Fast cloud image segmentation with superpixel analysis based convolutional networks

Abstract

Due to the various noises, the cloud image segmentation becomes a big challenge for atmosphere prediction. CNN is capable of learning discriminative features from complex data, but this may be quite time-consuming in pixel-level segmentation problems. In this paper we propose superpixel analysis based CNN (SP-CNN) for high efficient cloud image segmentation. SP-CNN employs image over-segmentation of superpixels as basic entities to preserve local consistency. SP-CNN takes the image patches centered at representative pixels in every superpixel as input, and all superpixels are classified as cloud or non-cloud part by voting of the representative pixels. It greatly reduces the computational burden on CNN learning. In order to avoid the ambiguity from superpixel boundaries, SP-CNN selects the representative pixels uniformly from the eroded superpixels. Experimental analysis demonstrates that SP-CNN guarantees both the effectiveness and efficiency in cloud segmentation.