During steel strip production, mechanical forces and environmental factors cause surface defects of the steel strip. Therefore, detection of such defects is key to the production of high quality products. Moreover, surface defects of the steel strip cause great economic losses to the high-tech industry. So far, few studies have explored methods of identifying the defects, and most of the currently available algorithms are not sufficiently effective. Therefore, we developed an end-to-end defect detection model based on YOLO-V3. Briefly, the anchor-free feature selection mechanism was utilized to select an ideal feature scale for model training, replace the anchor-based structure, and shorten the computing time. Next, specially designed dense convolution blocks were introduced into the model to extract rich feature information, which effectively improves feature reuse, feature propagation, and enhances the characterization ability of the network. The experimental results show that, compared with other comparison models, the improved model proposed in this study has higher performance. For instance, the proposed model yielded 71.3% mAP on the GC10-DET dataset, and 72.2% mAP on the NEU-DET dataset.

在钢带生产过程中，机械力和环境因素导致钢带表面缺陷。因此，检测此类缺陷是生产高质量产品的关键。而且，钢带的表面缺陷给高新技术产业带来了巨大的经济损失。到目前为止，很少有研究探索识别缺陷的方法，并且大多数当前可用的算法都不够有效。因此，我们开发了基于YOLO-V3的端到端缺陷检测模型。简而言之，利用anchor-free特征选择机制为模型训练选择理想的特征尺度，替代anchor-based结构，缩短计算时间。接下来，在模型中引入了专门设计的密集卷积块以提取丰富的特征信息，有效提高了特征重用、特征传播，增强了网络的表征能力。实验结果表明，与其他对比模型相比，本研究提出的改进模型具有更高的性能。例如，所提出的模型在 GC10-DET 数据集上产生了 71.3% 的 mAP，在 NEU-DET 数据集上产生了 72.2% 的 mAP。