Defect detection on rail lines is essential for ensuring safe and efficient transportation. Current image analysis methods with deep neural networks (DNNs) for defect detection often focus on the defects themselves while ignoring the related context. In this work, we propose a fusion model that combines both a targeted defect search and a context analysis, which is seen as a multimodal fusion task. Our model performs rule-based decision-level fusion, merging the confidence scores of multiple individual models to classify rail-line defects. We call the model “hybrid” in the sense that it is composed of supervised learning components and rule-based fusion. We first propose an improvement to existing vision-based defect detection methods by incorporating a convolutional block attention module (CBAM) in the you only look once (YOLO) versions 5 (YOLOv5) and 8 (YOLOv8) architectures for the detection of defects and contextual image elements. This attention module is applied at different detection scales. The domain-knowledge rules are applied to fuse the detection results. Our method demonstrates improvements over baseline models in vision-based defect detection. The model is open for the integration of modalities other than an image, e.g., sound and accelerometer data.
Multimodal fusion is a multimedia technique that has become popular in the wide range of tasks where image information is accompanied by a signal/audio. The latter may not convey highly semantic information, such as speech or music, but some measures such as audio signal recorded by mics in the goal to detect rail structure elements or defects. While classical detection approaches such as You Only Look Once (YOLO) family detectors can be efficiently deployed for defect detection on the image modality, the single modality approaches remain limited. They yield an overdetection in case of the appearance similar to normal structural elements. The paper proposes a new multimodal fusion architecture built on the basis of domain rules with YOLO and Vision transformer backbones. It integrates YOLOv8n for rapid object detection with a Vision Transformer (ViT) to combine feature maps extracted from multiple layers (7, 16, and 19) and synthesised audio representations for two defect classes: rail Rupture and Surface defect. Fusion is performed between audio and image. Experimental evaluation on a real-world railway dataset demonstrates that our multimodal fusion improves precision and overall accuracy by 0.2 points compared to the vision-only approach. Student's unpaired t-test also confirms statistical significance of differences in the mean accuracy.