Steel Surface Defect Detection Algorithm Based on Improved YOLOv8

• Main Authors: Che, Jie, Yao, Xinrui, Wang, Yuqi, Li, Bohan, Xu, Hangrui, Liu, Jian
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Convolution; deep learning; Defect detection; Feature extraction; Inference algorithms; Kernel; large kernel depthwise convolution; Object detection; Steel; Steel industry; surface defect detection; Training
• ISSN: 2766-8665
• DOI: 10.1109/ICNSC62968.2024.10760099

Surface defect detection technology is a vital component of the steel industry that has garnered significant attention from the academic community in recent times. While modern methods with deep learning-based object detection provide better detection accuracy than traditional approaches, production requirements are often not met by their inference time. This study suggested an improved method based on the industry-leading object detection algorithm YOLOv8 to address this issue and effectively detect steel surface defects while keeping a reasonable balance between speed and accuracy. First, a Stem module based on YOLOv8s is added to downsample the image and lessen the effect of redundant features in the original image. In the meantime, the backbone network gains a large kernel depthwise convolution module called InceptionNeXt+, which improves the model's ability to extract features and has a larger effective receptive field. Lastly, the base model's SPPF module is changed to a RsimSPPF module, which increases the model's accuracy and speeds up inference. The improved YOLOv8 model confirms the efficacy of the improved method with experimental results on the NEU-DET dataset achieving a mAP of 76.9%, increasing by 3.3 compared to the baseline model.