The surface quality of heavy rail directly affects the service life of heavy rail and plays a critical role in the safety of railway transportation.Surface defect inspection for cold heavy rail is an important part of quality control in heavy rail production,and its related research has attracted a lot of attention.In recent years,some researchers have proposed several methods based on computer vision to address problems with surface defect inspection for cold heavy rail.A lot of traditional image processing technics were adopted in their work,like object extraction,edge detection,etc.However,these methods cannot work well with the influence of environmental factors such as vibration,dust,lightness,which all usually appear during the inspection.Therefore,the manual visual inspection method is used for surface defect inspection in almost all workshops nowadays.Herein,we proposed a kind of system for cold state heavy rail surface defect inspection based on deep learning methods to improve inspection efficiency and ensure the quality of heavy rail.The computer vision module was designed for image collection of rail surface in this system.Afterward,the images will be passed to system control module for real-time online defect inspection.The defect detection algorithm based on Faster-RCNN was implemented and improved with the Tensorflow framework in system control module.Experiments showed that the system proposed in this paper took about 15 ms for a single image test,and both the detection accuracy and recall rate reached 90% according to the test on the data with two types of flaw,which are scar and dent.The system meets the requirements of real-time surface defect detection for cold heavy rail.The defect detection system proposed in this paper achieved the real-time on-line detection of cold heavy rail surface defects,which greatly improved production efficiency and saved production costs.In addition,the ideas of the detection system design and the method described for deep model optimization in this paper also have certain reference significance for the development of deep learning-based visual inspection systems in other fields. |