Research On Intelligent Defects Detection Of Typical Components Of High-Speed Railway From UAV Images

At present,the inspection of equipment along the high-speed railway mainly depend on manual inspection and inspection vehicle.The existing statistical data and inspection data show that the rail defects and catenary support devise(CSD)are two typical types of equipment,which have a significant impact on railway operation safety.However,the existing two kinds of inspection methods generally have some problems,such as poor night inspection conditions,low inspection frequency,restricted by comprehensive maintenance period,narrow inspection area and so on.Besides,the existing defect detection system still has some problems,such as poor image understanding ability,poor automatic analysis ability,defect recognition relies heavily on human assistance and so on.Therefore,in order to overcome the above problems,this paper proposes UAV image intelligent defect detection method for rail and CSD,and solves the haze problem often encountered in UAV outdoor inspection,so as to improve inspection means,enhance inspection efficiency,and improve the intelligent level of existing detection system.The main work of this paper is as follows:(1)This paper presents a novel end-to-end network for UAV-based railway images dehazing,and focuses on two key issues: network architecture and loss function.On the one hand,based on a pyramidal network structure,densely pyramidal residual network consists of dense residual block and enhanced residual blocks,which heavily exploits the feature maps of all preceding layers and considerably increased depth at different scale,respectively.On the other hand,a new loss function introducing structural similarity index is proposed to preserve more structural information,thereby restore the appealing perceptual quality of the hazy images.Finally,the superiority of the method can be proved by extensive experimental comparison.(2)This paper proposes a new rail boundary guidance network(RBGNet)for rail surface(RS)of high saliency and aspect ratio segmentation.Firstly,a novel architecture is proposed to fully utilize the complementarity between the RS and the rail edge,which injects high-level RS object information into shallow rail edge features by a progressive fused way for obtaining fine edge features.And then integrates the refined edge features to RS features at different high-level layers to predict the RS precisely.Secondly,an innovative hybrid loss consisting of Binary Cross Entropy,Structural Similarity,and Intersection-Over-Union is proposed and equipped into the RBGNet to supervise the network to learn the transformation between the input and ground truth,which further refine the RS location and edges.Finally,the superiority of the method is verified by comparing the experimental results on UAV images.(3)This paper presents an inspection approach for non-depth and shapeless RS defect,and focuses on two main issues: image improvement and defects detection: First,Local Weber-like Contrast is used to improve rail images,which homogenized backgrounds and highlighted RS defects under various sunlight intensity,because of its local nonlinear,illuminance independent.Second,Gray Stretch Maximum Entropy is utilized to stretch gray range and de-noising on rail images,and generates an optimized segmentation threshold to detect defects.Finally,the effectiveness of the proposed method is proved by comparing with the classical methods and evaluating the whole system.(4)Focusing on small sample,high redundancy CSD joint detection,a cascaded YOLO network based on cascaded multi-attention mechanism is proposed,which integrates the shallow edge features and deep semantic information,and establishes two yolo layers according to the component proportion,thereby improving the detection precision.In addition,the gridmask data segmentation method is utilized to eliminate the over-fitting for less volume and excessive similarity in dataset.Finally,the superiority of the proposed method is fully verified on the inspection vehicle and UAV image dataset.(5)Aiming at arbitrary direction,small object defect detection for the fasteners on CSDs(nut missing,split pin damage,etc.),the paper proposes a rotation Retina Net of which structure and anchor mechanism generate suitable candidate anchors for fasteners,avoiding unnecessary computational overhead.Moreover,the loss function introduces RIo U loss to eliminate inaccurate angle prediction of predicted boxes.Finally,the superiority of the proposed method is fully verified on the inspection vehicle and the UAV image dataset.(6)Aiming at the requirements of daily safety management of different typical equipment along the high-speed railway and the safety guarantee of high-speed railway operation environment,a UAV image inspection system is proposed,including UAV inspection scheme,load selection,route planning and inspection security.Finally,the paper presents UAV inspection system platform,and the actual cases.The actual case proves the effectiveness of the defect detection method proposed in this paper for different modes of typical equipment along the railway.