基于交叉可变特征融合和动态稀疏注意力YOLOv8的遥感森林野火检测模型

Remote Sensing Forest Wildfire Detection Model Based on Cross Variable Feature Fusion and Dynamic Sparse Attention YOLOv8

摘要: 为了解决传统火焰烟雾检测算法在森林树木遮挡与雨雾天气因素影响下存在漏检误检、准确性下降、小目标检测效果不佳的缺陷，提出了基于交叉可变特征融合和动态稀疏注意力YOLOv8的遥感森林野火检测模型。首先，针对火焰烟雾目标特征复杂的问题，在C2f模块中融合可变形卷积网络(DCNv3)实现特征融合，提升对网络图像中不同尺度火焰烟雾空间位置变化的细节感知能力，增强了网络在不同尺度下的特征表示能力。然后，在主干检测网络加入BiFormer注意力模块，达到抑制干扰信息，提升模型表征能力的效果。最后，引入小目标检测层，进一步提高了检测精度。改进后的算法相比于传统算法，mAP50值提高了1.3%，P值提高了1.5%，R值提高了0.4%。

Abstract: In order to solve the shortcomings of the traditional flame and smoke detection algorithm under the influence of forest tree occlusion and rain and fog weather factors, such as missing detection, false detection, reduced accuracy and poor detection effect of small targets, a remote sensing wildfire detection model based on cross-variable feature fusion and dynamic sparse attention YOLOv8 is proposed. Firstly, in order to solve the problem of complex features of flame smoke targets, the C2f module is fused with a Deformable Convolution Network v3 (DCNv3) to achieve feature fusion, which improves the detail perception ability of the spatial position changes of flame smoke at different scales in the network image, and enhances the feature representation ability of the network at different scales. Then, the attention module of BiFormer was added to the backbone detection network to suppress the interference information and improve the model representation ability. Finally, small object detection layer is introduced to further improve the detection accuracy. Compared with the traditional algorithm, the mAP50 value is increased by 1.3%, the P value is increased by 1.5%, and the R value is increased by 0.4%.