The detection of defects on material surface occuring after production with machine learning techniques

The effective quality control of materials in production process is so significant in terms of being cost-effective, labour-saving and time-saving. Steel used widely in production industry is one of the most important building materials of modern times. Therefore, the efficient detection of steel surface defects in production process is very important. Artificial intelligence-based applications will help to make steel production more efficient. In this study, a new deep learning-based approach has been developed that detects and classifies surface defects that occur in the steel production process. The first proposed methodology was created in four steps. In the first step, a deep learning model is designed that trains the residual and attention structures in parallel, thus increasing the classification performance. In the second step, deep features were extracted from the Parallel Attention-Residual Convolutional Neural Networks (PAR-CNN) model. The extracted features in the third step were selected by a new and simple algorithm based on matching the indexes obtained from the Neighbourhood Component Analysis (NCA) and Relief algorithms (NRMI). In the last process, classification was done with the Support Vector Machine (SVM) algorithm. The proposed methodology was used for dual and multi-class classification tasks and evaluated on two dataset in the Kaggle database named Severstal: Steel Defect Detection and NEU steel surface dataset. In the first classification task, defect-free and faulty steel surface images were classified and 97.90% classification accuracy was achieved. In the second classification task, the classification consisting of five classes was studied and 94.50% the classification accuracy was obtained. The second proposed approach consists of five stages. All stages were used to improve the classification performance. Spectogram images were used and the data were trained with a newly designed model called PAR-CNN model. An efficient and fast feature selection algorithm called Iterated Neighbourhood Component Analysis (INCA) was applied to reduce the computational cost and improve the classification performance. The training-validation accuracy scores for binary classification improved to 100% and 95.38%, respectively. For multiclass classification, the training-validation accuracy scores reached 99.21% and 91.89%, respectively.