Abstract: To address the issues of slow surface defect detection speed in industrial aluminum profiles,as well as low accuracy in detecting small target defects such as scratches and paint bubbles,a small target aluminum profile surface defect detection model based on improved YOLOv8n was proposed. Firstly,cross-scale feature fusion was employed to enhance the network's ability to perceive multiscale information,by integrating features from different scales. Secondly,the C2f structure of the YOLOv8 backbone network was improved by incorporating RFAConv attention convolution. This was done to design a new structure to enhance the network's feature extraction ability and improve the accuracy of small object detection. In the neck region,lightweight RepGhost modules were used to reduce parameter count and computational complexity. Finally,the original YOLOv8's CIoU loss function was replaced with WIoUV3 to improve detection accuracy and convergence speed.Extensive experiments on aluminum profile defect datasets showed the model's effectiveness. It achieved a 2.5 % increase in mAP@0.5,The frame rate has reached 110 fps,and reduced weight file size by 16.2 %,surpassing the original model's performance. Experiments on the VOC2012 dataset and the Northeastern University hot rolled strip surface defect dataset demonstrated the algorithm's robustness.The refined model effectively meets industrial demands for aluminum profile detection.

Key words: surface defect detection, attention convolution, feature fusion, YOLOv8, small object detection