Real-Time Defect Detection in Intelligent Manufacturing Systems Based on YOLOv8

Abstract

In high-end manufacturing, deep learning-based detection systems are gradually being used for automated quality control. In order to create a fast, real-time defect detection system, a robust backbone network and a powerful data preparation module were established in this study. The large intra-class variability, dynamic imaging environment, and strict requirements for real-time inference are serious issues faced by current work in industrial defect detection. Geometric transformations, photometric variations, and synthetic defect generation are types of data augmentation techniques that can enhance the model's generalization ability. To ensure stable convergence and avoid overfitting, the model training will use an adaptive learning rate scheduler, a composite loss function, and reasonable regularization methods. Architecture optimization: Multi-scale feature fusion and anchor-free detection heads improve the sensitivity to sparse surface defect recognition. Using the same method across multiple industries, each with its own products and defects. According to the experimental results, the average accuracy, recall rate, and processing speed of these two frameworks are all above average. The new pipeline runs stably on GPU servers and in edge cases. By using error and ablation analysis, the individual impact of the new components was revealed, providing a reference for future research. In summary, the aforementioned research indicates that the integrated framework can meet the high demands of large-scale automated quality inspection in modern manufacturing.