Feature-based domain disentanglement and randomization: A generalized framework for rail surface defect segmentation in unseen scenarios

Deep neural network has demonstrated high-level accuracy in rail surface defect segmentation. However, deploying these deep models in actual inspection situations results in generalizability deficits and accuracy degradation. This phenomenon is mainly caused by the appearance difference between training and test images. To alleviate this issue, we propose a feature-based domain disentanglement and randomization (FDDR) framework to improve the generalization of deep models in unseen datasets. Specifically, two encoders are introduced to decompose the defect image into domain-invariant structural features and domain-specific style features. Only the domain invariant features are used to identify the defects. Additionally, we design a shuffle whitening module to remove the style information from the domain-invariant features. Meanwhile, the extracted style features are used to train a style variational autoencoder to randomly generate novel defect styles. Then, the randomly generated style features are combined with the domain-invariant features to obtain new defect images, thus expanding the training sample. We validate the proposed FDDR framework in six defect segmentation datasets. Extensive experimental results show that FDDR demonstrates robust defect segmentation performance in unseen scenarios and outperforms other state-of-the-art domain generalization methods. The source code will be released at https://github.com/Rail-det/FDDR.

Automated summary

This paper proposes a feature-based domain disentanglement and randomization (FDDR) framework to improve the generalization of deep models in unseen datasets. The framework successfully addresses the appearance difference issue between training and test images by decomposing the defect image into domain-invariant structural features and domain-specific style features. It also utilizes randomly generated samples for training to further expand the training sample.