An Efficient Fire Detection Algorithm Based on Mamba Space State Linear Attention

The Mamba model, as a State Space Model (SSM), enhances the global receptive field and feature extraction capabilities of object detection models through an architecture inspired by Recurrent Neural Networks (RNNs). Compared to traditional Convolutional Neural Networks (CNNs) and Transformers, it excels in handling complex scale variations and multi-view interference, offering a novel approach for object detection tasks in dynamic environments such as fire detection. Therefore, this paper proposes an efficient attention mechanism, the Efficient-Mamba-Attention (EMA) module, which enhances input features through adaptive average pooling and SSM modules. By integrating the EMA module with the YOLOv9 architecture, a highly efficient fire detection algorithm is proposed, particularly suitable for complex scenarios. Additionally, the model introduces the ConvNeXtV2 block to strengthen the backbone network, compensating for the limitations of SSM models in local feature modeling. Finally, the introduction of dynamic non-monotonic focusing and distance penalty mechanisms optimizes the loss function, improving the accuracy of the detection bounding boxes and significantly enhancing the precision and robustness of the fire alarm tasks. Comparative experiments demonstrate that the proposed network excels in terms of AP50 and FPS, achieving an AP50 of 91.0% on the large-scale fire dataset (dataset1), and 87.2% on the small-scale fire dataset (dataset2), with FPS reaching 71 on dataset1 and 64 on dataset2. This demonstrates that the proposed method maintains high detection performance while achieving outstanding efficiency.