AMPNet: An advanced lightweight defect detection network for tiny steel sheets inside mobile phone in industrial scenarios

Mobile steel sheets, essential for smartphone structural reinforcement, often suffer from surface defects (irregular shapes/textures) caused by contaminants or manufacturing flaws. The Advanced Mobile Phone Steel Sheets Defect Detection Network (AMPNet) is a lightweight solution designed for real-time, high-accuracy defect detection in resource-limited industrial environments. AMPNet is structured around three main components specifically designed for tiny steel sheet defect detection. Firstly, contextual and spatial attention (CASA) merges strip convolution (capturing elongated defects) and a spatial attention module (SA Module) to prioritize critical spatial regions, enhancing defect feature extraction. Secondly, residual contextual and spatial (RCS) convolution integrates CASA with skip connections to mitigate gradient vanishing and improve multi-level feature fusion, boosting accuracy without computational overhead. Finally, the lightweight architecture utilizes the C2f-RVB backbone by replacing traditional convolutions with RepViTBlocks (RVB) and integrates the GhostHead detection head through a combination of GhostConv and coupled design, significantly slashing parameters and computational costs. Experimental results on the mobile phone steel sheet defect dataset demonstrate that AMPNet achieves 91.8% Average Precision at an Intersection over Union threshold of 0.5 (AP0.5) with only 3.0 million parameters (Params) and 5.8 billion floating point operations (FLOPs), outperforming a series of You Only Look Once (YOLO) models (e.g., YOLOv5 YOLO13), proving its efficacy and suitability for deployment on resource-limited embedded systems in industrial settings.