RepGhost-YOLO: A lightweight YOLOv11 variant for real-time chip defect detection

Semiconductor chip surface defect detection is a core task for industrial automated optical inspection (AOI). However, it remains challenging because the defects are often small (pixel area < 32 × 32), irregularly shaped, and have low contrast against repetitive background textures. Although the latest YOLOv11 has achieved over 90% accuracy in general object detection benchmarks, it still exhibits significant performance degradation in industrial chip defect scenarios, especially for tiny low-contrast defects and boundary localization accuracy, which cannot meet the strict requirements of AOI systems for zero missed detection and sub-pixel level measurement. In this study, we propose a lightweight object detector for real-time chip defect detection that jointly enhances feature extraction, feature re-weighting, and bounding-box regression. First, a RepGhostNet (a lightweight re-parameterized network) backbone is added to improve multi-scale feature representation and reduce computational redundancy via structural re-parameterization. Second, an improved coordinate-attention (ImCA) module is designed, which integrates dual pooling, cross-directional feature fusion, adaptive multi-scale gating, and high-frequency residual enhancement to strengthen boundary-sensitive feature representation. Third, we design an improved MPDIoU (Modified Pooled Distance Intersection over Union) loss, namely ImMPDIoU, which introduces scale-adaptive corner constraints, dynamic re-weighting, and numerical clamping to stabilize training and improve localization accuracy for multi-scale defects. Experiments with 5 independent runs (random seeds 0-4) on a self-built chip defect dataset and two public industrial defect datasets show that the proposed model achieves a mean Average Precision at IoU threshold 0.5 (mAP@0.5) of 99.11±0.08% and a mean Average Precision across IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95) of 67.55±0.12%, outperforming the YOLOv11 baseline by 2.41 and 11.16 percentage points, respectively. Extensive benchmarking against the latest 2024-2026 industrial lightweight SOTA detectors demonstrates that our model achieves the state-of-the-art accuracy-efficiency trade-off. Cross-domain generalization experiments on infrared small targets further verify the model's strong adaptability to non-defect detection scenarios. Grad-CAM visualization analysis confirms that the model focuses on actual defect regions rather than background artifacts, providing reliable interpretability for industrial deployment. The model maintains a lightweight complexity of 2.38M parameters and 9.46 GFLOPs, and achieves real-time inference speeds of 128 FPS on NVIDIA RTX 4090, 21 FPS on NVIDIA Jetson Xavier NX, and 8 FPS on NVIDIA Jetson Nano. Comprehensive end-to-end real-time pipeline validation confirms that the model can operate at 16.5 FPS on Jetson Xavier NX and 6.8 FPS on Jetson Nano in practical industrial AOI environments. These results demonstrate that our proposed design achieves an excellent balance between detection accuracy and computational efficiency, and is fully suitable for real-time deployment in industrial AOI systems.