计算机科学
卷积神经网络
编码器
连铸
变压器
人工神经网络
特征(语言学)
深度学习
一般化
人工智能
曲面(拓扑)
模式识别(心理学)
源代码
编码(集合论)
水准点(测量)
铸造
噪音(视频)
算法
特征提取
作者
Shanlin Li,Zhehan Chen,Xiaodong Zhang,Tao Yang,Dianning Gong
标识
DOI:10.1088/1361-6501/ae140a
摘要
Abstract In steel manufacturing, steel continuous casting billets are essential intermediate products, making surface defect detection (SDD) a critical task. To address the challenge of limited industrial datasets that hinder the accuracy of deep learning methods, this paper introduces the steel continuous casting billets SDD (SCCB-SDD) dataset, constructed from real production data. We also propose STCNet, a novel detection model that combines Transformer and Convolutional neural network architectures. Designed for industrial applications, STCNet employs a star-operation backbone TinystarNet for feature extraction. It further incorporates an intra-scale interaction encoder with convolutional additive self-attention to strengthen same-scale feature learning. An adaptive cross-scale fusion encoder integrates multi-scale information effectively. In our experiments, raw images from the SCCB-SDD dataset were manually annotated, with defect regions cropped into compact, information-dense samples to reduce background noise and accelerate model convergence. Results show that STCNet outperforms mainstream competing methods, achieving a mean Average Precision (mAP@50) of 92.1% with only 7.1 million parameters and 15.9 giga floating-point operations per second (GFLOPs). Furthermore, to validate the generalization capability of the model, we conducted evaluations on the GC10-DET dataset, where STCNet achieved 66.4% mAP@50 and 30.1% mAP@75, outperforming mainstream competing methods. These results demonstrate that STCNet achieves a strong balance between accuracy and efficiency, emphasizing its practical value for industrial billet defect detection. The source code and dataset are publicly available at https://github.com/Lislttt/STCNet .
科研通智能强力驱动
Strongly Powered by AbleSci AI