| 叶学义,韩卓,蒋甜甜,王佳欣,陈华华.多尺度反向校正增强和无损下采样的毫米波图像目标检测方法[J].电子测量与仪器学报,2025,39(4):50-61 |
| 多尺度反向校正增强和无损下采样的毫米波图像目标检测方法 |
| Multi scale reverse correction enhancement and lossless downsampling formillimeter wave image object detection method |
| |
| DOI： |
| 中文关键词: 隐匿目标检测 主动毫米波图像 多尺度反向校正特征增强 无损下采样 K-means++ |
| 英文关键词:hidden target detection active millimeter wave image multi scale inverse correction feature enhancement lossless downsampling K-means++ |
| 基金项目:国家自然科学基金项目（U19B2016，60802047）资助 |
|
|
| Author | Institution |
| Ye Xueyi | School of Communication Engineering，Hangzhou University of Electronic Science and Technology, Hangzhou 310018,China |
| Han Zhuo | School of Communication Engineering，Hangzhou University of Electronic Science and Technology, Hangzhou 310018,China |
| Jiang Tiantian | School of Communication Engineering，Hangzhou University of Electronic Science and Technology, Hangzhou 310018,China |
| Wang Jiaxin | School of Communication Engineering，Hangzhou University of Electronic Science and Technology, Hangzhou 310018,China |
| Chen Huahua | School of Communication Engineering，Hangzhou University of Electronic Science and Technology, Hangzhou 310018,China |
|
| 摘要点击次数: 54 |
| 全文下载次数: 78 |
| 中文摘要: |
| 针对毫米波图像中隐匿目标局部信噪比低导致检测障碍的问题，提出了一种基于多尺度反向校正增强和无损下采样的检测方法。首先设计了一种多尺度反向校正特征增强模块，在提取多尺度特征的多卷积核Res2Net上融合反向校正操作，实现大感受野区域对区域内相关小感受野区域卷积计算的反向校正，使得深度模型不仅能够获取更细粒度的特征，而且使宏观判别性表示贯穿多个尺度的特征信息；其次，利用非跨步卷积层的SPD-Conv实现无损下采样，缓解卷积下采样导致的信息丢失；最后，采用K-means++聚类算法生成适合隐匿目标检测任务的新锚框。实验在YOLO系列中选择了各方面性能都适中的YOLOv5s作为基础框架，针对现有的两种毫米波图像数据集（阵列图像集和线扫图像集）平均精度均值(mAP)mAP@0.5分别达到了96.21%和97.97%，相较于原版YOLOv5s以及YOLO其他系列等性能有显著提升。实验结果表明，该方法在不明显增加参数量和推理时间的同时，能够有效提升深度模型的检测性能。 |
| 英文摘要: |
| A detection method based on multi-scale inverse correction enhancement and lossless downsampling is proposed to improve the detection of hidden targets in millimeter wave images with low local signal-to-noise ratio. Firstly, a multi-scale reverse correction feature enhancement module was designed, which integrates the reverse correction operation on the Res2Net multi convolution kernel. This achieves the reverse correction of convolution calculation between large receptive field regions and related small receptive field regions, enabling finer-grained features across multiple scales. Secondly, utilizing non-step convolutional layers of SPD-Conv to achieve lossless downsampling and preserve more information. Finally, the K-means++ clustering algorithm generates new anchor boxes suitable for hidden object detection tasks. The experiment selected YOLOv5s, which has moderate performance in all aspects, as the basic framework, targeting two existing millimeter wave image datasets (array image dataset and line scan image dataset) mAP@0.5 reaching 96.21% and 97.97% respectively. Compared to the original YOLOv5s and other YOLO series, the performance has significantly improved. The experimental results show that this method can effectively improve the detection performance of deep models without significantly increasing the number of parameters and inference time. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
