| 邵自强,魏利胜,武涛.改进YOLOv11n的高效交通实例分割算法[J].电子测量与仪器学报,2026,40(2):95-106 |
| 改进YOLOv11n的高效交通实例分割算法 |
| Efficient traffic instance segmentation algorithmbased on improved YOLOv11n |
| |
| DOI： |
| 中文关键词: 交通场景 目标分割 小波变换 注意力机制 NMS算法 YOLOv11n |
| 英文关键词:traffic scene target segmentation wavelet transform attention mechanism NMS algorithm YOLOv11n |
| 基金项目:安徽省教育厅重大项目(KJ2020ZD39)、安徽省高等学校省级质量工程项目(2023cxtd057)资助 |
|
|
|
|
| 摘要点击次数: 435 |
| 全文下载次数: 251 |
| 中文摘要: |
| 针对交通场景下目标分割精度低和掩膜质量差的问题,提出一种改进YOLOv11n的高效交通实例分割算法。
首先,在主干网络的C3k2模块中融合小波变换卷积WTConv,构建C3k2-WTConv模块,以高效扩展感受野并增强低频特征提取;其次,设计特征交互增强AIFI-LA模块,降低快速空间金字塔池化（SPPF）的多尺度计算冗余,并提高处理长序列和保留关键特征信息能力;然后,提出特征重校准EMCSA模块,并嵌入至特征重组上采样算子（CARAFE）中,构建CARAFE-EMCSA模块重构上采样,以增强环境特征的捕获能力和特征图的整体判别性;最后,将Soft-NMS与DIoU-NMS相融合并替换原
非极大值抑制算法（NMS）,在保留更多高质量边界框的同时,利用相对位置信息进一步优化选择,提升边界框精度.实验结果表明,在Cityscapes数据集上,与原模型相比,边界框精度mAP@0.5和mAP@0.5:0.95值分别提高了9.2%和8.5%,分割掩膜精度mAP@0.5和mAP@0.5:0.95值分别提高了10.6%和8.8%,在BDD100K数据集上,边界框精度mAP@0.5和mAP@0.5:0.95值分别提高了5.1%和7.4%,分割掩膜精度mAP@0.5和mAP@0.5:0.95值分别提高了4.5%和6.6%.由此可知,所提方法在交通场景分割方面的有效性. |
| 英文摘要: |
| To address the problems of low target segmentation accuracy and poor mask quality in traffic scenes, an improved YOLOv11n efficient traffic instance segmentation algorithm, ETIS-YOLO, is proposed. Firstly, the C3k2-WTConv module is constructed by fusing the Wavelet Transform Convolution into the C3k2 module of the backbone network to efficiently expand the receptive field and enhance low-frequency feature extraction; Secondly, the feature interaction enhancement AIFI-LA module is designed to reduce the multi-scale computational redundancy of spatial pyramid pooling-fast (SPPF) and improve its ability to handle long sequences and preserve key feature information; Additionally, the feature recalibration EMCSA module is proposed and embedded into the up-sampling operator content aware reassembly of features （CARAFE） to form a CARAFE-EMCSA module, which reconstructs the up-sampling process to enhance the capture of contextual features and the overall discriminability of feature maps; Finally, Soft-NMS and DIoU-NMS are fused and replaced with the original non-maximum suppression （NMS）, which further optimizes the selection and improves the accuracy of the bounding boxes by utilizing relative position information while retaining more high-quality bounding boxes. The experimental results show that on the cityscapes dataset, the bounding box accuracy mAP@0.5 and mAP@0.5:0.95 values are improved by 9.2% and 8.5%, and the segmentation mask accuracy mAP@0.5 and mAP@0.5:0.95 values are improved by 10.6% and 8.8%, respectively, compared with the YOLOv11n model; on the BDD100K dataset, the bounding box accuracy mAP@0.5 and mAP@0.5:0.95 values are improved by 5.1% and 7.4%, and the segmentation mask accuracy mAP@0.5 and mAP@0.5:0.95 values are improved by 4.5% and 6.6%, respectively. It can be seen that the proposed method is effective in traffic scene segmentation. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
