Use of Synthetically Generated Training Data to Enhance Graffiti Detection on Passenger Trains

An Approach to Improving Model Robustness Through Targeted Extension of the Feature Space

Abstract

Graffiti on passenger trains results in high cleaning and service disruption costs and poses challenges for automated condition monitoring: the visual variability of graffiti is considerable, real-world training data is often limited, and class distributions are usually unbalanced in practice. This work investigates the extent to which synthetically generated training data can usefully expand the feature space and improve the robustness of a binary graffiti classification (classes: clean vs. graffiti). Four training regimes are compared: (i) baseline (real data only), (ii) reference setup, (iii) hybrid data (real backgrounds with synthetically overlaid graffiti), and (iv) the addition of fully synthetic examples. The models are trained using a standardised pipeline (image size 224 × 224, batch size 32, learning rate 10−3) and evaluated on a separate test split. The decision threshold is determined for each case on the validation split via an F1 sweep and subsequently transferred to the test split. In our experiments, the best configuration achieves an F1 score of up to 0.984 on the test split, with a ROC AUC of up to 1.000, whilst eliminating misclassifications of the positive class (no false negatives in the test split).