结合条件随机场与神经网络的语义分割方法综述

Semantic segmentation is the fundamental process of natural image understanding and analysis. It requires that the applied algorithm has a thorough understanding of objects inside the image, relationships of these objects, local and global context information of the image. Moreover, these algorithms should model the image on the basis of image context, structure, color, texture, position of pixels, super-pixels or the whole image in order to achieve fine-level segmentation. Considering the consistency constraint for the image segmentation task, most state-of-the-art methods employ the strategy of fusioning deep convolutional neural networks and probability graphical models. Depending on the fusion strategies, these methods could be classified into two categories. The first series of methods treat the graphical model and the neural network as separate modules, and adjust their parameters individually. They propose to use graphical models to post-process the output of the neural network for finer segmentation results. The other series of methods incorporate graphical models into neural networks, and alter their parameters collectively. These end-to-end methods could easily be plugged into existing systems and achieve better results on most datasets than the first series of methods.