基于Faster R-CNN的多任务增强裂缝图像检测方法

PDF下载HTML

Title:

Multi-task enhanced dam crack image detection based on Faster R-CNN

作者:

毛莺池, 唐江红, 王静, 平萍, 王龙宝

河海大学 计算机与信息学院，江苏 南京 211100

Author(s):

MAO Yingchi, TANG Jianghong, WANG Jing, PING Ping, WANG Longbao

College of Computer and Information, Hohai University, Nanjing 211100, China

关键词:

裂缝图像检测; Faster R-CNN; 多任务检测; 小目标检测; 迁移学习; 大坝安全; 区域建议网络; 小样本

Keywords:

crack image detection; Faster R-CNN; Multi-task detection; small targets detection; transfer learning; dam safety; RPN; small sample

分类号:

TP391

DOI:

10.11992/tis.201910004

摘要:

针对Faster R-CNN算法对多目标、小目标检测精度不高的问题，本文提出一种基于Faster R-CNN的多任务增强裂缝图像检测(Multitask Enhanced Dam Crack Image Detection Based on Faster R-CNN, ME-Faster R-CNN)方法。同时提出一种基于K-means的多源自适应平衡TrAdaBoost的迁移学习方法(multi-source adaptive balance TrAdaBoost based on K-means, K-MABtrA)辅助网络训练，解决样本不足问题。ME-Faster R-CNN将图片输入ResNet-50网络提取特征；然后将所得特征图输入多任务增强RPN模型，同时改善RPN模型的锚盒尺寸和大小以提高检测识别精度，生成候选区域；最后将特征图和候选区域发送到检测处理网络。K-MABtrA方法利用K-means聚类删除与目标源差别较大的图像，再在多元自适应平衡TrAdaBoost迁移学习方法下训练模型。实验结果表明：将ME-Faster R-CNN在K-MABtrA迁移学习的条件下应用于小数据集大坝裂缝图像集的平均IoU为82.52%，平均精度mAP值为80.02%，与相同参数设置下的Faster R-CNN检测算法相比，平均IoU和mAP值分别提高了1.06%和1.56%。

Abstract:

To improve the accuracy of the detection of multiple small targets using the faster R-CNN model, we propose a multi-task enhanced dam-crack image detection method based on faster R-CNN (ME-Faster R-CNN). In addition, to solve the problem of insufficient dam-crack samples, we propose a transfer learning method, multi-source adaptive balance TrAdaBoost based on K-means (K-MABtrA), to assist with network training. In the ME-Faster R-CNN, the ResNet-50 network is adopted to extract features from original images, obtain the feature map, and input a multi-task enhanced region-proposal-network module to generate candidate regions by adopting the appropriate size and dimensions of the anchor box. Lastly, the features map and candidate regions are processed to detect dam cracks. The K-MABtrA method first uses K-means clustering to delete unsuitable images. Then, models are trained using the multi-source adaptive balance TrAdaBoost method. Our experimental results show that the proposed ME Faster R-CNN with the K-MABtrA method can obtain an 82.52% average intersection over union (IoU) and 80.02% mean average precision (mAP). Compared with Faster R-CNN detection method using the same parameters, the average IoU and mAP values was increased by 1.06% and 1.56%, respectively.