the practical utility of this approach for real-world corrosion inspection. The observed limitations on images with poor lighting conditions or unusual background features indicate the need for further dataset expansion and the potential benefits of multimodal data acquisition (such as infrared imaging).

This paper presents a novel deep learning model for corrosion detection that significantly improves upon existing methods by incorporating Bayesian techniques to estimate uncertainty. The ensemble method demonstrates superior performance, providing high accuracy and reliable uncertainty estimates for improved decision-making in real-world applications. Future research should focus on expanding the dataset size and exploring the integration of additional data modalities, like infrared imaging, to address the limitations identified in this study. Further investigation into multi-task learning and specialized training regimes could further enhance model robustness and accuracy.

The primary limitation of this study is the relatively small size of the training dataset (225 images). Although impressive results were obtained, a larger, more diverse dataset would likely lead to improved model generalization and accuracy. The model also shows sensitivity to variations in lighting and image quality, indicating that addressing these factors is crucial for reliable performance across various real-world scenarios. The OoD dataset testing was limited and should be expanded.