结合训练加速及注意力机制的桥梁检测算法  

A combination bridge detection algorithm of training acceleration and attention mechanism

作　　者：余培东 王鑫[1] 江刚武[1] 刘建辉[1] 徐佰祺 YU Peidong;WANG Xin;JIANG Gangwu;LIU Jianhui;XU Baiqi(School of Data and Target Engineering,Strategic Support Force Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]战略支援部队信息工程大学数据与目标工程学院,河南郑州450001

出　　处：《海洋测绘》2021年第3期57-61,共5页Hydrographic Surveying and Charting

摘　　要：在解析YOLOv4算法基础上,针对应用YOLOv4算法检测遥感影像桥梁目标任务中出现的训练耗时严重及精度较低缺陷,从算法训练过程和结构模块两方面进行优化:使用多尺度训练以及fp16训练策略降低算法训练成本,并引入SE模块和CBAM模块两种注意力机制提升算法检测精度。消融实验结果表明:优化训练策略能够有效降低算法训练成本,同时提高目标检测精度;相比较CBAM模块,SE模块对算法训练成本增加较小却能收获显著的检测精度提升,优化训练策略并嵌入SE模块的算法,使高分桥梁数据集和DOTA桥梁数据集的平均准确率分别提升1.4%和3%。该优化算法兼具效率和精度优势,为桥梁目标检测难题提供有效解决方法。Based on the analysis of YOLOV4 algorithm, aiming at the defects of using YOLOV4 algorithm in detecting the bridge target task of remote sensing images, such as serious time-consuming training and low precision, optimization is carried out from two aspects of algorithm training process and structure module: Multi-scale training and FP16 training strategy were used to reduce the training cost of the algorithm, and two attention mechanisms, SE module and CBAM module, were introduced to improve the detection accuracy of the algorithm. The ablation experiment results show that the optimized training strategy can effectively reduce the training cost of the algorithm and improve the target detection accuracy. Compared with CPAM module, SE module has a small increase in algorithm training cost but can gain significant improvement in detection accuracy. Optimization of training strategy and embedded algorithm of SE module can improve the average accuracy of high-resolution bridge data set and DOTA bridge data set by 1.4% and 3%,respectively. The optimization algorithm proposed in this paper has the advantages of both efficiency and precision, which provides an effective solution to the problem of bridge target detection.

关 键 词：遥感影像 桥梁目标检测 YOLOv4算法 训练策略优化 注意力机制 

分 类 号：P237.2[天文地球—摄影测量与遥感]