MlcsNet: cross-space multi-scale method for detection of crane track defects in complex environments

Rail-mounted crane rail fastener defect detection is an essential component of routine safety inspections for cranes. However, due to the challenges posed by complex environments, conventional deep learning models for defect detection often suffer from high false-negative rates and suboptimal detection efficiency when identifying defects in crane rails. We propose a cross-space multi-scale method for the detection of crane track defects in complex environments to address these issues. In our work, we enhanced the feature extraction network of YOLOv8 by incorporating efficient multi-scale attention (EMA). Subsequently, the reparametrized gather-and-distribute neck (RepGDNeck) network was introduced as a feature fusion network to support cross-scale information integration. In addition, we applied EMA attention to the feature outputs at each scale of RepGDNeck. During the production of the dataset, we build a dataset of crane rail fastener defects (RFDDs). The experiments were conducted on the RFDD dataset we produced and the rail surface detect datasets (RSDD) public dataset. Experiments on the RFDD dataset show that the precision, recall, and mAP50 of our model reached 87.89%, 88.48%, and 88.27%, respectively. On the RSDD dataset, the precision, recall, and mAP50 of our model reached 87.12%, 93.12%, and 95.14%, respectively. Compared with other mainstream models, our model has better robustness and detection accuracy.