Parts Surface Defect Detection Algorithm Based on Improved <scp>YOLOv8s</scp>

ABSTRACT Inspection of surface defects in industrial components is vital to ensuring product quality and operational safety. This study presents an improved YOLOv8s‐based deep learning model designed to detect fatigue and linear cracks on part surfaces with high localization and classification accuracy. Unlike conventional manual inspection, the proposed system achieves real‐time detection while maintaining low computational cost. A custom dataset comprising high‐resolution images of concrete and metallic textures captured under diverse environmental conditions was used for training and evaluation. Data augmentation techniques such as Mosaic and Contrast Limited Adaptive Histogram Equalization (CLAHE) were employed to enhance generalization, and the model was trained for 100 epochs to ensure stable convergence. The enhanced YOLOv8s architecture integrates the Convolutional Block Attention Module (CBAM) and Ant Colony Optimization (ACO) for intelligent feature selection, resulting in improved learning efficiency and reduced overfitting. Experimental results show that the model achieves mAP@IoU = 0.5 of 0.9626 and [email protected]:0.95 of 0.7803, with validation box and class losses of 0.6849 and 0.5377, respectively. Precision and recall values of 0.926 and 0.923 demonstrate excellent detection completeness and low false positives. Comparative analysis with BsS‐YOLO and Efficient YOLOv8‐ES confirms the superior accuracy and efficiency of the proposed approach. Visual inspection further validates the model's robustness in identifying diverse crack patterns under challenging surface conditions. The improved YOLOv8s model thus offers a scalable, real‐time, and accurate solution for automated defect detection in industrial applications. Future work will focus on multi‐class defect recognition and deployment on lightweight edge‐computing devices.