Design and optimization of three class object detection modalities for manufacturing steel surface fault diagnosis and dimensionality classification

The main objective of this research is to create and improve three different object identification techniques for identifying surface flaws and categorising dimensions in steel that has been fabricated. RetinaNet, YOLOv3, and Faster R-CNN are the selected modalities in the experiment. The main goal is to evaluate these modalities' ability to detect and classify defects on steel surfaces in terms of accuracy, precision, recall, and F1 score. This assessment makes use of a varied collection of steel surface photos that show different kinds and sizes of faults. Training, validation, and testing sets make up the dataset's partitioning. The training set is used to train and optimise the three modalities, while the testing and validation sets are used to evaluate their performance. According to the study's findings, all three methods provide excellent of 0.92. RetinaNet comes in second with an F1 score of 0.89, followed by YOLOv3 with an F1 score of 0.87, while the Faster R-CNN modality obtains the greatest overall performance with an F1 score.