|
| 热轧带钢表面轻量级缺陷识别算法研究 |
| Research on Lightweight Defect Identification Algorithm of Hot Rolled Strip Surface |
| |
| DOI：10.16018/j.cnki.cn32-1650/n.202404007 |
| 中文关键词: 热轧带钢 轻量级模型 注意力模块 |
| 英文关键词: hot rolled strip lightweight model attention module |
| 基金项目: |
|
| 摘要点击次数: 50 |
| 全文下载次数: 43 |
| 中文摘要: |
| 热轧带钢是重要的生产材料之一，然而现有的基于深度学习的缺陷识别算法识别效率低、模型参数冗余，无法满足生产要求。本文对轻量级模型SqueezeNext的基础模块进行优化重构，通过加入不同分辨率的卷积通路提升模型特征提取能力。同时提出一种下采样空间通道注意力模块，将浅层信息施加注意力后传入模型深层，进一步提升识别精度。实验结果显示，最优模型在东北大学带钢表面缺陷数据集上的识别准确性达到98. 96%，优于同类轻量级模型，且相较于同准确率的其他模型，拥有更少的参数量和浮点运算量。 |
| 英文摘要: |
| Hot rolled strip is one of the important production materials. However, the existing defect identification algorithm based on deep learning has low identification efficiency and redundant model parameters, which can not meet the production requirements. In this paper, the basic module of the lightweight model SqueezeNext is optimized and reconstructed, and the feature extraction ability of the model is improved by adding convolution paths with different resolutions. At the same time, a down-sampling spatial channel attention module is proposed, which applies attention to shallow information and then transmits it to the deep layer of the model to further improve the recognition accuracy. The experimental results show that the identification accuracy of the optimal model on the strip surface defect data set of Northeastern University reaches 98. 96%, which is superior to similar lightweight models, and has less parameters and floating-point operations than other models with the same accuracy. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|