Performance Evaluation of CNN for Industrial Steel Defect Classification

Abstract

Surface inspection is a critical stage of steel manufacturing, and manual visual inspection is slow, subjective, and hard to scale. This study aims to construct a convolutional neural network (CNN) pipeline that classifies the six types of hot-rolled steel surface defects (crazing, inclusion, patches, pitted surface, rolled-in scale, and scratches) based on the public NEU surface-defect dataset (1,800 grayscale images, 300 images of each type). Three baseline networks, a basic CNN, a ResNet-style network and a MobileNet-style network, were first compared under a fixed 70/15/15 split; they reached 74.1%, 58.1% and 74.8% test accuracy, and the deeper network overfitted badly. To close this gap, the study designed a compact network built from batch-normalised depthwise-separable convolutions with global average pooling (66,198 parameters), trained from scratch with stronger augmentation and a cosine learning-rate schedule. Over three random seeds on the official split, it achieved 97.9 ± 0.2% accuracy (macro-F1 0.979), and stratified five-fold cross-validation on all 1,800 images gave 98.7 ± 0.7% (95% confidence interval 97.8–99.6%, macro-F1 0.987). Class-wise precision, recall, and F1, and the pooled confusion matrix show that inclusion and pitted surface remain the most confused pair. Full hyperparameters and the evaluation protocol are reported so that the experiments can be repeated.