LAM-YOLOv10: lightweight and multiscale feature enhancement for paint surface defect detection model

Defects on automotive paint surfaces not only compromise aesthetic quality but also diminish their protective capabilities. To address this issue, we propose the lightweight and multiscale feature enhancement You Only Look Once version 10 (LAM-YOLOv10) model, an advancement of YOLOv10s, specifically designed for rapid, accurate, and efficient detection of paint surface defects in industrial production line environments. The model integrates several innovative components: the dual-attention group normalization module, the triplet attention module, and the omni-dimensional dynamic convolution module. These elements facilitate defect detection through multiscale feature fusion, refined deep feature representation, and a lightweight overall network architecture. The LAM-YOLOv10 model achieves a mean average precision (mAP) at 50% intersection over union (mAP50) of 95.4% on the rearview mirror paint defect dataset, surpassing YOLOv10s by 0.9%. Validation on the NEU-DET dataset yields a mAP50 of 75.2%, indicating a 0.6% improvement over its predecessor. Notably, this model significantly reduces computational costs; floating-point operations and parameter counts are decreased by 20.8% and 11.3%, respectively. These optimizations render LAM-YOLOv10 highly suitable for online rapid defect detection within production lines where real-time performance and efficiency are paramount. The results demonstrate that LAM-YOLOv10 not only enhances detection accuracy but also provides a lightweight and efficient solution for industrial applications—enabling seamless integration into automated inspection systems with limited computational resources.