FMR-YOLO: An improved YOLOv8 algorithm for steel surface defect detection

Abstract

To address the insufficient feature extraction capability for steel surface defects in industrial production, as well as issues such as low detection speed and poor accuracy caused by large model parameters, a metal surface defect detection algorithm named FMR-YOLO, based on an improved YOLOv8n, is proposed. The algorithm incorporates a fast lightweight feature extraction structure, the number of parameters and computation of the model are reduced while preserving the spatial information, thus improving the target detection performance. A multi-scale feature fusion module is introduced, enabling the extraction of more comprehensive and richer features compared to traditional single-scale methods, to better support defect detection tasks. Additionally, a receptive field attention structure, Receptive Field Attention Neck, is designed in the Neck part to expand the model's receptive field and reduce computational complexity, significantly improving detection accuracy for small defects. This allows the model to effectively capture both global and local features in complex industrial scenarios. The effectiveness of the improved FMR-YOLO algorithm is validated on two industrial surface defect datasets: GC10-DET and NEU-DET. Experimental results show that the [email protected] detection accuracy has increased by 4.5% and 5.1% on the GC10-DET and NEU-DET datasets, respectively, with a parameter size of merely 2.7 M.