Abstract: With the problem that forest fire and its similar objects are difficult to distinguish, this paper presents a new forest fire image classification approach based on deep neural network of sparse autoencoder. Using an unsupervised learning algorithm sparse autoencoder to learn features of large number of small patches from some unlabeled images has completed the training for deep neural network, and then with the learned features, the features can be extracted from large scale images and be convolved and pooled. It uses pooled features to train the softmax classifier by softmax regression. Experimental results show that this new image classification approach can more effectively distinguish forest fire and its similar objects, red flag, red leaves, etc. than traditional neural network does.

Key words: sparse autoencoder, unsupervised learning, convolve and pooling, softmax regression

摘要： 针对林火与相似目标很难区分的问题，提出一种基于稀疏自编码深度神经网络的林火图像分类新方法。采用无监督的特征学习算法稀疏自编码从无标签图像小块中学习特征参数，完成深度神经网络的训练；利用学习到的特征从原始大小分类图像中提取特征并卷积和均值池化特征；对卷积和池化后的特征采用softmax回归来训练最终softmax分类器。实验结果表明，跟传统的BP神经网络相比，新方法能够更有效区分林火与红旗、红叶等类似物体。

关键词: 稀疏自编码, 无监督学习, 卷积与池化, softmax回归