Classifying Surface Fault in Steel Strips Using a Customized NasNet-Mobile CNN and Small Dataset

Abstract

Steel metal is an important product in ferrous manufacturing, and the manufacturing process has to be improved so that hot-rolled strip flaws may be correctly identified. Machine-learning- based automated visual inspection (AVI) systems have been created, however they lack crucial components, such as inadequate RAM, resulting in complexity and sluggish implementation. Long execution times also result in delays or incompleteness. A scarcity of faulty samples further complicates steel defect diagnosis due to the disparity between non-defective and defective pictures. To overcome these difficulties, a deep CNN model is built using the pre- trained NasNet-Mobile backbone architecture. The model, which uses 26 times less data than other papers' datasets, recognizes steel surface pictures with six faults with 99.30% accuracy, outperforming previous methods. This study is beneficial for surface fault classification when the sample size is small, the software is less effective, or time is limited. Avoiding these issues will improve safety and end product quality in the steel industry, saving time and money