Detection Technology Of Pipeline Internal Surface Defects Based On Machine Vision

With the acceleration of urbanization,the land resources on the urban ground are relatively tight.Based on this,the central and local governments attach great importance to the construction of integrated underground pipeline corridors.However,with the passage of time,various types of surface defects will appear in the underground pipelines everywhere.The surface defects of the pipeline will destroy the basic functions of the pipeline and cause huge economic losses to the country and individuals.Therefore,the detection of these defects is extremely Necessary.The specific research contents of this topic are as follows:1.Designed and completed a smart car based on the Raspberry Pi.The car can complete the operation of intelligent driving and taking pictures in the pipeline.At the same time,in order to realize human-computer interaction,the host computer interface has been written to remotely control the car to perform operations.2.Due to the less data of the pipeline surface defect image training set,GAN and DCGAN networks are used to generate a small number of new pipeline surface defect images to widen the training data set,and median filtering is used to eliminate image noise through experimental comparison.3.The fast-rcnn algorithm is improved in the surface defect detection of the first type of pipeline(high detection accuracy requirements).Through the selection and comparison of the underlying network,the improved Googlenet network is selected as the underlying network of the algorithm.The detection accuracy of small defects is replaced by ROIPooling with ROIAglin.Finally,in the experimental comparison,the hyperparameters in the algorithm are adjusted to improve the recognition rate of defect detection.4.In the second type of pipeline(low detection accuracy requirements)surface defect detection,in order to speed up the detection speed,an improved corner-net network is used,which is improved by introducing the center point prediction idea and center pooling and cascade corner pooling methods.Detection recognition rate.