Molten mark classification using multi-scale directional cross-scale attention fusion

In electrical fire investigation, it is critical to determine whether molten marks on wires were caused by a short circuit or by external heat. Existing analytical methods require substantial manual work and expensive equipment for preprocessing and classification. Recent convolutional neural network (CNN)-based approaches have demonstrated relatively good performance; however, their accuracy is limited by the insufficient utilization of various features of molten marks. To address this, we propose a molten mark classification method based on multi-scale directional cross-scale attention fusion. We first extract feature maps at four resolutions (scales) using a Swin Transformer. These features are then fused using a bi-directional feature pyramid network neck integrated with directional cross-scale attention modules for channel and spatial attention, conditioned on features from different scales. The fused features are passed to a classification head for final classification. For performance evaluation, we conducted extensive experiments with seven comparison models. The proposed method achieved an F1 score of 0.9644 and a Matthews correlation coefficient of 0.9257, improvements of 1.78 and 3.86 percentage points over a state-of-the-art model. Additionally, the proposed method exhibits only a 2.88% accuracy decrease on degraded images, demonstrating greater robustness than CNN-based models, which decrease by 11.31% on average.