A high-performance convolutional neural network for steel defects detection

Ensuring defect-free products is a critical consideration in the hot-rolled steel strips and steel terminal manufacturing industry and cannot be over-emphasized, which underpins the purpose of the study presented herein which is focused on developing an algorithm for the detection of defects for two (2) different classes of steel products. The study is divided into two (2) parts, part one being based on transfer learning is aimed at detecting defects on hot rolled steel strip surfaces, while part two aims to achieve defect detection on steel terminals using an optimized YOLO v3 algorithm. Problem statement one of the studies proposes a transfer learning-based method for detecting defects on hot-rolled steel strip surfaces. The method proposed in this study utilizes Deep Convolutional Neural Network (DCNN) using VGG-19 as a pre-trained model, with data augmentation to mitigate the effect of limited training data and also data-class imbalance. Experimental tests indicate that the method proposed herein using the VGG-19 pre-trained model has an excellent performance, with training and validation accuracies of 93.11% and 97.22%, respectively, which is more accurate than the KNN, ANN, and Faster RCNN model. Problem statement two of the study proposes an optimized YOLO v3 algorithm to detect three (3) most common classes of defects on steel terminals. The optimized YOLO v3 network proposed in the thesis was trained using steel terminal image datasets acquired in real-time from Hatko Electronics' production factory in Istanbul, and delivered a good performance in the region of 97.19% AP50, at 0.50 IOU Threshold, which is a higher than the mAP (mean Average Precision) results (57.9% AP50 @0.50) from the original YOLO v3 network, and higher than the mAP score of Faster-RCNN that is about 55.7% AP50, and, 50.4% AP50 for SSD513 (Single-shot Detector) network.