We conduct the experiments on an industrial image dataset NEU-CLS which has been extensively studied for evaluating DNN-based algorithms on surface defects

The dataset contains 1,800 grayscale images of size 200x200 pixels, with six types of surface defects: Crazing, Inclusion, Patches, Pitted Surface, Rolled-in Scale, and Scratches. Each type of defect has 300 images, with varying degrees of severity and different orientations.

We use a split of 70-15-15 for training, validation, and testing, respectively. We also apply data augmentation techniques such as rotation, translation, and flipping to increase the diversity of the training data and prevent overfitting.

We train several state-of-the-art DNN architectures, such as ResNet, DenseNet, and Inception, using the Adam optimizer and cross-entropy loss function. We compare their performance in terms of accuracy, precision, recall, and F1-score on the test set.

Our experiments show that these DNN models achieve high accuracy and F1-score, with ResNet-50 achieving the highest performance with an accuracy of 98.61% and an F1-score of 0.986. These results demonstrate the effectiveness of DNN-based approaches for surface defect detection and classification on industrial image datasets.