Omni-Dimensional Dynamic Convolution Meets Bottleneck Transformer: A Novel Improved High Accuracy Forest Fire Smoke Detection Model

The frequent occurrence of forest fires in recent years has not only seriously damaged the forests’ ecological environments but also threatened safety public life and property. Smoke, as main manifestation flame before it is produced, advantage a wide diffusion range that easily obscured. Therefore, timely detection fire smoke with better real-time for early warnings wins valuable time firefighting great significance applications development systems. However, existing methods still have problems, such low accuracy, slow speed, difficulty detecting from small targets. In order to solve aforementioned problems further achieve higher accuracy detection, this paper proposes an improved, new, high-accuracy model, OBDS. Firstly, address problem insufficient extraction effective features complex environments, introduces SimAM attention mechanism, which makes model pay more feature information suppresses interference non-targeted background information. Moreover, Omni-Dimensional Dynamic Convolution instead static convolution adaptively dynamically adjusts weights kernel, enables network extract key different shapes sizes. addition, traditional convolutional neural networks are capable capturing global information, Bottleneck Transformer Net (BoTNet) fully local images while improving target smoke, effectively reducing model’s computation, speed smoke. Finally, decoupling head improve up convergence model. Our experimental results show OBDS proposed significantly than mainstream computational complexity 21.5 GFLOPs (giga floating-point operations per second), improvement 4.31% compared YOLOv5 (YOLO, you look once) [email protected], reaching 92.10%, FPS (frames second) 54, conducive realization warning fires.