Explainable efficient and optimized feature fusion network for surface defect detection

The quality of the surface and plate form of hot-rolled strip steel, a crucial raw material to produce automobiles, household appliances, and other goods greatly influences the final products that end users make. The identification of surface flaws is crucial to the manufacture of steel strips. Furthermore, typical fault identification techniques have issue of poor detecting reliability, and lower accuracy is obtained by the explainable single pre-trained networks which led to the development of the feature fusion network (FFN). The major objective of the work is to design a traditional deep network model is enhanced by the application of a transfer learning model to detect surface flaws in steel strips. The use of pre-trained models reduces negative effects by drastically reducing training time and improving the accuracy of image classification. Transfer learning models such as VGG16, InceptionV3, and ResNet50 are used to train the Northeastern University-DETection (NEU-DET) Dataset which significantly reduces the time for the training. Generative adversarial network is used for data augmentation to increase the input images. An explainable artificial intelligence (XAI) classifier is applied to the pre-trained networks to understand the classification of the surface defects. A hybrid FFN (HFFN) is proposed which combines the features of pre-trained networks (VGG16, InceptionV3, and ResNet50) to accurately classify flaws in the hot-rolled strips surface. To reduce the features in the HFFN, particle swarm optimization (PSO) algorithm (PFFN) is used. On the NEU-DET, FFN by three-pre-trained model achieves 98.65%, 98.42%, 98.51%, and 98.54% for precision, recall, f-score, and accuracy respectively.