A ResNet-50 based method for aircraft skin surface defect classification integrating class balancing and data augmentation

Purpose This study aims to address the critical challenges of defect sample scarcity and extreme long-tail distributions in automated aircraft skin inspection, which frequently lead to conventional deep learning models missing critical minority defects. Design/methodology/approach An application-oriented classification pipeline was constructed based on a ResNet-50 backbone pre-trained on the standard ImageNet data set. Class-balanced sampling and weighted cross-entropy loss were used to reduce prediction bias towards majority classes, whilst geometric and colour enhancement strategies were used to increase the diversity of training samples under limited data conditions. Dropout regularisation was incorporated into the task-specific classification head, followed by the use of gradient-weighted class activation mapping to visualise the image regions contributing to the classification decision. The combined effect of these complementary strategies was evaluated through comparison, ablation, visualisation and sensitivity tests under simulated image-noise perturbations. Findings Experimental results on a proprietary aircraft skin data set demonstrate that the model achieves 96.60% accuracy and 92.12% multiclass balanced accuracy. These results indicate improved classification performance under the current proprietary patch-level data set and experimental protocol; however, they do not establish generalisation to independent aircraft fleets, data acquisition systems or full-image inspection scenarios. Originality/value The contribution of this study does not lie in proposing a new backbone architecture or learning algorithm, but in the application-specific configuration and systematic evaluation of complementary class-balancing, data-augmentation, regularisation and interpretability strategies for limited and imbalanced aircraft skin defect data. Through controlled ablation, repeated-run evaluation and minority-class-oriented performance analysis, the study provides empirical evidence on how these established techniques jointly influence classification performance under the present aircraft skin inspection setting.