| 万伟彤,李长峰,朱华波,陶友瑞.轻量化 CenterNet 网络的二维条码定位算法[J].电子测量与仪器学报,2022,36(5):128-135 |
| 轻量化 CenterNet 网络的二维条码定位算法 |
| Two-dimensional barcode positioning algorithm oflightweight CenterNet network |
| |
| DOI： |
| 中文关键词: 目标检测 二维条码定位 CenterNet 网络 轻量化网络 |
| 英文关键词:object detection two-dimensional barcode positioning CenterNet network lightweight network |
| 基金项目:国家重点研发计划项目（2020YFB2009400）资助 |
|
|
|
|
| 摘要点击次数: 1050 |
| 全文下载次数: 1150 |
| 中文摘要: |
| 针对复杂工业、物流运输场景中传统的二维条码定位算法效率和稳定性较低的问题,提出了一种基于轻量化的
CenterNet 网络的二维条码定位算法。 针对实际情况中二维条码尺寸变化问题,采用 CSPDarkNet53-tiny 作为主干网络并对其加
以修改。 添加 SPP 模块以提高网络精度,对 CenterNet 的上采样以及输出模块部分进行轻量化改造,使用 5×5 深度可分离卷积
代替普通卷积,并在训练时采用余弦退火学习率策略防止过拟合。 实验结果表明,在定位准确率仅比 YOLOv4-tiny 降低 0. 64%
的情况下,不仅能够避免传统算法准确率受背景影响大、鲁棒性不强等问题,而且实时推理速度可以达到 124 fps,可以更好的
用于低硬件配置下各种二维条码定位。 |
| 英文摘要: |
| Aiming at the low efficiency and stability of the traditional two-dimensional bar code positioning algorithm in complex industrial
and logistics transportation scenarios, a two-dimensional bar code positioning algorithm based on lightweight CenterNet network is
proposed, a lightweight CenterNet detection algorithm is proposed. In view of the size change of two-dimensional bar code in the actual
situation, CSPDarknet53-tiny is used as the backbone network and modified SPP module is added to improve the accuracy of the
network. The upsampling and detection head of CenterNet are lightweight transformed. 5 × 5 depth separable convolution is used to
replace ordinary convolution. The change strategy of learning rate during training adopts cosine annealing learning rate to prevent over
fitting. The experimental results show that the positioning accuracy is only 0. 64% lower than YOLOv4 tiny. It not only avoids the
problems that the accuracy of the traditional algorithm is greatly affected by the background and the robustness is not strong, the real-time
reasoning speed also reaches 124 fps, which can be better used for all kinds of two-dimensional bar code location under low hardware
configuration. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
