Abstract: During the manufacturing process of aluminum profiles,defects such as scratches and dirt spots may occur on the surface due to factors like materials or processing techniques,directly affecting the usability of aluminum profiles. It analyzes the characteristics of surface defects in aluminum profiles and compares existing deep learning-based object detection algorithms. Based on the YOLOv8 network model,an aluminum profile surface defect detection algorithm integrating GhostBottleneck and attention mechanism is proposed. By introducing Ghost into the Bottleneck layer and replacing some of the convolutional structures in the backbone network with DWConv,the complexity of the model is reduced while ensuring detection accuracy. Furthermore,the ECA attention mechanism is added to the YOLOv8 detection head module to enhance the detection accuracy of the model. It conducts experimental verification,and the experimental results show that the accuracy of the improved algorithm reaches 0.932,representing a 5.9 % improvement compared to the basic YOLOv8 algorithm. Moreover,the number of model operation parameters is reduced by 24 %. The overall performance meets the industrial requirements for the accuracy and speed of defect detection in aluminum profiles.

Key words: aluminum profile, surface defect detection, attention mechanism, YOLOv8, model lightweighting, GhostBottleneck