基于深度学习的林火图像识别算法及实现

发布时间：2018-06-15 20:48

本文选题：林火识别 + 图像处理　；参考：《北京林业大学》2016年硕士论文

【摘要】：森林火灾的发生对经济、生态环境等会产生巨大影响。世界各国对林火监测系统的研究都非常重视。传感器监测方法受环境影响较大,不适宜应用在大尺度空间的森林火灾监测上。传统图像型监测方法需要对图像进行必要处理,人工提取特征,特征选择成为了能否达到理想效果的关键因素。针对上述情况,本文将林火识别与机器学习领域的深度学习算法相结合,将其中的卷积神经网络模型应用在林火识别上。深度网络可以自动提取输入图像特征,通过层与层之间的传递,将底层特征组合形成高层的抽象特征,避免了传统方法中人工提取特征的复杂性和盲目性。深度卷积神经网络特有的局部感受野和权值共享技术减少了参数的数目,降低了算法训练的难度,下采样的使用增强了网络容忍图像畸变的能力。实验结果证明,该方法取得了较为理想的效果。本文的主要工作如下：(1)通过实验和网上搜集构建林火数据库。(2)人工提取图像特征,使用目前比较常用的方法如支持向量机、径向基函数网络、BP神经网络对特征进行识别,并对实验结果进行分析。(3)在对卷积神经网络深入研究的基础上,对森林火灾图像进行有无火灾的辨别,针对夜晚和白天背景不同的情况,设计不同的网络进行识别,对不同的结构,不同的参数进行比较分析。最终得到的夜晚林火识别模型正确率为95.71%,白天林火识别模型正确率为98%,与人工提取特征的传统图像型监测方法相比,具有显著优势。
[Abstract]:The occurrence of forest fires has a great impact on the economy, the ecological environment and so on. All countries of the world have paid great attention to the research of forest fire monitoring system. The sensor monitoring method is greatly influenced by the environment and is not suitable for application in the monitoring of forest fire in large scale space. The traditional image monitoring method needs the necessary processing of the image and artificial extraction. Feature selection is the key factor to achieve the ideal effect. In this case, this paper combines the forest fire recognition with the depth learning algorithm in the machine learning field, and applies the convolution neural network model in the forest fire recognition. The depth network can automatically extract the feature of the input image, through the layer and the layer. In order to avoid the complexity and blindness of the artificial extraction of the traditional methods, the specific local receptive field and weight sharing technology of the deep convolution neural network reduce the number of parameters and reduce the difficulty of the algorithm training. The use of lower sampling enhances the network tolerance of image distortion. The experimental results show that the method has achieved better results. The main work of this paper is as follows: (1) build forest fire database through experiment and online collection. (2) artificial extraction of image features, using the commonly used methods such as support vector machine, radial basis function network, BP neural network to identify the characteristics, and the actual The results are analyzed. (3) on the basis of the thorough research on the convolution neural network, there is no fire discrimination on the forest fire images. According to the different circumstances in the night and the daytime, different networks are designed and the different structures and different parameters are compared and analyzed. The final recognition model of the night forest fire is correct. The accuracy rate of forest fire recognition model is 98% during the daytime, and has a significant advantage compared with the traditional image monitoring method based on artificial feature extraction. 95.71%.
【学位授予单位】：北京林业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】