Detection and prediction of defects in steel sheets based on Haar-ResUNet hybrid neural network: a deep learning approach

Production of steel as an essential raw material in all kinds of industries is an important challenge, both qualitatively and quantitatively. The quality of the product affects the final performance. Cracks and cuts on the surface, thinning and non-uniform injection can be the most important factor in lowering the stress tolerance in a steel sheet. These abrasions occur when the plastic deformation accumulated in the material during its construction exceeds the limit. It's shaping goes beyond and leads to breaking the thickness of the part. Although classical solutions have been able to detect defects on steel surfaces, the best option is solutions based on deep learning. Convolutional and recurrent neural networks can accurately detect areas of abrasion or corrosion and wounds on the surface of steel; But in real-time diagnostics, the accuracy of the diagnosis decreases. With the aim of increasing the detection accuracy and reducing the calculation time, this study proposes a completely new Haar-ResUNet hybrid solution, which reduces the number of calculations by using the feature extraction pattern and quick and definitive removal of non-functional areas in a sample image. The findings show that the surface of the steel sheet is not like the human face in an image, where an almost accurate estimate of the distance between the facial components can be seen on it, and the usual errors observed on the surface of the steel sheet can be spread and scattered. This causes areas of the image to be removed that may have one or two surface scratches and lowers the accuracy of the algorithm. Also, rapid removal of improbable regions from the image changes the average density and doubles the calculations. Because this step must be done twice. But these two problems have been solved by providing suggestions. Finally, the accuracy of Res-UNet algorithm alone and in previous models is reported to be ۹۳%. This model is recursive and is a good choice for classification, but by adding Viola-Jones input layers to this algorithm, the accuracy of the final Haar-ResUNet model reaches approximately ۹۷%.

کلیدواژه ها:

Steel Sheet ، Defect Detection ، Image Processing ، Res-UNet ، Haar