摘要
为了保护生态环境和人员安全,周期性地对煤矿采空区的地裂缝进行检测至关重要。传统的地裂缝人工巡检方法耗时长、效率低下,且无法保障巡检人员安全。航拍视角下的地裂缝与路面、桥梁等裂缝具有相似特征,都呈现出狭长形状。但是,一般煤矿采空区地裂缝分布在山区,其背景和噪声更为复杂,导致传统图像处理方法性能不佳。因此,提出一种基于混合域注意力变形卷积网络的地裂缝检测方法。针对地裂缝的狭长特性,引入变形卷积,在特征提取中自适应地确定感受野的范围。混合域注意力机制为特征图中不同通道和不同空间位置的特征信息赋予相应权值,来强化特征图中特定通道和空间位置对地裂缝检测的贡献程度。其中,通道域注意力模块利用通道池化,并经过卷积和激活函数为每个通道生成0~1的权值,强化了特定通道对检测的贡献;空间域注意力模块利用空间池化,并结合变形卷积训练得到每个空间位置的权值,使得模型能更有效地获取空间信息。基于此,给出注意力机制引导的地裂缝检测一般框架,应用无人机搭载高清摄像头采集采空区图像进行地裂缝检测。实验中,所提方法与一阶段检测模型SSD300、SSD512和RetinaNet相比,平均精度分别提升了0.246,0.101和0.034,与多阶段检测模型Faster R-CNN和Cascade R-CNN相比,精准率分别提升了0.300和0.271。自我对比实验中,引入混合域注意力提升了0.195的精准率和0.038的平均精度;较大尺寸的输入图像各方面性能更高。结果表明,所提方法通过结合变形卷积与注意力机制,相比于其它检测方法准确率更高,训练过程更平稳。
Periodical detection of ground crack over mine goaf is necessary for environmental protection and human safety.Traditional detection method is time-consuming,low efficiency,and cannot guarantee the safety of inspector.Ground crack in aerial images share similar characteristics with road,bridge and so on,they all are shallow and long.However,the ground crack is normally located in mountain area,which has complicated background,results in poor performance by traditional image processing.Therefore,an aerial image detection method of ground crack over goaf based on deformable convolutional network with hybrid domain attention is proposed.In view of the narrow and long characteristics of ground crack,deformable convolution is introduced to adaptively determine the range of receptive field in feature extraction.Hybrid domain attention mechanism assigns the corresponding weight to the feature information of different channels and spatial locations in the feature map,with the purpose to strengthen the ground crack detecting contribution of specific channels and spatial locations in the feature map.Channel domain attention mechanism uses channel pooling and generates weights from 0 to 1 for each channel through convolution and activation function,which strengthens the detection contribution of specific channel.Spatial domain attention mechanism uses spatial pooling and generates weights for each spatial location after training with deformable convolution,making model more efficient to capture spatial information.Therefore,the attention guided common detection framework of ground crack is proposed,unmanned aerial vehicle with high-resolution is employed to capture goaf images and detect ground cracks.In experiments,the average precision of the proposed method is improved by 0.246,0.101 and 0.034 compared with single stage detection model SSD300,SSD512 and Retinanet,respectively.Compared with multi-stage detection model Faster R-CNN and Cascade R-CNN,the precision is improved by 0.300 and 0.271,respectively.In the self-comparison experiments,the precision and average precision are improved by 0.195 and 0.038 by introducing hybrid domain attention,and higher performances in all aspects are achieved with input image of larger size.The results show that the proposed method are more stable to training,and has better detection accuracy compared with other detection methods by combining deformable convolution and attention mechanism.
作者
程健
叶亮
郭一楠
王瑞彬
CHENG Jian;YE Liang;GUO Yinan;WANG Ruibin(Research Institute of Mine Big Data,China Coal Research Institute,Beijing 100013,China;School of Electromechanical and Information Engineering,China University of Mining and Technology(Beijing),Beijing 100083,China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China;State Key Laboratory of Coal Mining and Clean Utilization,Beijing 100013,China)
出处
《煤炭学报》
EI
CAS
CSCD
北大核心
2020年第S02期993-1002,共10页
Journal of China Coal Society
基金
辽宁省自然基金计划资助项目(2020-KF-22-02)
国家自然科学基金资助项目(61973305)
中国煤炭科工集团有限公司科技创新创业资金专项重点资助项目(2019-2-ZD002)
关键词
混合域注意力
变形卷积网络
地裂缝检测
航拍图像
煤矿采空区
hybrid domain attention
deformable convolutional network
detection of ground crack
aerial image
goaf
