Surface defect detection of stay cable sheath based on autoencoder and auxiliary anomaly location

As the critical protection component of stay cables, the stay cable sheath’s reliability relates to cable-stayed bridges’ operation safety. Due to the scarcity of defect samples, the subtlety of defect characteristics, and the high annotation cost in the actual industrial scene, existing methods are still challenging in automating the stay cable sheath’s surface defect detection. To solve these problems, a semi-supervised deep learning method based on autoencoder and assisted anomaly location (AEAL) is proposed. The defect detection and localization task can be performed end-to-end with only normal samples used for training. The model learns the differences between normal and abnormal samples by constructing effective positive and negative sample pairs. Meanwhile, it fuses and fine-adjusts image features in combination with the proposed auxiliary anomaly location module and multi-scale feature fusion module, thereby achieving accurate location of defects. The experimental results demonstrate that AEAL outperforms other advanced unsupervised and semi-supervised defect detection networks, yielding superior defect location results. It achieves the highest pixel-level detection accuracy in the dataset of stay cable sheath surfaces, making it well-suited for practical applications.