一种改进YOLOv5的多尺度像素林火识别算法

Multi-scale forest fire recognition using improved YOLOv5 algorithm

林火破坏程度大、蔓延速度快的特点给森林生态环境和人类带来极大的危害。深度学习技术可以学习和自适应提取林火特征,捕获的林火图像中火焰的像素尺寸不同,林火提取的特征也不同。为了能够识别复杂背景下不同像素尺度的火灾,笔者提出了一种改进的YOLOv5林火识别方法,通过在YOLOv5的检测网络加入解耦头,解决林火图像输出变量时分类和回归的冲突问题,加快网络收敛速度提高识别精度;在网络中引进CBAM注意力机制,更关注林火信息同时提升识别精度;在Neck网络引入加权双向特征金字塔网络(BiFPN),替换原有的路径聚合网络(PANet),对不同维度的林火特征进行融合,进行特征筛选,增强特征表示能力。实验结果表明,该林火识别算法在自制的林火数据集上进行训练和验证模型,检测性能上均优于YOLOv5算法,在准确率、召回率、平均精度分别提升了5.2%,3.0%,3.4%,mAP@.5:.95提升了4.6%,并且在不同尺寸林火目标的识别精度上均有提升。研究结果对林火识别性能提升有着积极意义。

The characteristics of forest fires with high destruction and spreading speed pose a serious threat to forest ecology and human beings. All traditional methods of manual inspection, sensor-based detection, satellite remote sensing and computer vision detection have their obvious limitations. Deep learning techniques can learn and adaptively extract the features of forest fires, however, the different flame sizes in captured forest fire images cannot learn effective information for recognition. In order to recognize fires of different scales in complex backgrounds, this paper proposes an improved YOLOv5 forest fire recognition method, which solves the conflict between the classification and regression in the output variables of forest fire images by adding Decoupled Head to the detection network of YOLOv5. The addition of the decoupling header not only brings the improvement to the model recognition accuracy, but also significantly accelerates the model convergence speed. For the case of fire target with low pixels and easy information loss, in order to make the model better focus on the detection of local information and improve the accuracy of extracting image features. The CBAM attention mechanism was introduced into the network. After adding CBAM attention mechanism as the network level deepens, the features between different levels were lost. Therefore, a weighted bi-directional feature pyramid network(BiFPN) was introduced in Neck network to replace the original path aggregation network(PANet) to fuse different dimensional forest fire features for feature screening and enhance feature representation to prevent the feature loss. The experimental results show that the forest fire recognition algorithm outperforms the YOLOv5 algorithm in both detection performance when the model is trained and validated on a lab-made forest fire dataset, with improvements of 5.2%, 3.0%, and 3.4% in accuracy, recall, and average precision, respectively. Furthermore, mAP@.5:.95 is increased by 4.6%, and the recognition accuracy of forest fire targets in different sizes is enhanced. The research results have positive implications for forest fire recognition performance improvement and possess promising application prospects.