Image matting in the perception granular deep learning

In the past decade, proposed by Geoffrey Hinton, deep learning has been proved its powerful ability in processing data from lower level to higher level and gradually composes more and more semantic concepts by unsupervised feature learning for single modalities (e.g., text, images or audio). Usually a multi scale pyramid structure is applied in a layered deep learning neural network. But how to design a multi scale pyramid structure is still an open problem. At the same time, granular computing (GrC) has been an active topic of research in machine learning and computer vision. In this paper, inspired by the original insight of granular computing proposed by Zadeh, a generalized image-matting approach is defined in the framework of a novel Granular Deep Learning(GDL), in which the information similarity, proximity and functionality are very important for feature learning. We show that layered deep learning can be formally represented as a framework of a granular system defined by fuzzy logic. In this way, the pyramids or hierarchical structure of a layered deep learning neural network can be easily designed in such a granular system, i.e., the convolution pyramids or hierarchical convolutional factor analysis in the deep learning can be viewed as special cases of granular computing. The experiments show the effectiveness of our approach in the task of foreground and background separating.