A Lightweight Detection System for Aluminum Ingot Surface Defects

Abstract

Online inspection of surface defects (e.g., flash and slag inclusion) in aluminum ingot casting lines is crucial for ensuring product quality and production efficiency. However, this task faces multiple industrial challenges, including scarce real-world data, small defect scales, and high efficiency requirements. To address these issues, this work proposes a lightweight online detection model named YOLOv8n-Ghost-Simple Parameter-Free Attention Module (SimAM). First, a specialized dataset for aluminum ingot surface defects from a real production line is constructed. Targeted data augmentation strategies are designed to overcome the small-sample learning challenge. Subsequently, an improved lightweight detection model is proposed. By synergistically integrating the Ghost Convolution (GhostConv) lightweight convolution and the parameter-free SimAM attention mechanism, the model effectively enhances the feature discrimination capability for tiny defects while reducing model complexity. Experiments show that the proposed model achieves a mean Average Precision (mAP@0.5) of 63.8% on the in-house dataset, representing an 8.3% improvement over the baseline model (YOLOv8n). The parameter count and computational cost are reduced by 8.6% and 6.1%, respectively, while the inference speed is increased by 17.83%. A preliminary factory pilot application has verified the engineering utility of the system, demonstrating its stable integration into the aluminum ingot production line and its positive effect on improving product qualification rates. This study provides a systematic and empirically-validated technical pathway, which specifically addresses the compounded challenges of scarce data, tiny targets, and stringent efficiency demands, for intelligent online quality inspection in the aluminum ingot and similar metal casting industries.