关键词: 全卷积神经网络, 跳跃连接, 条件随机场, 图像分割

Abstract: As traditional image classification methods are susceptible to noise, this paper proposes a method for forest image recognition and classification based on fully convolutional neural network, which does not require preprocessing. The fully convolutional neural network model can extract features by the convolution layers, then image initial resolution is recovered by the unpooling layers and deconvolution layers. Due to the skip architecture, the model not only overcomes the problem of the gradient vanishing or explosion, but also simplifies the fully convolutional neural network. In order to prevent the over-fitting of the network, Dropout size is adopted. The hiding units are activated randomly. At the later stage, the details of the object’s edge is restored via full-connection conditional random field, which can optimize the final prediction result. The method sets a good performance at the forest image segmentation.

Key words: fully convolutional neural network, skip architecture, conditional random field, image segmentation