| 摘 要: 针对编织袋生产中纹理背景干扰强、材料反光及缺陷微小不规则等问题，提出一种高精度轻量化检测算法SHI-DETR。首先，构建了基于反向移动块(Inverted Residual Mobile Block，iRMB)的动态骨干网络，降低计算开销并增强特征提取鲁棒性；其次，引入高低频注意力(Hi-Lo Attention)替代多头自注意力，提升小尺度缺陷判别力并抑制反光干扰；最后，设计小目标增强金字塔 (Small Object Upgrade Pyramid，SOUP)，融合SPDConv与CSP-OmniKernel模块，强化对微小瑕疵的感知。实验表明，该算法在自建数据集上的mAP50达90.6%，较RT-DETR提升2.9%，参数量减少12.5%，在提高精度的同时显著降低了算力需求，具有较高的工程应用价值。 |
| 关键词: 目标检测 缺陷检测 RT-DETR 特征融合 高精度轻量化模型 |
|
中图分类号:
文献标识码:
|
| 基金项目: 浙江省高层次人才特殊支持计划(2023R5212) |
|
| Woven Bags Defect Detection Algorithm Improved Based on RT-DETR |
|
Dai Chengying, SHI Weimin, ZHANG Yongchao, SUN Lei
|
Zhejiang Sci-Tech University
|
| Abstract: To address challenges in woven bag production, such as strong textural background interference, material reflectivity, and minute irregular defects, a high-precision and lightweight detection algorithm named SHI-DETR is proposed. Firstly, a dynamic backbone network based on Inverted Residual Mobile Blocks (iRMB) is constructed to reduce computational overhead while enhancing the robustness of feature extraction. Secondly, Hi-Lo Attention is introduced to replace multi-head self-attention, enhancing the discrimination of small-scale defects and effectively suppressing reflective interference. Finally, a Small Object Upgrade Pyramid (SOUP) is designed by integrating SPDConv and CSP-OmniKernel modules to strengthen the perception of minute imperfections. Experimental results demonstrate that the proposed algorithm achieves an mAP50 of 90.6% on a proprietary dataset, surpassing the baseline RT-DETR by 2.9% while reducing the parameter count by 12.5%. The results indicate that the model significantly reduces computational demand while improving accuracy, demonstrating high value for engineering applications. |
| Keywords: Object detection Defect detection RT-DETR Feature fusion High-precision lightweight model |
