Sparse deformable convolution and high-resolution fusion for contact wire bolt defect detection
The prolonged operation of the train induces vibrations that can result in detrimental conditions,such as loosening and detachment of the catenary bolts,thereby significantly compromising traffic safety.Aiming at the problems that the catenary bolt disease detection of high-speed railway was easy to be interfered with by complex background and the bolt loosening disease was difficult to detect,a method of catenary bolt disease detection based on sparse deformable convolution and high-resolution fusion was proposed.Firstly,a feature extraction network composed of sparse dynamic deformable convolution was constructed to capture the shape features of bolts at different scales by increasing the range of receptive field,so as to strengthen the model's ability to extract the features of small-size objects of bolts.Then,a multi-scale feature pyramid was designed for high-resolution feature fusion module,which fully fused the high-resolution feature maps of the deep features and shallow features of the bolt to improve the utilization of the multi-scale feature map.Secondly,a bolt loosening discrimination method based on connected component statistics was proposed,and the bolt loosening disease state detection was completed by counting the number of connected components of the truncated bolt.Finally,from the high-speed railway contact net bolt detection test,it was concluded.Compared with the Mask R-CNN detection method before improvement,the mean average precision of the proposed method is increased by 41.4 percentage points.The average recall is increased by 27.3 percentage points.The pixel accuracy is increased by 28.11 percentage points.The F1-score is 83.4%.At the same time,the detection efficiency of the catenary bolt network model is tested,and the floating-point calculation efficiency of the model is improved by 36.23%compared with that of Mask R-CNN.The comparative tests of catenation bolt detection in different scenarios show that the proposed method has good adaptability and accuracy.The results can provide a more accurate method for the detection of nut loosening and missing disease,and have certain reference significance for the later intelligent detection of catenation.
国家科技期刊平台
NSTL
万方数据
