Аннотация
Fire and smoke constitute critical hazards in industrial and urban environments, necessitating reliable early detection systems capable of real-time operation. Although deep learning-based object detection methods have demonstrated considerable potential, existing fire and smoke detection studies predominantly apply default training configurations or modify network architectures, while the systematic investigation of fine-tuning optimization, specifically the interaction effects among optimizer selection, layer freezing depth, learning rate policy, and domain-targeted augmentation, remains insufficiently explored for dual-class fire and smoke detection. This paper proposes a structured four-component fine-tuning framework for the YOLOv9c architecture that jointly optimizes: Stochastic Gradient Descent (SGD) with momentum as a generalization-favoring optimizer, selective freezing of the first 10 backbone layers to preserve transferable features, cosine annealing with linear warmup for controlled domain adaptation, and a fire- and smoke-targeted augmentation pipeline calibrated to the visual variability of fire and smoke phenomena. The key contribution lies not in the individual techniques, which are established in the literature, but in the systematic ablation-based quantification of their individual and interactive contributions, an analysis absent from prior work in this domain. Experiments are conducted on 35,000 manually annotated images sourced from public repositories and Kazakhstan industrial surveillance facilities. The fine-tuned model achieves 78.0% precision, 76.0% recall, and 77.0% F1-score at 30.2 frames per second (FPS), outperforming Faster R-CNN, ResNet-based, YOLOv8m, and YOLOv10m baselines. Ablation analysis reveals that SGD with momentum yields the largest individual improvement (+4.2% F1-score), with the cumulative improvement (+6.3%) exceeding the sum of individual contributions, indicating positive interaction effects among the optimization components.