HD-SDNet: A Forest Fire Smoke Detection Model Based on Hierarchical Deformable Transformer

Forest fires pose a serious threat to human life and property, making accurate and rapid detection of forest fire smoke crucial for fire prevention. Traditional convolutional neural network (CNN)-based smoke detection methods rely heavily on manually designed components and exhibit limited capability in detecting small or indistinct smoke in complex forest environments. To improve detection accuracy and efficiency, this study proposes an early forest fire smoke detection model based on a hierarchical deformable Transformer (HD-SDNet). The model integrates a Dilated Context Fusion and Contrastive (DCFC) module and a Hierarchical Deformable Transformer Module (HD-DETR) to enhance the perception of small or faint smoke features. Furthermore, to improve detection precision and reduce false positives and missed detections, we introduce an Iterative Bounding Box Optimization (IBBO) method that generates bounding boxes that tightly enclose smoke targets. The proposed method is validated using a custom dataset containing forest fire smoke images at various scales. Extensive experimental results demonstrate that the improved model achieves significantly higher accuracy than existing state-of-the-art methods. The proposed HD-SDNet is well-suited for early forest fire smoke detection, exhibiting high detection accuracy for smoke targets across different scales.