CIA-YOLO: an improved steel cable defect detection model based on YOLOv11

To tackle small-scale features and blurred boundaries in steel cable defect detection, we propose CIA-YOLO, an enhanced YOLOv11–based model for high-precision industrial inspection. CIA-YOLO integrates three improvements: (1) a Convolutional Block Attention Module (CBAM) combining channel and spatial attention for finer feature extraction; (2) a dynamic-scaling Inner-IoU loss function enhancing robustness and localization accuracy; and (3) an optimized Adaptive Kernel Convolution (AKConv) with a refined C3k2 module for stronger multi-scale modeling. On a dataset of broken wire, corrosion, and wear defects, CIA-YOLO achieved mAP@0.5 of 88.5%, 97.7%, and 99.5%, and Recall of 88.4%, 96.4%, and 99.8%, respectively. Overall, it recorded a mAP@0.5 of 95.2% and Recall of 94.8%, notably with 99.8% Recall on small wear defects. Compared to baseline YOLOv11 variants, CIA-YOLO delivers superior accuracy and faster inference, enabling real-time, in-line quality monitoring and safety assurance in engineering settings.