Research On Defect Recognition And Location Of Gondola Car Body Based On Deep Learning

The gondola car body is affected by factors such as corrosive environment and sharp cargo impact,which will produce different degrees of damage and defects,thereby reducing the transportation safety and the service life of the gondola car.Therefore,it is necessary to check the state of open car body and repair defects in the process of partial repair and factory repair.At present,in the repair work of gondola car body defects,it mainly relies on manual completion of defect inspection around the car and subsequent repair tasks,which has problems such as low degree of automation,high labor intensity,and low efficiency.Therefore,in this paper,the research on the identification and localization of gondola car body defects based on deep learning is carried out,in order to realize the rapid identification and localization of gondola car body defects for subsequent cutting and repairing.The main research contents are as follows:（1）A set of datasets suitable for gondola car body defect recognition and localization algorithms is constructed.Firstly,the defect images of the gondola car body are collected in the maintenance workshop to determine the labeling category of the dataset.Then,in view of the image quality problem caused by the environment noise in the maintenance workshop,the image is denoised and preprocessed.In order to avoid overfitting and reduce the unbalanced proportion of samples,image processing methods such as rotation,flipping and brightness adjustment are used for data amplification,and a sufficient and balanced image data are obtained.Finally,the data is marked and divided to obtain standard gondola car body defect dataset.（2）Three more advanced target detection models,Faster R-CNN,YOLO and SSD,were built to study the detection effects of different models on the gondola car body defect dataset.The detection speed,average precision and mean average precision of defects were compared through experiments.Experiments show that YOLOv5 has the best performance on the gondola car body defect dataset,but the average precision of scratches,deformations,and holes still has a lot of room for improvement.（3）A gondola car body defects recognition algorithm based on optimized YOLOv5 is proposed.Firstly,the CA attention module is embedded in the Backbone network of original YOLOv5 to improve the feature extraction ability? Then reconstruct the detection head of the original YOLOv5 network,and use feature maps of more sizes for result prediction? Finally,based on the consideration of detection speed and practical application scenarios,a pruning method for the lightweight model was proposed.Experiments show that the detection effect of the optimized YOLOv5 is better than that of the original YOLOv5.The m AP value is increased from 87.47% of the original YOLOv5 to 91.11%,an increase of 3.64%,and the detection speed reaches 10.17 ms/frame,which can achieve detection accuracy and speed.At the same time,the model occupies less memory and is easier to embed into small devices.（4）A method for locating defects of gondola car body in maintenance workshop based on monocular vision is proposed.Build an gondola car body defect identification and positioning environment based on the optimized YOLOv5 algorithm and monocular vision technology,and use the camera calibration experiment to determine the camera parameters.The spatial location of the defect.Experiments show that the error of defect location is within 6.4 mm,and the single-frame processing time of the defect identification and location algorithm is 38.53 ms,which meets the requirements of engineering accuracy and real-time performance,and lays the foundation for further research on the gondola car body the vision system of the gondola car body defect inspection robot.