Semantic segmentation of remote sensing image based on deep fusion networks and conditional random field

Abstract

Image semantic segmentation refers to segmenting an image into several groups of pixel regions with different specific semantic meanings and identifying the categories of each region. In recent years, the common semantic segmentation methods that are based on Convolutional Neural Networks(CNN) have realised the pixel-to-pixel image semantic segmentation. They can avoid the problems of artificial design and selection of features in traditional image semantic segmentation methods. As a result of the pooling operation and lack of context information, the detailed information of images is neglected, the precision of the final image semantic segmentation result is low and the segmentation edge is inaccurate. Therefore, this study proposes a semantic segmentation method for remote sensing image on the basis of Deep Fusion Networks(DFN) combined with a conditional random field model.The method initially builds a DFN model in a Fully Convolutional Network(FCN) framework with a deconvolutional fusion structure.On the one hand, the multiscale features can be extracted through the deep networks, which can avoid the artificial design and selection of features to improve the generalisation ability of the model. On the other hand, the multiscale information is used in the model with the help of the deconvolutional fusion structure. The processing accuracy of the model is also improved by fusing the shallow detail information and deep semantic information. Fundamentally, the fully connected conditional random field is introduced to supplement the spatial context information towards precisely locating the boundary and obtaining final semantic segmentation results.From this study, we can draw the following conclusions:(1)With the increase in the depth of the fusion layer, detailed information becomes abundant, the semantic segmentation results become refined and the edge contour becomes close to the label image;(2) The fully connected conditional random field model synthesises the global and local information of the remote sensing image and further improves the efficiency and accuracy of the final semantic segmentation results.