基于GSR-YOLOv7的探地雷达回波图像识别的研究

Research on GPR Echo Image Recognition Based on GSR-YOLOv7

摘要:

针对探地雷达回波图像特征不突出，工作人员现场解译时间长，准确率不高的问题，提出一种基于YOLOv7改进的回波图像识别方法GSR-YOLOv7。首先使用GhostConv替换YOLOv7卷积层的卷积核减小参数量；引入SimAM注意力机制提高特征学习能力；并通过对感受野模块RFB卷积核的复用，增大感受野；同时引入改进EIoU损失函数提高模型分类能力和回归精度。基于道路回波图像数据集的实验结果表明，改进后的模型MAP50达到了97.57%，MAP50:95达到了73.13%，较YOLOv7分别提高2.13%和8.46%，模型大小减小了35%。本研究提出的GSR-YOLOv7模型对于回波目标的检测效果较好，能适用于移动端系统，对于体积小、算力低的平台具有较大应用价值。

Abstract:

In order to address the problems of indistinct features, time-consuming field interpretation, and low accuracy in Ground Penetrating Radar (GPR) echo image recognition, this study proposes an improved method called GSR-YOLOv7 based on YOLOv7. The GSR-YOLOv7 method includes several modifications. First, GhostConv is used to replace the convolutional kernels in the YOLOv7 convolutional layers, thus reducing the number of parameters. Second, the SimAM attention mechanism is introduced to improve the feature learning capabilities. Third, the reuse of the receptive field block (RFB) convolutional kernel is implemented to enlarge the receptive field. Finally, an improved EIoU loss function is employed to improve model classification and regression accuracy. Experimental results on a dataset of road echo images demonstrate the effectiveness of the proposed approach. The GSR-YOLOv7 model achieves a MAP50 score of 97.57% and a MAP50:95 score of 73.13%, showing improvements of 2.13% and 8.46%, respectively, over YOLOv7. In addition, the model size is reduced by 35%. The GSR-YOLOv7 model exhibits excellent detection performance for echo targets and is suitable for use on mobile systems. It has significant value for platforms with limited processing power and small form factors.