Study On Identification Of Corrosion Extent For Steel Plates Based On Convolution Neural Network

Corrosion,as the main damage of steel structure bridges,not only affects the appearance of the bridge but also harms the safety of the traffic.If the anti-corrosion measures are not done in time,it will cause a change in the geometric characteristics of the steel members,thus reducing the mechanical properties of the members.For the detection of corrosion,the visual method requires the inspectors to visually inspect the appearance of key components,which is dangerous and inefficient in the working environment.Moreover,the results are subjective and difficult to quantify.In recent years,with the development of computer vision,image recognition methods based on convolutional neural networks have been widely applied.To quantitatively evaluate the corrosion of corroded steel components,this study adopted the convolution neural network(CNN)model for classifying the corrosion extent of steel plates.First of all,accelerated corrosion tests and outdoor exposure tests were carried out.The steel plates were placed in the salt spray test chamber and on the actual bridge in Fukui Prefecture,Japan,to make it corroded.Secondly,the high-resolution images of corroded steel plates were obtained by using the camera,and they were divided into small pieces labeled with corrosion degrees.In this way,the dataset of corrosion was established.Then,based on CNN,the VGG-GAP model for evaluating the corrosion extent was built.The model was then trained to obtain the ability to distinguish the corrosion grades of the small pieces.Next,the sliding window algorithm and weighted average method were used to scan the whole steel plates,and the ranges of corrosion thinning value of the whole steel plates were evaluated.Finally,the accuracy of this method was tested on the newly recycled steel plate corrosion images collected from bridges in Fukui Prefecture,Japan.The test results of the VGG-GAP model on 17 corroded steel plate images in Fukui prefecture indicated that among the 17 samples,ranges of corrosion thinning of 10 samples were correctly evaluated by the proposed model VGG-GAP.VGG-GAP model can learn to extract the features of the non-corroded steel and corroded steel.All the thinning values of new samples and completely corroded samples have been correctly evaluated.For the rest of the corrosion samples,the predicted corrosion thinning values were relatively larger,but there was only a slight deviation,which is conducive to the maintenance workers to develop more safe and reliable anti-corrosion measures.Compared with the traditional corrosion removal methods,it can be concluded that the CNN model VGG-GAP can effectively and accurately evaluate the ranges of corrosion thinning of the steel plates from digital images.