Industrial Defect Detection Model Based on Multi-Scale Feature Extraction and Fusion

In the realm of industrial manufacturing, identifying product defects presents numerous obstacles, mainly focusing on the accuracy, efficiency, dependence on manual inspection, and the presence of a large amount of noise and interference in the detected image that affects the detection effect. In order to cope with the many complex situations faced by surface defect detection, a new model named IDDM is proposed, and its performance is proved to be good through comparative experiments and ablation experiments, and the model has been substantially enhanced by building upon the foundational framework of YOLOv8. IDDM model introduces the MBCC2f module and the GFPN module. By embedding multiple MBConv modules in the C2f structure, the MBCC2f module realizes multilevel feature extraction and fusion, thereby improving the accuracy and stability of detection in complex industrial environments. The GFPN module introduces the Spatial Pyramid Pooling (SPP) module into CSPStage. This enhancement boosts the model's capability to identify defects across various scales by leveraging multi-scale pooling techniques. The experimental outcomes demonstrate that the IDDM model achieves a detection accuracy of 97.1%, the recall rate reaches 97.3%, the mAP50 reaches 97.4%, and the mAP50:95 reaches 55.4%, which are better than other YOLO models in terms of evaluation performance indicators. The IDDM model not only shows good performance in terms of accuracy, recall, and mAP metrics, but also achieves adaptability to complex industrial environments by introducing the MBCC2f module for multilevel feature extraction and fusion and the GFPN module for multi-scale defect detection, contributing to more efficient and accurate defect detection. At the same time, this research demonstrates the potential of today's deep learning-based methods to revolutionize industrial defect detection, providing scalable and reliable solutions for a wide range of industrial applications.