Scholarlink

<p>Surface quality represents a critical factor in steel plate manufacturing. Current deep learning methods for steel surface defect detection generally demand extensive labeled training data. In practical applications, unlabeled defect data can be acquired more economically. Therefore, this paper presents Semi-SSDDet, a semi-supervised model for steel surface defect detection. The model utilizes unlabeled defect data to improve generalization capacity and detection accuracy. We develop an adaptive pseudo-label threshold mechanism that dynamically selects appropriate values based on real-time model predictions, improving pseudo-label quality for training. To address class imbalance caused by extensive unlabeled data usage, we design a dynamic label assignment strategy. Moreover, we propose a novel detector based on the YOLO series of architectures with two key innovations: (1) Dilated block enhances the retention of contextual defect feature information by stacking dilated convolutions with varying dilation rates; (2) Haar wavelet downsampling preserves critical defect details during the downsampling phase. Experimental results on NEU-DET and GC10-DET benchmarks demonstrate the framework achieves 0.603 and 0.441 mAP50 respectively when trained with merely 15% labeled data. Implementation code is available at <a href="https://github.com/ghlerrix/Semi-SSDDet">https://github.com/ghlerrix/Semi-SSDDet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semi-SSDDet: An adaptive semi-supervised steel surface defect detection

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Guanghu Liu,
  • Maoxiang Chu,
  • Rongfen Gong

摘要

Surface quality represents a critical factor in steel plate manufacturing. Current deep learning methods for steel surface defect detection generally demand extensive labeled training data. In practical applications, unlabeled defect data can be acquired more economically. Therefore, this paper presents Semi-SSDDet, a semi-supervised model for steel surface defect detection. The model utilizes unlabeled defect data to improve generalization capacity and detection accuracy. We develop an adaptive pseudo-label threshold mechanism that dynamically selects appropriate values based on real-time model predictions, improving pseudo-label quality for training. To address class imbalance caused by extensive unlabeled data usage, we design a dynamic label assignment strategy. Moreover, we propose a novel detector based on the YOLO series of architectures with two key innovations: (1) Dilated block enhances the retention of contextual defect feature information by stacking dilated convolutions with varying dilation rates; (2) Haar wavelet downsampling preserves critical defect details during the downsampling phase. Experimental results on NEU-DET and GC10-DET benchmarks demonstrate the framework achieves 0.603 and 0.441 mAP50 respectively when trained with merely 15% labeled data. Implementation code is available at https://github.com/ghlerrix/Semi-SSDDet.