Hybrid CNN–Transformer–GRU framework with improved residual shrinkage network and DCCS optimisation for steel surface defect classification

Automated inspection of steel surface defects is an important task in intelligent manufacturing systems to ensure product quality and reduce manual inspection efforts. This study proposes a hybrid deep learning framework for accurate classification of metal surface defects by integrating convolutional feature extraction, residual shrinkage learning, transformer-based attention modelling, and recurrent sequence analysis. Initially, greyscale defect images from the NEU metal surface defect dataset are processed using a convolutional neural network (CNN) combined with Batch Normalisation and Improved Residual Shrinkage Network (IRSN) blocks to extract noise-resilient and discriminative local features. The IRSN mechanism performs adaptive soft thresholding to suppress noise and enhance relevant defect patterns. Transformer encoder with multi-head attention learns long-range relationships among extracted features. In parallel, a Gated Recurrent Unit (GRU) branch models sequential dependencies in the feature representation. The outputs from both branches are fused and passed through fully connected layers for multi-class defect classification. A Dual-Condition Cost-Sensitive (DCCS) optimisation strategy is applied to automatically tune hyperparameters. The proposed model attains an accuracy of 93.06% and 98.40% on the NEU metal surface defect dataset and Severstal Steel Defect Detection dataset, respectively. The proposed model underscores an improvement of around 0.06% and 1.2% over the recent benchmark model.