Effect of PolyLoss Function on Steel Defect Detection

Steel defect detection is the process of locating and classifying defects in steel surfaces or structures in order to guarantee the quality and integrity of the steel product. Convolutional neural networks (CNN), in particular, have demonstrated remarkable performance in the detection of steel defects. In cases where the data set is not large enough or CNN-based models are too complicated, CNN-based models with transfer learning method can produce very good results together. In this study, we use the transfer learning method to classify defects using three pre-trained CNN models: MobileNet, DenseNet169, and Xception. We train CNN models using the categorical cross-entropy loss function and the polyloss cross-entropy loss function to evaluate the effectiveness of functions. We use the NEU and SEVERSTAL datasets to evaluate models. The results show that validation accuracy improves in models trained with both NEU (0.94% on MobileNet, 1.01% on DenseNet169, and 0.71% on Xception) and SEVERSTAL (0.99% on MobileNet, 1.84% on DenseNet169, and 0.31% on Xception) using PolyLoss cross-entropy loss function