一种改进的林火实例分割深度学习模型

An Improved Deep Learning Model for Forest Fire Instance Segmentation

林火识别是森林防火中的关键环节,对于早期火灾扑救和森林资源保护具有重要意义。本文提出了一种利用卷积神经网络对林火进行实例分割的模型。相比于原始模型Mask R-CNN,该模型的主要变化如下:(1)对主干特征提取网络和掩膜生成网络进行重构;(2)简化了目标分类和边界框回归的过程;(3)将非极大值抑制算法替换为新的区域分组和滤波算法。在本研究中,使用了5000张林火图像及其手动分割生成的掩膜。大量实验结果表明:所提出的优化模型在略微降低mIoU(81.44%)和mAP(60.52%)的情况下,能够大幅度提高识别效率(9FPS)。因此,本文的模型可以较好地实现对林火进行实例分割,并为相关技术提供参考和借鉴。

Forest fire recognition is a key link in forest fire prevention,which is significant for early fire fighting and forest resource protection.In this paper,a model for instance segmentation of forest fire using convolutional neural networks is proposed.Compared with the original model Mask R-CNN,the main changes of this model are as follows:(1)The backbone feature extraction network and mask generation network are reconfigured;(2)The process of object classification and bounding box regression is simplified;(3)The non-maximum suppression algorithm is replaced with a new region grouping and filtering algorithm.In this work,5000 forest fire images and manually-segmented masks are used.Extensive experimental results show that the proposed model is able to improve the recognition efficiency(10FPS)substantially with slightly reduced mIoU(85.43%)and mAP(65.51%).Therefore,the model in this paper can better realize forest fire instance segmentation,and provide reference for related technologies.