Research On Steel Plate Rust Category Identification Method Based On Deep Learning Transformer Network

As the main type of steel structure damage,rust not only has an impact on its appearance image,but also leads to the loss of structural component section and the reduction of material strength,resulting in a decrease in the cold brittleness resistance of steel,and a sudden brittle fracture occurs without obvious signs of deformation,especially under the action of impact load,which is more dangerous,affecting the safety of steel structure,and seriously affecting its durability,so the detection of steel structure rust state is particularly important.The traditional steel structure rust detection is mainly visual inspection,although it has the advantages of low threshold,easy operation and low cost,but there are low efficiency,high cost,long detection cycle,easy to miss inspection,not suitable for large-range,large-area and high-altitude personnel difficult to reach the part of the steel structure batch detection and other problems.In recent years,with the vigorous development of emerging industries such as information science and artificial intelligence,machine vision technology has been applied to image processing and non-destructive testing on a large scale.Therefore,this project plans to propose a steel plate rust recognition and classification method based on Transformer network from the image perspective and with the help of Transformer network,which is widely used in the field of image processing.The corrosion status of steel plate is detected and classified by major transformation,replacing manpower to complete the corresponding detection and identification work.By comparing the advantages,disadvantages and application scenarios of common backbone network models of Transformer networks,the Cross Former network model(Cross Former: Multifunctional Visual Transformer Network Model Based on Cross-scale Attention)is selected as the steel plate rust recognition classification model.In this paper,the steel plate corrosion type database and the steel plate uniform corrosion depth database are constructed.Firstly,through the original data acquisition and data amplification,screening and enhancement of five typical steel plate corrosion types in industrial environment,a database of 60,000 different types of steel plate corrosion types was constructed.At the same time,through the indoor neutral salt spray test,the photo data of uniform rust apparent morphology of Q345 and Q235 steel plates of different thicknesses were obtained,and the XRD composition analysis of steel plate corrosion products,the corrosion thickness of steel plate and the apparent morphology characteristics of steel plate were analyzed,and the correlation between uniform corrosion depth and apparent morphology characteristics of steel plate was obtained,and the uniform corrosion apparent morphology of steel plate with four corrosion depth gradients was divided.By image preprocessing,a database of 60,000 uniform rust original images of steel plates with four rust depth gradients was constructed.In order to realize the rust recognition and classification of steel plates,a steel plate rust type recognition classification model and a steel plate uniform rust depth recognition classification model based on deep learning(DL,Deep Learning)Cross Former network are proposed,and the Cross Former neural network model is constructed based on Python,and the steel plate rust type database and the steel plate uniform rust depth database are trained and tested respectively,and the hyper parameters of the optimized network model are adjusted to obtain the optimal model.Through the tests:(1)the steel plate rust type recognition classification model based on Cross Former network can achieve 96.86% of the steel plate rust recognition classification accuracy,and compared with the Swin-T network model,the recognition accuracy of Cross Former neural network model is significantly higher than that of the Swin-T network model,which can effectively identify different rust types of steel plates;(2)The accuracy of steel plate uniform rust depth recognition classification model based on Cross Former network can reach 71.79%,which is much higher than 42% of Swin-T network model.In order to verify the practical application of steel plate corrosion type identification classification model and steel plate uniform rust depth recognition classification model,this paper collects the image data of steel plate rust type in the hot stewing workshop of Xiang gang Steel Mill and the uniform corrosion image data of steel plate with four kinds of corrosion depth gradient for model test verification,and the results show that:(1)The verification accuracy of steel plate corrosion type identification classification model is 96.80%,which can effectively identify steel plate rust type and has a high practical application type.(2)The verification accuracy of the uniform rust depth recognition classification model of steel plate is 40%,which is much less than the model test accuracy of 71.79%,which proves that the practical application of the uniform rust depth recognition classification model of steel plate is not high,and the analysis reasons are mainly divided into the structural problems of the model itself,the picture data of the uniform corrosion depth database of steel plate established by salt spray experiment is too idealistic,and the apparent characteristics of steel plate rust in actual engineering are affected by the environment.