高级检索
    周浪, 樊坤, 瞿华, 吕媛媛, 张正宜. 基于Sparse-DenseNet模型的森林火灾识别研究[J]. 北京林业大学学报, 2020, 42(10): 36-44. DOI: 10.12171/j.1000-1522.20190371
    引用本文: 周浪, 樊坤, 瞿华, 吕媛媛, 张正宜. 基于Sparse-DenseNet模型的森林火灾识别研究[J]. 北京林业大学学报, 2020, 42(10): 36-44. DOI: 10.12171/j.1000-1522.20190371
    Zhou Lang, Fan Kun, Qu Hua, Lü Yuanyuan, Zhang Zhengyi. Forest fire identification based on Sparse-DenseNet model[J]. Journal of Beijing Forestry University, 2020, 42(10): 36-44. DOI: 10.12171/j.1000-1522.20190371
    Citation: Zhou Lang, Fan Kun, Qu Hua, Lü Yuanyuan, Zhang Zhengyi. Forest fire identification based on Sparse-DenseNet model[J]. Journal of Beijing Forestry University, 2020, 42(10): 36-44. DOI: 10.12171/j.1000-1522.20190371

    基于Sparse-DenseNet模型的森林火灾识别研究

    Forest fire identification based on Sparse-DenseNet model

    • 摘要:
        目的  森林火灾的频繁发生给森林防治工作带来很大的难度,传统的森林火灾识别算法存在准确率低、处理效率不够高等问题,同时由于森林火灾图像数据本身具有很强的复杂性,需要从识别精度和泛化能力等多方面进行综合考虑,因此本文将利用稀疏化的DenseNet模型展开森林火灾的识别研究。
        方法  首先,对DenseNet模型进行稀疏化改造,通过随机屏蔽Dense Block模块中节点的方式来产生稀疏化效果,使得算法具备减轻过拟合、缓解梯度消失以及加快收敛速度等优点。其次,在林区进行图像采集时,由于摄像设备与被采集物体之间的相对运动以及光影作用,会出现图片数据被干扰的情况,因此本文利用python相关的图片处理工具对图片进行变换,从而对图片数据集进行相应的扩充,使其能够契合实际的应用场景。最后,本文将Sparse-DenseNet模型与其他经典深度学习模型在森林火灾数据集以及cifar10数据集上的表现进行对比,观察其效果。
        结果  Sparse-DenseNet模型拥有在结构上更加轻量的特点,并且训练更快,避免过拟合的效果更好,在森林火灾数据集和标准数据集cifar10上都具有较好的表现。
        结论  本文所提出的Sparse-DenseNet模型在森林火灾识别问题上,可以有效优化传统模型存在的问题,并取得良好的识别效果,其准确率可达到99.33%,优于DesenNet的98.15%,并且相同轮次训练时间只有DenseNet训练时间的3/4左右。

       

      Abstract:
        Objective  The frequent occurrence of forest fire brings great difficulty to forest prevention and control. Traditional forest fire identification algorithms have problems such as low accuracy and insufficient processing efficiency. At the same time, forest fire image data itself has strong complexity. It is necessary to comprehensively consider the recognition accuracy and generalization ability. For this reason, this paper uses the Sparse-DenseNet model to carry out the identification research of forest fire.
        Method  Firstly, the DenseNet model was sparsely transformed, and the sparse effect was generated by randomly shielding the nodes in the Dense Block module, so that the algorithm had the advantages of reducing over-fitting, mitigating gradient disappearance and accelerating convergence speed. Secondly, when image acquisition was carried out in the forest area, due to the relative motion between the camera device and the object being collected and the effect of light and shadow, the image data may be disturbed. Therefore, this paper uses python image processing tools to transform the image. The image dataset was expanded accordingly to fit the actual application scenario. Finally, this paper compares the performance of the Sparse-DenseNet model with other classical deep learning models on the forest fire dataset and the cifar10 dataset, and observes its effects.
        Result  The Sparse-DenseNet model had the characteristics of being lighter in structure, training faster, better effect of avoiding over-fitting. It had a good performance in the forest fire data set and the standard data set cifar10.
        Conclusion  The Sparse-DenseNet model proposed in this paper can effectively optimize the problems existing in traditional forest models and achieve good recognition results. The accuracy rate can reach 99.33%, which is better than DesenNet’s 98.15%, and the training time is only about three-quarters of DenseNet in same round.

       

    /

    返回文章
    返回