首页|融合改进Transformer的车辆部件检测方法

融合改进Transformer的车辆部件检测方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
为有效解决车辆部件检测中模型由于特征提取不充分以及候选框未能充分利用导致的错检、漏检等问题,提出了融合改进 Transformer 的车辆部件检测方法.首先将多头自注意力和双层路由注意力结合,提出了关键区域多头自注意力(KR-MHSA);然后将基线模型(Mask R-CNN)中ResNet的最后一层与KR-MHSA进行残差融合,提升了模型的基础特征提取能力;最后通过改进的Swin Transformer对模型生成的候选框进行特征学习,使模型更好地理解不同候选框之间的差异和相似性.实验在构建的 59 类车辆部件数据集上进行,对比实验结果证明,本文模型在检测和分割效果上均优于其他先进实例分割模型.相较于基线模型,检测准确率提高了 4.47%,分割准确率提高了 4.4%,有效地解决了车辆部件检测中特征提取不足和候选框未充分利用导致的错检、漏检和实例分割精度较低的问题,使保险公司能够更准确、更高效地更换损坏的部件,提高索赔效率.
The vehicle parts detection method enhanced with Transformer integration
To effectively address issues such as false detections and missed detections caused by insufficient feature extraction and inadequate utilization of candidate boxes in vehicle component detection models,an improved Transformer-based method for vehicle component detection was proposed.Firstly,by combining multi-head self-attention and bi-layer routing attention,a key region multi-head self-attention(KR-MHSA)mechanism was introduced.Secondly,the final layer of ResNet in the baseline model(Mask R-CNN)was integrated with KR-MHSA using residual fusion,enhancing the basic feature extraction capabilities of the model.Finally,the improved Swin Transformer was employed for feature learning on the candidate boxes generated by the model,enabling the model to better understand the differences and similarities between various candidate boxes.Experiments conducted on a constructed dataset of 59 vehicle component categories demonstrated that the proposed model outperformed other state-of-the-art instance segmentation models in both detection and segmentation performance.Compared to the baseline model,the detection accuracy improved by 4.47%,and the segmentation accuracy improved by 4.4%.This effectively resolved the issues of insufficient feature extraction and inadequate utilization of candidate boxes in vehicle component detection,leading to more accurate and efficient replacement of damaged parts by insurance companies,thus improving claims processing efficiency.

vehicle partsdeep learninginstance segmentationMask R-CNNfeature extractionmulti-head self-attentionbi-level routing attention

翟永杰、李佳蔚、陈年昊、王乾铭、王新颖

展开 >

华北电力大学自动化系,河北 保定 071003

车辆部件 深度学习 实例分割 Mask R-CNN 特征提取 多头自注意力 双层路由注意力

国家自然科学基金项目河北省自然科学基金面上项目中央高校基本科研业务费专项资金项目中央高校基本科研业务费专项资金项目

62373151F20235020102023JC0062024MS136

2024

10.11996/JG.j.2095-302X.2024050930

图学学报
中国图学学会

图学学报

CSTPCD北大核心
影响因子：0.73
ISSN：2095-302X
年,卷(期)：2024.45(5)