Intelligent representation method of image flatness for cold rolled strip

Real flatness images are the bases for flatness detection based on machine vision of cold rolled strip.The characteristics of a real flatness image are analyzed,and a lightweight strip location detection(SLD)model with deep semantic segmentation networks is established.The interference areas in the real flatness image can be eliminated by the SLD model,and valid information can be retained.On this basis,the concept of image flatness is proposed for the first time.An image flatness representation(IFAR)model is established on the basis of an autoencoder with a new structure.The optimal structure of the bottleneck layer is 16×16×4,and the IFAR model exhibits a good representation effect.Moreover,interpretability analysis of the representation factors is carried out,and the difference and physical meaning of the representation factors for image flatness with different categories are analyzed.Image flatness with new defect morphologies(bilateral quarter waves and large middle waves)that are not present in the original dataset are generated by modifying the representation factors of the no wave image.Lastly,the SLD and IFAR models are used to detect and represent all the real flatness images on the test set.The average processing time for a single image is 11.42 ms,which is suitable for industrial applications.The research results provide effective methods and ideas for intelligent flatness detection technology based on machine vision.

Data-driven flatness intelligent representation method of cold rolled strip

A high-accuracy flatness prediction model is the basis for realizing flatness control.Real flatness is typically reflected as the strain distribution,which is a vector.However,it is difficult to obtain ideal results if the real flatness is directly used as the output value of the flatness intelligent prediction model.Thus,it is necessary to seek an abstract representation method of real flatness.For this reason,two new intelligent flatness representation models were proposed based on the autoencoder of unsupervised learning theory:the flatness autoencoder representation(FAR)model and the flatness stacked sparse autoencoder representation(FSSAR)model.Compared with the traditional Legendre fourth-order polynomial representation model,the representation accuracies of the FAR and FSSAR models are significantly improved,better representing the flatness defects,like the double tight edge.The optimal number of bottleneck layer neurons in the FAR and FSSAR models is 5,which means that five basic patterns can accurately represent real flatness.Compared with the FAR model,the FSSAR model has higher representation accuracy,although the flatness basic pattern is more abstract,and the physical meaning is not clear enough.Furthermore,the accuracy of the FAR model is slightly lower than that of the FSSAR model.However,it can automatically learn the flatness basic pattern with a very clear physical meaning for both the theoretical and real flatness,which is an optimal intelligent representation method for flatness.