ISRM: introspective self-supervised reconstruction model for rail surface defect detection and segmentation

Abstract

The problems of intrinsic imbalance of the sample and interference from complex backgrounds limit the performance of existing deep learning methods when applied to the detection and segmentation of rail surface defects. To address these issues, an introspective self-supervised reconstruction model (ISRM) is proposed, which only requires normal samples in the training phase and incorporates the concept of self-supervised learning into an introspective autoencoder. The training framework of ISRM first extracts general features using a pretrained Feature Extractor. Subsequently, a Feature Transformer transfers the features to the target domain. Next, a synthetic defect embedder embeds Bessel-Gaussian random defects into the feature space. Finally, the asymmetric autoencoder reconstructs the rail surface features back into image space. The transformation of pretrained features into target-oriented features helps mitigate domain bias. Since defects exhibit higher commonality in the feature space relative to the image space, embedding synthetic defects into the feature space effectively improves training efficiency. Moreover, the adversarial training architecture enhances the clarity of reconstructed images. The impact of core parameters on the model performance is analyzed through ablation experiments. The results from comparative experiments demonstrate that ISRM achieves 98.5% and 97.2% accuracy on defect detection and segmentation tasks, respectively, reducing the error rate by 11.8% and 3.4% compared to the current state-of-the-art model.