Study On Application Of Strip Steel Surface Defect Detection Based On Deep Learning

As an indispensable raw material in the development process of modern industry,strip steel has its surface defects greatly affecting the performance and service life of strip steel.Therefore,it is necessary to study a method that can effectively detect the surface defects of the steel strip.The defect detection algorithm is researched to solve the problem that the sample data on the surface of the steel strip is uneven and the supervised detection algorithm depends on a large amount of labeled information.The main work completed includes:(1)For the problem that the supervised detection algorithm relies on a large number of marking information,an unsupervised strip steel surface defect detection method based on image reconstruction is proposed.Firstly,the image reconstruction method based on DCGAN is proposed to achieve complete sample reconstruction.Then the DA-DOCS method is proposed to optimize the parameter combination of OCSVM,and uses the reconstruction error as input to train OCSVM to realize the anomaly detection of the strip steel surface.Experimental results show that compared with AE and AAE,the proposed method has lower reconstruction error and better reconstruction performance.The precision,recall and F1-score of the proposed detection method were 97.3%,98.1% and 97.7% respectively,which was better than these defect detection methods based on AE and AAE.(2)Aiming at the problem of unbalanced sample data of strip steel surface defects,a method for strip steel surface defect detection based on data enhancement is proposed.First of all,the samples-heuristic GAN model is proposed to enhance the defect data,which prompts the discriminator to specify generate the sample distribution of the counter-example,and provides the generator with a beneficial gradient update to realize the enhancement of the defect data.Then,the Dense Net model based on selective kernel and squeeze excitation is proposed to reduce information redundancy and improve detection accuracy.The experimental results on the strip data set show that the proposed method has better enhancement performance than DCGAN,WGAN,LSGAN.The enhanced detection precision,recall and F1-score were increased by 3.2%,1.8% and 2.5% respectively,which provides a new solution for strip steel surface defect detection.