基于Mask-RCNN的林火火焰区域提取方法研究
摘要：森林火灾在森林三大自然灾害（火灾、虫害、病害）中居首位，是森林、森林生态系统和人类造成毁灭性的灾难。近年，世界各国均投入大量的人力、物力、财力，重视森林防火监测工作。许多的传统分类方法根据火焰燃烧时的特征选取火焰的圆形度、尖角个数、面积的变化和质心点的运动等特征作为区分火焰的特征，然后再用向量式样本训练分类器完成识别。对于此类研究，应用这些传统特征的火焰检测率并不高，且特征表达形式不够抽象，不能够很好的适应用多种复杂环境。本课题所提出的一种基于Mask-RCNN 的林火火焰区域的自动提取方法，便可实现对林火的自动精确分割和提取。相对于传统直接考虑火焰特征的模型来说，本课题所提出的识别方法存在以下三方面优势：1）Mask-RCNN模型易于提取出难以量化的抽象特征，使其检测精度高；2）其受环境背景影响较小，且它不仅可对图像中的目标进行检测，还可以对每一个目标给出一个高质量的分割结果；3）因为其训练速度快，更加有望实现对林火监测的实时性要求。
关键词：林火识别；目标检测；Mask-RCNN。
Research on Extraction Method of Forest Fire Flame Region Based on Mask-RCNN
Abstract: Forest fire ranks first among the three natural disasters (fires, pests, and diseases) in the forest. It is a devastating disaster caused by forests, forest ecosystems and humans. In recent years, countries in the world have invested a lot of manpower, material resources and financial resources to pay attention to forest fire prevention monitoring. Many traditional classification methods select the roundness of the flame, the number of sharp corners, the change in area, and the movement of the centroid point as the characteristics of the flame according to the characteristics of the flame when it is burning, and then use the vector sample training classifier to complete the recognition. . For this kind of research, the flame detection rate using these traditional features is not high, and the feature expression is not abstract enough to adapt to a variety of complex environments. This topic proposed an automatic extraction method of forest fire flame area based on Mask-RCNN, which can realize automatic and accurate segmentation and extraction of forest fire. Compared with traditional models that directly consider flame features, the recognition method proposed in this topic has the following three advantages: 1) The Mask-RCNN model is easy to extract abstract features that are difficult to quantify, so that its detection accuracy is high; 2) It suffers The environmental background has little influence, and it can not only detect the targets in the image, but also give a high-quality segmentation result for each target; 3) Because of its fast training speed, it is more expected to realize real-time forest fire monitoring Sexual requirements.
Keywords：Forest fire recognition; target detection; Mask-RCNN.
前言
随着我国林业信息化建设的不断深入，解决实时性的林火监测问题已成为了亟待解决的问题。已有的研究成果中，传统算法中大量介入人本身的经验，并没有办法完整的考虑到林火图像的特性，加之环境的复杂性，必然导致采用统一不变的规则无法达到有效识别结果。深度网络由于其对特征的抽象性，使得处理图像上的效果上远优于传统的机器学习算法。为了解决林业信息化的瓶颈，设计出更有效的智能方法成为十分重要且亟待解决的问题。目前基于深度学习的目标检测，大大提高了图像识别的准确度和识别效率。使得没有固定的几何特征,如形状、颜色、大小等形态的火在基于图像分块的处理上成为可能。本课题从深度学习、计算机视觉技术出发，在基于Mask-RCNN模型的基础上，结合图像特点，为林火图像的识别技术提供技术方案。
研究的目的与意义
